Datasets:

neeto-lab

/

arxiv-struct-100k

like 0

Follow

Neeto Lab 3

[ [ "0.0773", "0.01625", "0.10144", "0.01625", "0.10144", "0.03087", "0.0773", "0.03087" ], [ "0.157", "0.01381", "0.4844", "0.015434", "0.4844", "0.0333", "0.157", "0.03168" ], [ "0.08215", "0.05768", "0.9155", "0.05768", "0.9155", "0.0723", "0.08215", "0.0723" ], [ "0.08093", "0.0788", "0.9165", "0.0788", "0.9165", "0.09344", "0.08093", "0.09344" ], [ "0.0773", "0.0967", "0.918", "0.09503", "0.918", "0.1129", "0.0773", "0.11456" ], [ "0.08093", "0.117", "0.913", "0.117", "0.913", "0.1316", "0.08093", "0.1316" ], [ "0.0785", "0.1349", "0.9165", "0.1373", "0.9165", "0.1536", "0.0785", "0.1519" ], [ "0.08093", "0.1567", "0.4927", "0.1567", "0.4927", "0.1714", "0.08093", "0.1714" ], [ "0.1135", "0.1941", "0.918", "0.1958", "0.918", "0.2161", "0.1135", "0.2153" ], [ "0.08215", "0.2169", "0.913", "0.2169", "0.913", "0.2316", "0.08215", "0.2316" ], [ "0.08215", "0.2372", "0.9165", "0.2372", "0.9165", "0.2517", "0.08215", "0.2517" ], [ "0.0785", "0.2551", "0.9165", "0.2551", "0.9165", "0.2722", "0.0785", "0.2722" ], [ "0.08093", "0.2754", "0.9165", "0.2754", "0.9165", "0.29", "0.08093", "0.29" ], [ "0.08093", "0.2957", "0.918", "0.2957", "0.918", "0.3103", "0.08093", "0.3103" ], [ "0.08215", "0.3152", "0.913", "0.3152", "0.913", "0.3298", "0.08215", "0.3298" ], [ "0.0773", "0.3323", "0.9155", "0.333", "0.9155", "0.3508", "0.0773", "0.3494" ], [ "0.08093", "0.3525", "0.9165", "0.3525", "0.9165", "0.3704", "0.08093", "0.3704" ], [ "0.11475", "0.3826", "0.9165", "0.385", "0.9165", "0.4028", "0.11475", "0.4004" ], [ "0.08093", "0.4062", "0.9165", "0.4062", "0.9165", "0.421", "0.08093", "0.421" ], [ "0.08093", "0.4258", "0.9155", "0.4258", "0.9155", "0.4402", "0.08093", "0.4402" ], [ "0.0785", "0.4436", "0.918", "0.4436", "0.918", "0.4614", "0.0785", "0.4614" ], [ "0.0773", "0.4622", "0.9155", "0.4631", "0.9155", "0.481", "0.0773", "0.48" ], [ "0.08093", "0.4841", "0.9155", "0.4841", "0.9155", "0.4988", "0.08093", "0.4988" ], [ "0.08215", "0.5034", "0.8384", "0.5034", "0.8384", "0.518", "0.08215", "0.518" ], [ "0.11475", "0.5425", "0.918", "0.5435", "0.918", "0.5615", "0.11475", "0.5596" ], [ "0.0773", "0.562", "0.913", "0.563", "0.913", "0.581", "0.0773", "0.58" ], [ "0.08215", "0.584", "0.9165", "0.584", "0.9165", "0.5986", "0.08215", "0.5986" ], [ "0.08093", "0.605", "0.9165", "0.6035", "0.9165", "0.618", "0.08093", "0.619" ], [ "0.08215", "0.624", "0.9155", "0.624", "0.9155", "0.6387", "0.08215", "0.6387" ], [ "0.08215", "0.6436", "0.9165", "0.6436", "0.9165", "0.658", "0.08215", "0.658" ], [ "0.0773", "0.6597", "0.529", "0.661", "0.529", "0.679", "0.0773", "0.6777" ], [ "0.11475", "0.7", "0.9165", "0.7017", "0.9165", "0.7197", "0.11475", "0.7183" ], [ "0.08093", "0.722", "0.9116", "0.722", "0.9116", "0.737", "0.08093", "0.737" ], [ "0.08093", "0.743", "0.9165", "0.743", "0.9165", "0.758", "0.08093", "0.758" ], [ "0.0785", "0.7603", "0.913", "0.7603", "0.913", "0.7773", "0.0785", "0.7773" ], [ "0.08093", "0.782", "0.9155", "0.782", "0.9155", "0.796", "0.08093", "0.796" ], [ "0.08093", "0.802", "0.9116", "0.802", "0.9116", "0.8164", "0.08093", "0.8164" ], [ "0.0785", "0.8213", "0.2246", "0.8213", "0.2246", "0.836", "0.0785", "0.836" ], [ "0.1171", "0.8506", "0.913", "0.8516", "0.913", "0.869", "0.1171", "0.868" ], [ "0.0785", "0.87", "0.9165", "0.8716", "0.9165", "0.8896", "0.0785", "0.8877" ], [ "0.08093", "0.891", "0.918", "0.891", "0.918", "0.909", "0.08093", "0.909" ], [ "0.08215", "0.912", "0.918", "0.912", "0.918", "0.927", "0.08215", "0.927" ], [ "0.08093", "0.9316", "0.868", "0.9316", "0.868", "0.9463", "0.08093", "0.9463" ] ]

[ "4", "Y.Zou,Z.Luo, and J.-B.Huang", "alone, however, is not sufficient due to the ambiguity on textureless regions and", "occlusion boundaries.Hence, the network training is often unstable and requires", "careful hyper-parameter tuning of the loss functions. Our approach builds upon", "existing unsupervised losses fortraining our depth and flow networks.We show", "thatthe proposed cross-task consistencyloss provides a sizableperformance", "boost over individuallytrained models.", "Methods exploiting geometry cues. Recently, a number of work exploits", "the geometric relationship between depth,camera pose,and flowforlearning", "depthorflowmodels[60,65,68,73l.Thesemethodsfirst estimate the depthof", "the input images. Together with the estimated camera poses between two con-", "secutive frames,thesemethods\"synthesize\" the flowfield of rigid regions.The", "synthesizedflowfromdepthandposecaneither beusedforflowprediction in", "rigid regions60,65,68,48 asisor usedfor viewsynthesistotraindepthmodel us", "ingmonocularvideos[73].Additional cuessuchassurfacenormal [67],edge[66],", "physical constraints [59] can be incorporated to further improve the performance.", "These approaches exploit the inherent geometric relationship between struc-", "ture and motion.However, the errors produced by either the depth or the camera", "pose estimation propagate to flow predictions.Our key insight is thatfor rigid", "regions the estimated flow (from flow prediction network) and the synthesized", "rigid flow (from depth and camera pose networks) should be consistent. Con-", "sequently,coupled training allows both depth and flow networksto learn from", "eachotherand enforcegeometricallyconsistentpredictions ofthescene", "Structure from motion. Joint estimation of structure and camera pose", "from multiple images of a given scene is a long-standing problem [46,15,64]", "Conventional methods can recover (semi-)dense depth estimation and camera", "posethroughkeypointtracking/matching.Theoutputsof thesealgorithms can", "potentially be used to help train a flow network, but not the other way around.", "Our work differs as we are also interested in learning a depth network to recover", "dense structure from a single input image.", "Multi-task learning. Simultaneously addressing multiple tasks through", "multi-task learning [52] has shown advantages over methods that tackle indi", "vidual onesJ70l.For examples,joint learning of videosegmentation and optical", "flow through layered models [6,56] or feature sharing [9l helps improve accu", "racyat motion boundaries.Single-view depthmodel learning can also benefit", "from joint trainingwith surfacenormal estimation35,67or semantic segmen", "tation[13.30]", "Our approach tackles the problems of learning both depth and flowmodels", "Unlike existing multi-task learning methods that often require direct supervision", "using ground truth training data for each task, our approach instead leverage", "meta-supervision to couplethetraining of depth and flow models.While our", "models are jointly trained, they can be applied independently at test time." ]

{ "arxiv_id": "1809.01649", "language": "en", "page_idx": 3 }

[ [ "0.262", "0.01706", "0.836", "0.01706", "0.836", "0.03087", "0.262", "0.03087" ], [ "0.901", "0.01625", "0.9214", "0.01625", "0.9214", "0.03087", "0.901", "0.03087" ], [ "0.704", "0.06012", "0.768", "0.06012", "0.768", "0.06415", "0.704", "0.06415" ], [ "0.7114", "0.0666", "0.7417", "0.0666", "0.7417", "0.078", "0.7114", "0.078" ], [ "0.261", "0.08124", "0.268", "0.08124", "0.268", "0.08527", "0.261", "0.08527" ], [ "0.4734", "0.11127", "0.506", "0.11127", "0.506", "0.11536", "0.4734", "0.11536" ], [ "0.6387", "0.212", "0.656", "0.212", "0.656", "0.2161", "0.6387", "0.2161" ], [ "0.79", "0.2185", "0.837", "0.2185", "0.837", "0.2258", "0.79", "0.2258" ], [ "0.6196", "0.2291", "0.674", "0.2291", "0.674", "0.2494", "0.6196", "0.2494" ], [ "0.0833", "0.2883", "0.8804", "0.2883", "0.8804", "0.303", "0.0833", "0.303" ], [ "0.08215", "0.3079", "0.8755", "0.3079", "0.8755", "0.3225", "0.08215", "0.3225" ], [ "0.08093", "0.3274", "0.864", "0.3274", "0.864", "0.342", "0.08093", "0.342" ], [ "0.08093", "0.3477", "0.878", "0.3477", "0.878", "0.3623", "0.08093", "0.3623" ], [ "0.08575", "0.3665", "0.907", "0.3665", "0.907", "0.381", "0.08575", "0.381" ], [ "0.08093", "0.3843", "0.907", "0.3843", "0.907", "0.4014", "0.08093", "0.4014" ], [ "0.08093", "0.4045", "0.8623", "0.4045", "0.8623", "0.4192", "0.08093", "0.4192" ], [ "0.08093", "0.4248", "0.818", "0.4248", "0.818", "0.4395", "0.08093", "0.4395" ], [ "0.0773", "0.442", "0.8804", "0.4412", "0.8804", "0.459", "0.0773", "0.4597" ], [ "0.08093", "0.4631", "0.882", "0.4631", "0.882", "0.4775", "0.08093", "0.4775" ], [ "0.08093", "0.4824", "0.9116", "0.4824", "0.9116", "0.497", "0.08093", "0.497" ], [ "0.08215", "0.503", "0.843", "0.503", "0.843", "0.5176", "0.08215", "0.5176" ], [ "0.08093", "0.5225", "0.906", "0.5225", "0.906", "0.536", "0.08093", "0.536" ], [ "0.08215", "0.541", "0.907", "0.541", "0.907", "0.5557", "0.08215", "0.5557" ], [ "0.08093", "0.5615", "0.9", "0.5615", "0.9", "0.576", "0.08093", "0.576" ], [ "0.0773", "0.5776", "0.8745", "0.58", "0.8745", "0.5977", "0.0773", "0.595" ], [ "0.08093", "0.6", "0.89", "0.6", "0.89", "0.6143", "0.08093", "0.6143" ], [ "0.08215", "0.619", "0.5337", "0.619", "0.5337", "0.633", "0.08215", "0.633" ], [ "0.577", "0.619", "0.8745", "0.619", "0.8745", "0.633", "0.577", "0.633" ], [ "0.08093", "0.6387", "0.8936", "0.6387", "0.8936", "0.6523", "0.08093", "0.6523" ], [ "0.08215", "0.658", "0.703", "0.658", "0.703", "0.673", "0.08215", "0.673" ], [ "0.08215", "0.7124", "0.10266", "0.7124", "0.10266", "0.7256", "0.08215", "0.7256" ], [ "0.128", "0.71", "0.818", "0.71", "0.818", "0.728", "0.128", "0.728" ], [ "0.0761", "0.744", "0.3574", "0.7456", "0.3574", "0.7627", "0.0761", "0.762" ], [ "0.0785", "0.7734", "0.9165", "0.775", "0.9165", "0.792", "0.0785", "0.791" ], [ "0.08215", "0.7954", "0.9155", "0.7954", "0.9155", "0.81", "0.08215", "0.81" ], [ "0.08093", "0.8154", "0.9165", "0.8154", "0.9165", "0.8296", "0.08093", "0.8296" ], [ "0.08093", "0.835", "0.9155", "0.835", "0.9155", "0.8496", "0.08093", "0.8496" ], [ "0.08093", "0.853", "0.9155", "0.853", "0.9155", "0.87", "0.08093", "0.87" ], [ "0.0785", "0.8716", "0.918", "0.8716", "0.918", "0.8896", "0.0785", "0.8896" ], [ "0.0773", "0.89", "0.778", "0.891", "0.778", "0.909", "0.0773", "0.908" ], [ "0.1171", "0.909", "0.918", "0.9116", "0.918", "0.9287", "0.1171", "0.927" ], [ "0.08093", "0.9316", "0.9155", "0.9316", "0.9155", "0.9463", "0.08093", "0.9463" ] ]

[ "UnsupervisedJointLearningusingCross-Task Consistency", "5", "iordhanaae", "Depth", "", "2inidE", "", "OoticalFl", "Bpecwnig", "Fig.3:Overview of our unsupervised joint learning framework.Our", "framework consists of threemajormodules:(1)a DepthNetforsingle-view", "depth estimation:(2) a Pose Net that takes two stacked input frames and", "estimates therelative camerapose between thetwo input frames;and (3)a", "FlowNetthat estimatesdenseoptical flowfieldbetweenthetwo inputframes", "Given a pair of input images It and It+i sampled from an unlabeled video, we", "first estimate the depth of each frame, the 6D camerapose, and the dense", "forward and backward flows.Using thepredicted scene depth and the", "estimated camera pose, we can synthesize 2D forward and backward optical", "flows(referredas riqidflow)bybackprojectingtheinduced3Dforward and", "backward sceneflows(Section 3.2).As we do not haveground truth depth and", "flowmapsfor supervision,weleveragestandardphotometric and spatial", "smoothnesscoststoregularizethenetworktraining(Section 3.3,notshownin", "thisfigurefor clarity).To enforcethe consistency offlowand depth prediction", "inbothdirections,weexploit theforward-backward consistency(Section3.4)", "and adopt the valid masks derived from it to filter out invalid regions (e.g.,", "occlusion/dis-occlusion)forthephotometricloss.Finallyweproposeanovel", "cross-networkconsistencyloss(Section3.5", "encouragingtheopticalflow", "estimation(fromtheFlowNet)andtherigidflow(fromtheDepthandPose", "Net)to be consistent to each otherwithin in valid regions.", "3", "Unsupervised Joint Learning of Depth and Flow", "3.1Method overview", "Our goal is to develop an unsupervised learning framework for jointly training", "thesingle-view depth estimation networkand theopticalflowpredictionnetwork", "using unlabeled video sequences.Figure 3 shows the high-level sketch of our", "proposedapproach.Given two consecutive frames (It,It+i)sampledfrom an", "unlabeled video, we first estimate depth of frame It and It+i, and forward-", "backward optical flow fields between frame It and It+i. We then estimate the", "6D camera pose transformation between the two frames (It, It+i):", "Withthepredicted depthmap and theestimated 6D camerapose,we can", "producethe3Dsceneflowinducedfromcameraego-motionandbackproject" ]

{ "arxiv_id": "1809.01649", "language": "en", "page_idx": 4 }

[ [ "0.1234", "0.1194", "0.881", "0.1194", "0.881", "0.133", "0.1234", "0.133" ], [ "0.3938", "0.1401", "0.6094", "0.1401", "0.6094", "0.155", "0.3938", "0.155" ], [ "0.22", "0.1729", "0.7827", "0.1729", "0.7827", "0.187", "0.22", "0.187" ], [ "0.3477", "0.1895", "0.6533", "0.1895", "0.6533", "0.2031", "0.3477", "0.2031" ], [ "0.4148", "0.2037", "0.588", "0.2037", "0.588", "0.2173", "0.4148", "0.2173" ], [ "0.3494", "0.2173", "0.656", "0.2173", "0.656", "0.2322", "0.3494", "0.2322" ], [ "0.4634", "0.2505", "0.538", "0.2505", "0.538", "0.2625", "0.4634", "0.2625" ], [ "0.1184", "0.2654", "0.8833", "0.2654", "0.8833", "0.279", "0.1184", "0.279" ], [ "0.12177", "0.2798", "0.882", "0.2798", "0.882", "0.2903", "0.12177", "0.2903" ], [ "0.1184", "0.2903", "0.8833", "0.2903", "0.8833", "0.304", "0.1184", "0.304" ], [ "0.1184", "0.304", "0.8833", "0.3027", "0.8833", "0.3164", "0.1184", "0.3176" ], [ "0.12006", "0.3176", "0.882", "0.3176", "0.882", "0.328", "0.12006", "0.328" ], [ "0.1184", "0.33", "0.885", "0.33", "0.885", "0.3433", "0.1184", "0.3433" ], [ "0.11755", "0.3425", "0.8833", "0.3416", "0.8833", "0.3557", "0.11755", "0.357" ], [ "0.11755", "0.3557", "0.8833", "0.3552", "0.8833", "0.3694", "0.11755", "0.3706" ], [ "0.1184", "0.3694", "0.6465", "0.3694", "0.6465", "0.383", "0.1184", "0.383" ], [ "0.1184", "0.3943", "0.882", "0.3943", "0.882", "0.4092", "0.1184", "0.4092" ], [ "0.11755", "0.408", "0.3552", "0.4067", "0.3552", "0.4216", "0.11755", "0.4229" ], [ "0.2183", "0.437", "0.3862", "0.4382", "0.3862", "0.4536", "0.2183", "0.4531" ], [ "0.511", "0.4407", "0.881", "0.4407", "0.881", "0.4536", "0.511", "0.4536" ], [ "0.5146", "0.4548", "0.8833", "0.4548", "0.8833", "0.4685", "0.5146", "0.4685" ], [ "0.1184", "0.4697", "0.4902", "0.4697", "0.4902", "0.4827", "0.1184", "0.4827" ], [ "0.5146", "0.4685", "0.8833", "0.4685", "0.8833", "0.4822", "0.5146", "0.4822" ], [ "0.1184", "0.484", "0.4902", "0.484", "0.4902", "0.4976", "0.1184", "0.4976" ], [ "0.511", "0.481", "0.8833", "0.4822", "0.8833", "0.4963", "0.511", "0.4954" ], [ "0.1184", "0.4976", "0.4902", "0.4976", "0.4902", "0.511", "0.1184", "0.511" ], [ "0.511", "0.4954", "0.8833", "0.4963", "0.8833", "0.511", "0.511", "0.5103" ], [ "0.1184", "0.5127", "0.4902", "0.5127", "0.4902", "0.5254", "0.1184", "0.5254" ], [ "0.513", "0.511", "0.8833", "0.511", "0.8833", "0.525", "0.513", "0.525" ], [ "0.1184", "0.5254", "0.4902", "0.5254", "0.4902", "0.5405", "0.1184", "0.5405" ], [ "0.511", "0.5254", "0.8833", "0.5254", "0.8833", "0.539", "0.511", "0.539" ], [ "0.1184", "0.5405", "0.4895", "0.5405", "0.4895", "0.554", "0.1184", "0.554" ], [ "0.513", "0.539", "0.8833", "0.539", "0.8833", "0.554", "0.513", "0.554" ], [ "0.1184", "0.5547", "0.492", "0.5547", "0.492", "0.5684", "0.1184", "0.5684" ], [ "0.5146", "0.554", "0.8833", "0.554", "0.8833", "0.568", "0.5146", "0.568" ], [ "0.1184", "0.568", "0.4878", "0.568", "0.4878", "0.5806", "0.1184", "0.5806" ], [ "0.513", "0.5684", "0.8833", "0.5684", "0.8833", "0.582", "0.513", "0.582" ], [ "0.11755", "0.582", "0.4895", "0.582", "0.4895", "0.5957", "0.11755", "0.5957" ], [ "0.5146", "0.5845", "0.882", "0.5845", "0.882", "0.594", "0.5146", "0.594" ], [ "0.1184", "0.597", "0.4902", "0.597", "0.4902", "0.611", "0.1184", "0.611" ], [ "0.51", "0.5957", "0.8833", "0.5957", "0.8833", "0.6094", "0.51", "0.6094" ], [ "0.12177", "0.6133", "0.4895", "0.6133", "0.4895", "0.6235", "0.12177", "0.6235" ], [ "0.51", "0.611", "0.8833", "0.61", "0.8833", "0.6235", "0.51", "0.6245" ], [ "0.1184", "0.626", "0.4902", "0.626", "0.4902", "0.6396", "0.1184", "0.6396" ], [ "0.511", "0.6245", "0.8833", "0.6245", "0.8833", "0.638", "0.511", "0.638" ], [ "0.1184", "0.64", "0.4902", "0.64", "0.4902", "0.654", "0.1184", "0.654" ], [ "0.513", "0.6396", "0.8833", "0.6396", "0.8833", "0.653", "0.513", "0.653" ], [ "0.12177", "0.655", "0.4878", "0.655", "0.4878", "0.665", "0.12177", "0.665" ], [ "0.513", "0.654", "0.876", "0.654", "0.876", "0.6675", "0.513", "0.6675" ], [ "0.1184", "0.6685", "0.4902", "0.6685", "0.4902", "0.6816", "0.1184", "0.6816" ], [ "0.511", "0.6675", "0.882", "0.6675", "0.882", "0.681", "0.511", "0.681" ], [ "0.11755", "0.6816", "0.4902", "0.681", "0.4902", "0.6953", "0.11755", "0.697" ], [ "0.513", "0.683", "0.8833", "0.683", "0.8833", "0.697", "0.513", "0.697" ], [ "0.1184", "0.697", "0.4902", "0.697", "0.4902", "0.7104", "0.1184", "0.7104" ], [ "0.511", "0.697", "0.8833", "0.697", "0.8833", "0.7104", "0.511", "0.7104" ], [ "0.1184", "0.7104", "0.4895", "0.709", "0.4895", "0.7236", "0.1184", "0.7246" ], [ "0.5146", "0.7134", "0.882", "0.7134", "0.882", "0.7236", "0.5146", "0.7236" ], [ "0.11755", "0.7246", "0.4902", "0.7246", "0.4902", "0.7383", "0.11755", "0.7383" ], [ "0.513", "0.7256", "0.885", "0.7256", "0.885", "0.7393", "0.513", "0.7393" ], [ "0.1184", "0.7393", "0.4902", "0.7393", "0.4902", "0.753", "0.1184", "0.753" ], [ "0.513", "0.7407", "0.8833", "0.7407", "0.8833", "0.7544", "0.513", "0.7544" ], [ "0.12177", "0.7544", "0.4878", "0.7544", "0.4878", "0.764", "0.12177", "0.764" ], [ "0.513", "0.7544", "0.885", "0.7544", "0.885", "0.767", "0.513", "0.767" ], [ "0.1184", "0.7686", "0.4878", "0.7686", "0.4878", "0.782", "0.1184", "0.782" ], [ "0.5146", "0.7686", "0.882", "0.7686", "0.882", "0.782", "0.5146", "0.782" ], [ "0.12006", "0.782", "0.4902", "0.782", "0.4902", "0.796", "0.12006", "0.796" ], [ "0.513", "0.783", "0.882", "0.783", "0.882", "0.7964", "0.513", "0.7964" ], [ "0.12006", "0.7964", "0.4902", "0.7964", "0.4902", "0.81", "0.12006", "0.81" ], [ "0.511", "0.796", "0.885", "0.7964", "0.885", "0.811", "0.511", "0.81" ], [ "0.1184", "0.81", "0.4902", "0.81", "0.4902", "0.8237", "0.1184", "0.8237" ], [ "0.513", "0.811", "0.8833", "0.811", "0.8833", "0.8247", "0.513", "0.8247" ], [ "0.1184", "0.8247", "0.492", "0.8247", "0.492", "0.8384", "0.1184", "0.8384" ], [ "0.513", "0.826", "0.8833", "0.826", "0.8833", "0.839", "0.513", "0.839" ], [ "0.11755", "0.8374", "0.492", "0.8384", "0.492", "0.8525", "0.11755", "0.8516" ], [ "0.513", "0.8384", "0.882", "0.839", "0.882", "0.854", "0.513", "0.8525" ], [ "0.12006", "0.8525", "0.492", "0.8525", "0.492", "0.866", "0.12006", "0.866" ], [ "0.5146", "0.854", "0.8613", "0.854", "0.8613", "0.8677", "0.5146", "0.8677" ], [ "0.1184", "0.8677", "0.4878", "0.8677", "0.4878", "0.8813", "0.1184", "0.8813" ], [ "0.1184", "0.883", "0.4878", "0.883", "0.4878", "0.8936", "0.1184", "0.8936" ], [ "0.529", "0.8813", "0.885", "0.8813", "0.885", "0.8945", "0.529", "0.8945" ], [ "0.1184", "0.8965", "0.4895", "0.8965", "0.4895", "0.91", "0.1184", "0.91" ], [ "0.5425", "0.8945", "0.825", "0.8955", "0.825", "0.91", "0.5425", "0.909" ] ]

[ "AsNER - Annotated Dataset and Baseline for Assamese Named", "Entity recognition", "Dhrubajyoti Pathak, Sukumar Nandi, Priyankoo Sarmah", "Indian Institute of Technology Guwahati", "North Guwahati, India", "[drbj153, sukumar, priyankoo]@itg.ac.in", "Abstract", "We present the AsNER, a named entity annotation dataset for low resource Assamese language with a baseline", "Assamese NERmodel. The dataset contains about 99k tokens comprised of text from the speech of the Prime", "Minister of India and Assamese play. It also contains person names, location names and addresses. The proposed", "NER dataset is likely to be a significant resource for deep neural based Assamese language processing.We", "benchmark the dataset by training NER models and evaluating using state-of-the-art architectures for supervised", "named entity recognition (NER) such as Fasttext, BERT, XLM-R, FLAIR, MuRIL etc. We implement several", "baseline approaches with state-of-the-art sequence tagging Bi-LSTM-CRF architecture. The highest F1-score", "among all baselines achieves an accuracy of 80.69% when using MuRIL as a word embedding method. The", "annotated dataset and the top performing model are made publicly available.", "Keywords:NER dataset, Language Resources, Assamese NER, Assamese Language, Named Entity", "Recognition, NER model, AsNER", "1.Introduction", "tasks on a large amount of data. Suitable anno-", "tated corpora for Assamese NER, name dictionar-", "Named Entity Recognition (NER) aims to clas-", "ies, morphological analyzers, dependency parser,", "sify text in a sentence into predefined classes such", "POS taggers etc., are not yet publicly available", "as person, location, organization etc. It basically", "for suitable use in downstream tasks. Although,", "identifies a word or phrase that can be considered", "Assamese has a very old and rich literary history,", "as names from a set of documents. NER plays", "technology development in NLP is still in a nascent", "a vital role in preprocessing task of various natu-", "stage. The WikiAnn NER dataset (Pan et al.,", "ral language processing (NLP) applications such as", "2017) is the only publicly available dataset that", "information retrieval (Neudecker, 2016), text un-", "contains annotated NER data for 282 languages", "derstanding (Zhang et al., 2019), automatic text", "that exist in Wikipedia: however, the size is not", "summarization (Larsen, 1999), question answering", " large enough to train a neural model. The lack of", "(Molla et al.,2006),machinetranslation (Babych", " a suitable dataset may be the reason that we could", "and Hartley, 2003), and knowledge base construc-", "not find any study about DL-based NER for As-", "tion (Etzioni et al., 2005) etc. Recent advances in", "samese. All the previous studies on Assamese NER", "deep neural network (DL) in NLP exhibit success", "are based on rule-based or statistical approaches.", "in various domains. DL-based NER systems with", " In this paper, we introduce a novel Assamese NER", "minimal feature engineering have been flourish-", "dataset, AsNER, comprising of annotated sen-", "ing. Over the past few years, several studies have", "tences with five entity classes. It also contains a", "been reported success in deep learning-based NER", "large number of person names,locations and or-", "model and achieved state-of-the-art performance", "ganization names. The corpus is built from the", "in resource-rich language (Chiu and Nichols, 2016;", "Assamese translation of the speeches of the Prime", "Lample et al.2016: Akbiket al..2019)The DL-", "Minister of India, available online. Along with the", "based model requires high quality and large anno-", "annotation, preprocessing of the dataset and man-", "tated datasets for training and evaluation. There-", "ual evaluation of the proposed AsNER, we ver-", "fore, datasets play an essential role in extracting", "ify the effectiveness of our dataset by using it to", "the linguistic features of a language to achieve high", "train neural-based NER models. To the best of our", "performance in downstream tasks. On the other", "knowledge, AsNER is the first attempt to develop", "hand, data annotation for a language remains a", "and evaluate NER dataset for Assamese language.", "time consuming and expensive process. It is a", "The summary of our contribution is as follows:", "major challenge for many resource-poor languages,", "such as Assamese (Glottocode:assa1263),as it", ": We prepare and release a novel NER dataset", "requires language experts to perform annotation", " for low-resourced Assamese language." ]

{ "arxiv_id": "2207.03422", "language": "en", "page_idx": 0 }

[ [ "0.1335", "0.0748", "0.4902", "0.0736", "0.4902", "0.0885", "0.1335", "0.08905" ], [ "0.513", "0.0748", "0.885", "0.0748", "0.885", "0.08905", "0.513", "0.08905" ], [ "0.1511", "0.0903", "0.4902", "0.0903", "0.4902", "0.10394", "0.1511", "0.10394" ], [ "0.513", "0.08905", "0.8833", "0.08905", "0.8833", "0.1027", "0.513", "0.1027" ], [ "0.1528", "0.10394", "0.3594", "0.10394", "0.3594", "0.11755", "0.1528", "0.11755" ], [ "0.513", "0.10394", "0.885", "0.10394", "0.885", "0.11755", "0.513", "0.11755" ], [ "0.5146", "0.1194", "0.5684", "0.1194", "0.5684", "0.1312", "0.5146", "0.1312" ], [ "0.1335", "0.1271", "0.4878", "0.1271", "0.4878", "0.1401", "0.1335", "0.1401" ], [ "0.5146", "0.1318", "0.885", "0.1318", "0.885", "0.1455", "0.5146", "0.1455" ], [ "0.1503", "0.1414", "0.44", "0.1414", "0.44", "0.155", "0.1503", "0.155" ], [ "0.513", "0.1467", "0.885", "0.1467", "0.885", "0.1603", "0.513", "0.1603" ], [ "0.1318", "0.1621", "0.4895", "0.1621", "0.4895", "0.177", "0.1318", "0.177" ], [ "0.5146", "0.1621", "0.882", "0.1621", "0.882", "0.1729", "0.5146", "0.1729" ], [ "0.1486", "0.1776", "0.4165", "0.177", "0.4165", "0.1912", "0.1486", "0.1924" ], [ "0.513", "0.1746", "0.885", "0.1746", "0.885", "0.1882", "0.513", "0.1882" ], [ "0.511", "0.187", "0.8867", "0.1895", "0.8867", "0.2037", "0.511", "0.2019" ], [ "0.12006", "0.1996", "0.4902", "0.1996", "0.4902", "0.2131", "0.12006", "0.2131" ], [ "0.513", "0.2037", "0.8867", "0.2037", "0.8867", "0.2173", "0.513", "0.2173" ], [ "0.11755", "0.212", "0.4902", "0.2131", "0.4902", "0.228", "0.11755", "0.2268" ], [ "0.511", "0.2162", "0.885", "0.2173", "0.885", "0.2322", "0.511", "0.231" ], [ "0.1184", "0.228", "0.4895", "0.228", "0.4895", "0.2411", "0.1184", "0.2411" ], [ "0.513", "0.2322", "0.885", "0.2322", "0.885", "0.2452", "0.513", "0.2452" ], [ "0.1184", "0.2423", "0.4902", "0.2423", "0.4902", "0.2559", "0.1184", "0.2559" ], [ "0.5146", "0.2477", "0.882", "0.2477", "0.882", "0.2583", "0.5146", "0.2583" ], [ "0.12006", "0.2583", "0.4878", "0.2583", "0.4878", "0.2683", "0.12006", "0.2683" ], [ "0.513", "0.26", "0.8833", "0.26", "0.8833", "0.2737", "0.513", "0.2737" ], [ "0.11755", "0.2708", "0.4878", "0.2695", "0.4878", "0.2832", "0.11755", "0.284" ], [ "0.513", "0.2756", "0.8833", "0.2756", "0.8833", "0.2861", "0.513", "0.2861" ], [ "0.12006", "0.2852", "0.4895", "0.2852", "0.4895", "0.2986", "0.12006", "0.2986" ], [ "0.511", "0.288", "0.885", "0.288", "0.885", "0.3018", "0.511", "0.3018" ], [ "0.12006", "0.2998", "0.4895", "0.2998", "0.4895", "0.3135", "0.12006", "0.3135" ], [ "0.511", "0.3018", "0.882", "0.3018", "0.882", "0.3154", "0.511", "0.3154" ], [ "0.11755", "0.3123", "0.4878", "0.3123", "0.4878", "0.326", "0.11755", "0.326" ], [ "0.513", "0.3176", "0.882", "0.3176", "0.882", "0.3308", "0.513", "0.3308" ], [ "0.11755", "0.326", "0.44", "0.3267", "0.44", "0.3416", "0.11755", "0.3403" ], [ "0.5146", "0.3333", "0.881", "0.3333", "0.881", "0.3433", "0.5146", "0.3433" ], [ "0.1838", "0.3486", "0.419", "0.3508", "0.419", "0.3682", "0.1838", "0.3665" ], [ "0.5146", "0.3457", "0.8833", "0.3457", "0.8833", "0.3594", "0.5146", "0.3594" ], [ "0.513", "0.3599", "0.882", "0.3599", "0.882", "0.3735", "0.513", "0.3735" ], [ "0.1184", "0.3706", "0.4895", "0.3694", "0.4895", "0.3843", "0.1184", "0.3855" ], [ "0.513", "0.3748", "0.881", "0.3748", "0.881", "0.3855", "0.513", "0.3855" ], [ "0.12006", "0.3855", "0.4895", "0.3855", "0.4895", "0.3984", "0.12006", "0.3984" ], [ "0.513", "0.3884", "0.885", "0.3884", "0.885", "0.402", "0.513", "0.402" ], [ "0.1184", "0.3997", "0.4895", "0.3997", "0.4895", "0.4133", "0.1184", "0.4133" ], [ "0.517", "0.4026", "0.529", "0.4067", "0.523", "0.4146", "0.511", "0.4104" ], [ "0.1184", "0.4146", "0.4878", "0.4146", "0.4878", "0.4275", "0.1184", "0.4275" ], [ "0.11755", "0.427", "0.4895", "0.4275", "0.4895", "0.4412", "0.11755", "0.4407" ], [ "0.5786", "0.427", "0.817", "0.427", "0.817", "0.4407", "0.5786", "0.4407" ], [ "0.12006", "0.4424", "0.4895", "0.4424", "0.4895", "0.456", "0.12006", "0.456" ], [ "0.5146", "0.4453", "0.885", "0.4453", "0.885", "0.459", "0.5146", "0.459" ], [ "0.1184", "0.4573", "0.4902", "0.4573", "0.4902", "0.4702", "0.1184", "0.4702" ], [ "0.513", "0.4602", "0.8833", "0.4602", "0.8833", "0.4739", "0.513", "0.4739" ], [ "0.12177", "0.4727", "0.4895", "0.4727", "0.4895", "0.4827", "0.12177", "0.4827" ], [ "0.51", "0.4727", "0.885", "0.4744", "0.885", "0.4893", "0.51", "0.4868" ], [ "0.12006", "0.485", "0.4902", "0.485", "0.4902", "0.4988", "0.12006", "0.4988" ], [ "0.511", "0.488", "0.8833", "0.488", "0.8833", "0.503", "0.511", "0.503" ], [ "0.1184", "0.4995", "0.4895", "0.4995", "0.4895", "0.513", "0.1184", "0.513" ], [ "0.5146", "0.5044", "0.881", "0.5044", "0.881", "0.514", "0.5146", "0.514" ], [ "0.1184", "0.513", "0.4902", "0.513", "0.4902", "0.527", "0.1184", "0.527" ], [ "0.513", "0.5166", "0.885", "0.5166", "0.885", "0.53", "0.513", "0.53" ], [ "0.12177", "0.5293", "0.4895", "0.5293", "0.4895", "0.539", "0.12177", "0.539" ], [ "0.513", "0.531", "0.8833", "0.531", "0.8833", "0.5444", "0.513", "0.5444" ], [ "0.1184", "0.5415", "0.4265", "0.5415", "0.4265", "0.5547", "0.1184", "0.5547" ], [ "0.5146", "0.5464", "0.8833", "0.5464", "0.8833", "0.557", "0.5146", "0.557" ], [ "0.12006", "0.5557", "0.4895", "0.5557", "0.4895", "0.5693", "0.12006", "0.5693" ], [ "0.511", "0.558", "0.882", "0.558", "0.882", "0.572", "0.511", "0.572" ], [ "0.1184", "0.571", "0.4878", "0.571", "0.4878", "0.5845", "0.1184", "0.5845" ], [ "0.513", "0.5737", "0.885", "0.5737", "0.885", "0.5874", "0.513", "0.5874" ], [ "0.1184", "0.5845", "0.4895", "0.5845", "0.4895", "0.5986", "0.1184", "0.5986" ], [ "0.5146", "0.589", "0.882", "0.589", "0.882", "0.5996", "0.5146", "0.5996" ], [ "0.1184", "0.5986", "0.4902", "0.5986", "0.4902", "0.6123", "0.1184", "0.6123" ], [ "0.511", "0.601", "0.6987", "0.6016", "0.6987", "0.615", "0.511", "0.6143" ], [ "0.11755", "0.6123", "0.4902", "0.6133", "0.4902", "0.627", "0.11755", "0.626" ], [ "0.5146", "0.616", "0.8833", "0.616", "0.8833", "0.63", "0.5146", "0.63" ], [ "0.12006", "0.627", "0.4895", "0.627", "0.4895", "0.64", "0.12006", "0.64" ], [ "0.513", "0.6313", "0.8833", "0.6313", "0.8833", "0.6445", "0.513", "0.6445" ], [ "0.1184", "0.641", "0.4895", "0.641", "0.4895", "0.655", "0.1184", "0.655" ], [ "0.513", "0.6445", "0.885", "0.6445", "0.885", "0.6577", "0.513", "0.6577" ], [ "0.11755", "0.655", "0.4902", "0.6562", "0.4902", "0.6694", "0.11755", "0.6685" ], [ "0.5146", "0.659", "0.885", "0.659", "0.885", "0.673", "0.5146", "0.673" ], [ "0.1184", "0.6694", "0.4878", "0.6685", "0.4878", "0.6816", "0.1184", "0.683" ], [ "0.511", "0.6714", "0.882", "0.6714", "0.882", "0.685", "0.511", "0.685" ], [ "0.1184", "0.684", "0.4895", "0.684", "0.4895", "0.6978", "0.1184", "0.6978" ], [ "0.5146", "0.6885", "0.6514", "0.6885", "0.6514", "0.6987", "0.5146", "0.6987" ], [ "0.12006", "0.6987", "0.4517", "0.6987", "0.4517", "0.712", "0.12006", "0.712" ], [ "0.513", "0.6997", "0.8833", "0.7007", "0.8833", "0.7153", "0.513", "0.7144" ], [ "0.215", "0.7227", "0.419", "0.7227", "0.419", "0.737", "0.215", "0.737" ], [ "0.5146", "0.7173", "0.882", "0.7173", "0.882", "0.728", "0.5146", "0.728" ], [ "0.1873", "0.7236", "0.2225", "0.7236", "0.2225", "0.7354", "0.1873", "0.7354" ], [ "0.513", "0.73", "0.8833", "0.73", "0.8833", "0.7437", "0.513", "0.7437" ], [ "0.12006", "0.741", "0.4902", "0.741", "0.4902", "0.755", "0.12006", "0.755" ], [ "0.5146", "0.7446", "0.879", "0.7446", "0.879", "0.755", "0.5146", "0.755" ], [ "0.1184", "0.755", "0.4878", "0.755", "0.4878", "0.7686", "0.1184", "0.7686" ], [ "0.5146", "0.7603", "0.882", "0.7603", "0.882", "0.771", "0.5146", "0.771" ], [ "0.12006", "0.771", "0.4895", "0.771", "0.4895", "0.7837", "0.12006", "0.7837" ], [ "0.513", "0.7725", "0.8833", "0.7725", "0.8833", "0.786", "0.513", "0.786" ], [ "0.1184", "0.7837", "0.4878", "0.7837", "0.4878", "0.7974", "0.1184", "0.7974" ], [ "0.5146", "0.788", "0.6396", "0.788", "0.6396", "0.799", "0.5146", "0.799" ], [ "0.1184", "0.799", "0.4895", "0.799", "0.4895", "0.8125", "0.1184", "0.8125" ], [ "0.513", "0.8003", "0.885", "0.8003", "0.885", "0.814", "0.513", "0.814" ], [ "0.12177", "0.814", "0.486", "0.814", "0.486", "0.8247", "0.12177", "0.8247" ], [ "0.513", "0.8154", "0.885", "0.8154", "0.885", "0.829", "0.513", "0.829" ], [ "0.12006", "0.8267", "0.4895", "0.8267", "0.4895", "0.8403", "0.12006", "0.8403" ], [ "0.11584", "0.8413", "0.4878", "0.8413", "0.4878", "0.855", "0.11584", "0.855" ], [ "0.534", "0.8413", "0.7866", "0.843", "0.7866", "0.857", "0.534", "0.8555" ], [ "0.534", "0.855", "0.75", "0.857", "0.75", "0.8716", "0.534", "0.869" ], [ "0.1427", "0.869", "0.4182", "0.869", "0.4182", "0.883", "0.1427", "0.883" ], [ "0.532", "0.868", "0.7407", "0.8706", "0.7407", "0.8843", "0.532", "0.882" ], [ "0.12006", "0.8843", "0.2083", "0.8843", "0.2083", "0.8945", "0.12006", "0.8945" ], [ "0.532", "0.8813", "0.797", "0.882", "0.797", "0.8965", "0.532", "0.8955" ], [ "0.1394", "0.8955", "0.4265", "0.8955", "0.4265", "0.909", "0.1394", "0.909" ], [ "0.513", "0.8955", "0.6035", "0.8965", "0.6035", "0.9077", "0.513", "0.9067" ] ]

[ ". We present an evaluation of the AsNER by", "Knowledge Base properties. The size of WikiAnn", "employing a state-of-the-art sequence tagging", "is not too large to train a neural model. The pub-", "BiLSTM-CRF architecture.", "licly available Assamese NER comprises of 12.5k", "tokens.", ". Lastly, we report the performance of eight dif-", "The major part of the dataset is taken from the", "ferent NER models trained on AsNER.", "translated speech of the Prime Minister published", ". The dataset and the best performing trained", "by Press Information Bureau, Govt.of India3", "model are made available publicly'.", "Text from Assamese Wikisource4 and Wikipedia", "5 is also included. Apart from that we include", "This paper is organized as follows.We present a", "person names, location names and organization", "brief overview of the Assamese language in section", " names into the dataset. These persons, locations", "2. We describe our annotated dataset and anno-", "and organization's names, are originally in En-", "tation process in Section 3 and 4 respectively. We", "glish text collected from various sources, and trans-", "illustrate experiment details used to evaluate our", "lated to Assamese using Microsoft translator6. The", "dataset in Section 5. We discuss the results of the", "statistics of sentence and entity count from all the", "performance of different models in Section 6. In", " sources is presented in table 1. The final dataset", "7, we report the challenges encountered in the an-", "comprises 24,040 sentences and approximately 99k", "notation process as well as in training the models.", "tokens. We used 80% of the dataset for training,", "Finally, we conclude our paper in section 8.", "whileapproximately10%of dataset is used for vali-", " 2. Assamese language", "dation and 10% for testing. The additional person,", "location and organization names are proportionally", "Assamese or Asamiya, pronounced, /oxomija/ is", "distributed in the training, validation and test sets.", "an Indo-Aryan language spoken mainly in Assam,", "We summarise the statistics of the dataset in table", "a state of northeast India. It also refers to the na-", "A", "tive of Assam, whose mother tongue is Assamese.", "In this work, through Assamese, we refer to the", "4.Dataset Annotation", "Assamese language. It is a descendant of Magadhi", "The annotated dataset is prepared in two phase;", "Prakrit and bears affinities with Bengali, Hindi and", "first by a POS tagger to tag all the words in the", "Odia.Modern Assamese uses the Assamese script,", "sentence with POS tag. After that, we use the", "which is developed from the Brahmi script. As-", "tagged sentences for further named entity tagging", "samese script is similar to Bengali script except", "by three native Assamese speakers with one lin-", "for two characters, where Assamese differing from", "guist annotator. In second phase, only the nouns", "Bengaliinoneletter()forthe/r/sound,andan", "and the numbers are kept for further checking. The", "extra letter () for the /w/ or /v/ sound.", "rest of the words of the sentence are annotated as", "Assamese is a highly inflectional, morphologically", "“not a named entity\" Table 3 shows the distribu-", "rich and agglutinating language. The rich linguis-", "tion of different classes in the dataset. Conflicts", "tics feature of the language becomes the most chal-", "have been resolved manually. We discuss the con-", "lenging tasks in language processing. The morpho-", "flicting cases in section 7", "logical structure of words changes due to affixation,", "We focus on five classes of named entity during an-", "derivation and compounding. Affixes play an im-", "notation, Location(LOC), Person (PER), Organi-", "portant role in word formation in Assamese. Af-", "zation (ORG), Miscellaneous (MISC) and Number", "fixes are used extensively in the formation of nouns,", "(NUM) (Sang and De Meulder, 2003; Chinchor and", "pronouns, and in the inflection of verbs with re-", "Robinson, 1997: Strotgen and Gertz, 2013: Hvin-", "spect to number, person, tense, aspect and mood.", "gelby et al., 2020).", "Assamese is also a free word order Language.", "LoC includes locations like regions (villages,", "Dataset description", "towns, cities), roads (street name, highway), and", "3.", "natural locations (National park, forest reserve,", "There are a few organized sources of monolingual", "river. garden).as well as both public and commer.", "corpora available for most of the Indian languages.", "cial places like tourist sites, museums, hospitals.", "Among the Indian languages, the Assamese corpus", "airports, stations, markets, play-grounds, restau-", "is one of the smallest in size. The WikiAnn NER?", "rants, hotels etc.", "dataset is the only publicly available dataset for", "PER consists of first, middle and last names of", "Assamese NER.It is created from Wikipedia by", "people, animals, fictional characters, as well as", "transferring named entity labels from English to", "other languages by utilizing cross language link and", "3https://pib.gov.in/indexd.aspx/", "4https://as.wikisource.org/", "lhttps://anonymous.4open.science/r/", "5https://as.wikipedia.org/", "AsNER-04B3/", "6https://www.microsoft.com/en-us/", "2https://elisa-ie.github.io/wikiann/", "translator/" ]

{ "arxiv_id": "2207.03422", "language": "en", "page_idx": 1 }

[ [ "0.3728", "0.0736", "0.6323", "0.0736", "0.6323", "0.0873", "0.3728", "0.0873" ], [ "0.3242", "0.10156", "0.4382", "0.10156", "0.4382", "0.114", "0.3242", "0.114" ], [ "0.503", "0.10156", "0.581", "0.10156", "0.581", "0.1134", "0.503", "0.1134" ], [ "0.601", "0.10156", "0.679", "0.10156", "0.679", "0.1134", "0.601", "0.1134" ], [ "0.3267", "0.1152", "0.4492", "0.1152", "0.4492", "0.1289", "0.3267", "0.1289" ], [ "0.503", "0.1164", "0.54", "0.1164", "0.54", "0.1289", "0.503", "0.1289" ], [ "0.5977", "0.114", "0.6396", "0.114", "0.6396", "0.13", "0.5977", "0.13" ], [ "0.3284", "0.1289", "0.486", "0.1289", "0.486", "0.1426", "0.3284", "0.1426" ], [ "0.4995", "0.1271", "0.5356", "0.13", "0.532", "0.1443", "0.4963", "0.1414" ], [ "0.5977", "0.1289", "0.6323", "0.1289", "0.6323", "0.1443", "0.5977", "0.1443" ], [ "0.5015", "0.1443", "0.54", "0.1443", "0.54", "0.1567", "0.5015", "0.1567" ], [ "0.5977", "0.1437", "0.6465", "0.1437", "0.6465", "0.1592", "0.5977", "0.1592" ], [ "0.5015", "0.1567", "0.5415", "0.1567", "0.5415", "0.1729", "0.5015", "0.1729" ], [ "0.3284", "0.1455", "0.4307", "0.1455", "0.4307", "0.1567", "0.3284", "0.1567" ], [ "0.3284", "0.1592", "0.4165", "0.1592", "0.4165", "0.1704", "0.3284", "0.1704" ], [ "0.5977", "0.1567", "0.6396", "0.1567", "0.6396", "0.1729", "0.5977", "0.1729" ], [ "0.4995", "0.1704", "0.534", "0.1704", "0.534", "0.1859", "0.4995", "0.1859" ], [ "0.5977", "0.1692", "0.6396", "0.1716", "0.638", "0.187", "0.596", "0.1853" ], [ "0.4014", "0.2007", "0.601", "0.2007", "0.601", "0.2144", "0.4014", "0.2144" ], [ "0.3267", "0.1716", "0.461", "0.1716", "0.461", "0.1853", "0.3267", "0.1853" ], [ "0.3787", "0.2257", "0.4492", "0.228", "0.4475", "0.2434", "0.378", "0.2411" ], [ "0.4744", "0.2257", "0.525", "0.228", "0.523", "0.2423", "0.4727", "0.2405" ], [ "0.538", "0.2268", "0.5786", "0.2268", "0.5786", "0.2423", "0.538", "0.2423" ], [ "0.5938", "0.2245", "0.638", "0.2268", "0.636", "0.2423", "0.592", "0.2405" ], [ "0.367", "0.2411", "0.461", "0.2411", "0.461", "0.2559", "0.367", "0.2559" ], [ "0.476", "0.2411", "0.525", "0.2411", "0.525", "0.2559", "0.476", "0.2559" ], [ "0.544", "0.2423", "0.575", "0.2423", "0.575", "0.257", "0.544", "0.257" ], [ "0.594", "0.2411", "0.636", "0.2411", "0.636", "0.257", "0.594", "0.257" ], [ "0.476", "0.2546", "0.525", "0.2546", "0.525", "0.2708", "0.476", "0.2708" ], [ "0.54", "0.2546", "0.58", "0.2546", "0.58", "0.2708", "0.54", "0.2708" ], [ "0.596", "0.257", "0.634", "0.257", "0.634", "0.2695", "0.596", "0.2695" ], [ "0.4768", "0.2695", "0.525", "0.2695", "0.525", "0.284", "0.4768", "0.284" ], [ "0.3787", "0.2559", "0.451", "0.2559", "0.451", "0.2708", "0.3787", "0.2708" ], [ "0.3745", "0.2708", "0.4534", "0.2708", "0.4534", "0.2852", "0.3745", "0.2852" ], [ "0.5415", "0.2708", "0.5786", "0.2708", "0.5786", "0.2832", "0.5415", "0.2832" ], [ "0.5938", "0.2695", "0.636", "0.2695", "0.636", "0.2852", "0.5938", "0.2852" ], [ "0.2568", "0.2998", "0.7466", "0.2998", "0.7466", "0.3142", "0.2568", "0.3142" ], [ "0.5054", "0.3237", "0.602", "0.326", "0.602", "0.3416", "0.5054", "0.339" ], [ "0.3704", "0.3308", "0.4626", "0.3333", "0.4626", "0.3474", "0.3704", "0.3457" ], [ "0.4702", "0.3374", "0.5225", "0.339", "0.5205", "0.3552", "0.4685", "0.3528" ], [ "0.538", "0.3403", "0.574", "0.3403", "0.574", "0.3528", "0.538", "0.3528" ], [ "0.592", "0.3403", "0.634", "0.3403", "0.634", "0.354", "0.592", "0.354" ], [ "0.3652", "0.354", "0.4114", "0.354", "0.4114", "0.3682", "0.3652", "0.3682" ], [ "0.4768", "0.354", "0.5264", "0.354", "0.5264", "0.3682", "0.4768", "0.3682" ], [ "0.5366", "0.354", "0.571", "0.354", "0.571", "0.3694", "0.5366", "0.3694" ], [ "0.5903", "0.354", "0.6265", "0.354", "0.6265", "0.3694", "0.5903", "0.3694" ], [ "0.3687", "0.3665", "0.413", "0.3682", "0.4114", "0.3843", "0.367", "0.3818" ], [ "0.476", "0.3682", "0.525", "0.3682", "0.525", "0.383", "0.476", "0.383" ], [ "0.5366", "0.3682", "0.571", "0.3682", "0.571", "0.3843", "0.5366", "0.3843" ], [ "0.592", "0.3682", "0.6265", "0.3682", "0.6265", "0.3843", "0.592", "0.3843" ], [ "0.5366", "0.3818", "0.5728", "0.3818", "0.5728", "0.3984", "0.5366", "0.3984" ], [ "0.592", "0.3818", "0.6323", "0.3818", "0.6323", "0.398", "0.592", "0.398" ], [ "0.4768", "0.383", "0.509", "0.383", "0.509", "0.398", "0.4768", "0.398" ], [ "0.372", "0.383", "0.4165", "0.383", "0.4165", "0.398", "0.372", "0.398" ], [ "0.3704", "0.3967", "0.4207", "0.3967", "0.4207", "0.412", "0.3704", "0.412" ], [ "0.476", "0.3967", "0.517", "0.3967", "0.517", "0.412", "0.476", "0.412" ], [ "0.5366", "0.3967", "0.5728", "0.3967", "0.5728", "0.412", "0.5366", "0.412" ], [ "0.5903", "0.3967", "0.6265", "0.3967", "0.6265", "0.412", "0.5903", "0.412" ], [ "0.3704", "0.4104", "0.4182", "0.4104", "0.4182", "0.4258", "0.3704", "0.4258" ], [ "0.4768", "0.412", "0.509", "0.412", "0.509", "0.427", "0.4768", "0.427" ], [ "0.538", "0.411", "0.5728", "0.411", "0.5728", "0.427", "0.538", "0.427" ], [ "0.5938", "0.412", "0.614", "0.412", "0.614", "0.4258", "0.5938", "0.4258" ], [ "0.3728", "0.4229", "0.3914", "0.4258", "0.3855", "0.4412", "0.367", "0.4382" ], [ "0.476", "0.4258", "0.525", "0.4258", "0.525", "0.4407", "0.476", "0.4407" ], [ "0.54", "0.427", "0.5786", "0.427", "0.5786", "0.4395", "0.54", "0.4395" ], [ "0.5903", "0.4258", "0.636", "0.4258", "0.636", "0.4407", "0.5903", "0.4407" ], [ "0.546", "0.4631", "0.6055", "0.4631", "0.6055", "0.4785", "0.546", "0.4785" ], [ "0.6455", "0.4644", "0.687", "0.4644", "0.687", "0.4797", "0.6455", "0.4797" ], [ "0.546", "0.4785", "0.5884", "0.4785", "0.5884", "0.4946", "0.546", "0.4946" ], [ "0.12006", "0.4673", "0.1704", "0.4673", "0.1704", "0.4785", "0.12006", "0.4785" ], [ "0.1184", "0.4797", "0.4878", "0.4797", "0.4878", "0.4934", "0.1184", "0.4934" ], [ "0.6465", "0.4785", "0.69", "0.4785", "0.69", "0.4946", "0.6465", "0.4946" ], [ "0.546", "0.4905", "0.5903", "0.4922", "0.5884", "0.509", "0.544", "0.5073" ], [ "0.6455", "0.4934", "0.6885", "0.4934", "0.6885", "0.509", "0.6455", "0.509" ], [ "0.546", "0.5073", "0.586", "0.5073", "0.586", "0.5225", "0.546", "0.5225" ], [ "0.11755", "0.4946", "0.4895", "0.4934", "0.4895", "0.5083", "0.11755", "0.509" ], [ "0.1184", "0.509", "0.4895", "0.509", "0.4895", "0.5225", "0.1184", "0.5225" ], [ "0.6465", "0.5083", "0.6665", "0.5083", "0.6665", "0.524", "0.6465", "0.524" ], [ "0.11755", "0.524", "0.486", "0.524", "0.486", "0.5376", "0.11755", "0.5376" ], [ "0.546", "0.524", "0.5576", "0.524", "0.5576", "0.535", "0.546", "0.535" ], [ "0.6455", "0.5215", "0.665", "0.5215", "0.665", "0.5376", "0.6455", "0.5376" ], [ "0.1184", "0.5376", "0.451", "0.5376", "0.451", "0.5503", "0.1184", "0.5503" ], [ "0.1184", "0.552", "0.4878", "0.552", "0.4878", "0.5654", "0.1184", "0.5654" ], [ "0.513", "0.5474", "0.885", "0.5474", "0.885", "0.561", "0.513", "0.561" ], [ "0.511", "0.56", "0.7773", "0.5625", "0.7773", "0.5767", "0.511", "0.5747" ], [ "0.1184", "0.5664", "0.4895", "0.5664", "0.4895", "0.58", "0.1184", "0.58" ], [ "0.1184", "0.58", "0.4878", "0.58", "0.4878", "0.5933", "0.1184", "0.5933" ], [ "0.5156", "0.5884", "0.5684", "0.5884", "0.5684", "0.5986", "0.5156", "0.5986" ], [ "0.561", "0.5874", "0.831", "0.5874", "0.831", "0.601", "0.561", "0.601" ], [ "0.11755", "0.594", "0.4878", "0.5933", "0.4878", "0.607", "0.11755", "0.608" ], [ "0.1184", "0.608", "0.4895", "0.608", "0.4895", "0.6216", "0.1184", "0.6216" ], [ "0.511", "0.604", "0.882", "0.605", "0.882", "0.619", "0.511", "0.618" ], [ "0.1184", "0.6226", "0.4895", "0.6226", "0.4895", "0.6357", "0.1184", "0.6357" ], [ "0.513", "0.619", "0.8833", "0.619", "0.8833", "0.633", "0.513", "0.633" ], [ "0.11755", "0.6353", "0.4878", "0.6353", "0.4878", "0.6484", "0.11755", "0.6484" ], [ "0.511", "0.633", "0.882", "0.633", "0.882", "0.6465", "0.511", "0.6465" ], [ "0.1184", "0.652", "0.294", "0.652", "0.294", "0.665", "0.1184", "0.665" ], [ "0.513", "0.648", "0.885", "0.648", "0.885", "0.661", "0.513", "0.661" ], [ "0.11755", "0.663", "0.4895", "0.6646", "0.4895", "0.6787", "0.11755", "0.6777" ], [ "0.511", "0.662", "0.8833", "0.662", "0.8833", "0.676", "0.511", "0.676" ], [ "0.12006", "0.681", "0.2141", "0.681", "0.2141", "0.6924", "0.12006", "0.6924" ], [ "0.513", "0.676", "0.8833", "0.676", "0.8833", "0.6895", "0.513", "0.6895" ], [ "0.1184", "0.6934", "0.461", "0.6934", "0.461", "0.707", "0.1184", "0.707" ], [ "0.513", "0.6904", "0.885", "0.6904", "0.885", "0.7036", "0.513", "0.7036" ], [ "0.1184", "0.707", "0.4895", "0.708", "0.4895", "0.7217", "0.1184", "0.72" ], [ "0.513", "0.705", "0.885", "0.705", "0.885", "0.7188", "0.513", "0.7188" ], [ "0.11755", "0.7217", "0.4895", "0.72", "0.4895", "0.7354", "0.11755", "0.7363" ], [ "0.511", "0.7188", "0.882", "0.7188", "0.882", "0.7324", "0.511", "0.7324" ], [ "0.1184", "0.7363", "0.3208", "0.7363", "0.3208", "0.75", "0.1184", "0.75" ], [ "0.51", "0.7324", "0.885", "0.73", "0.885", "0.7476", "0.51", "0.75" ], [ "0.11755", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.11755", "0.764" ], [ "0.511", "0.7476", "0.8833", "0.7476", "0.8833", "0.761", "0.511", "0.761" ], [ "0.1184", "0.767", "0.4895", "0.767", "0.4895", "0.781", "0.1184", "0.781" ], [ "0.513", "0.7627", "0.8833", "0.7627", "0.8833", "0.7754", "0.513", "0.7754" ], [ "0.11755", "0.78", "0.4895", "0.78", "0.4895", "0.7935", "0.11755", "0.7935" ], [ "0.511", "0.775", "0.882", "0.775", "0.882", "0.788", "0.511", "0.788" ], [ "0.1696", "0.796", "0.4895", "0.796", "0.4895", "0.8086", "0.1696", "0.8086" ], [ "0.5146", "0.7905", "0.8833", "0.7905", "0.8833", "0.804", "0.5146", "0.804" ], [ "0.1184", "0.7964", "0.157", "0.7964", "0.157", "0.808", "0.1184", "0.808" ], [ "0.11584", "0.808", "0.2107", "0.81", "0.2107", "0.8247", "0.11584", "0.8223" ], [ "0.5146", "0.804", "0.796", "0.804", "0.796", "0.818", "0.5146", "0.818" ], [ "0.615", "0.8345", "0.6514", "0.8345", "0.6514", "0.8467", "0.615", "0.8467" ], [ "0.6484", "0.832", "0.7827", "0.8345", "0.7827", "0.8486", "0.6484", "0.8467" ], [ "0.2451", "0.8374", "0.288", "0.8374", "0.288", "0.854", "0.2451", "0.854" ], [ "0.1503", "0.835", "0.1855", "0.835", "0.1855", "0.8516", "0.1503", "0.8516" ], [ "0.6484", "0.8374", "0.6543", "0.8374", "0.6543", "0.843", "0.6484", "0.843" ], [ "0.1511", "0.8516", "0.2024", "0.8516", "0.2024", "0.8677", "0.1511", "0.8677" ], [ "0.2468", "0.8516", "0.2913", "0.8516", "0.2913", "0.8677", "0.2468", "0.8677" ], [ "0.513", "0.8516", "0.882", "0.8525", "0.882", "0.8677", "0.513", "0.866" ], [ "0.1486", "0.866", "0.2083", "0.8643", "0.2091", "0.88", "0.1503", "0.882" ], [ "0.2477", "0.8677", "0.2678", "0.8677", "0.2678", "0.882", "0.2477", "0.882" ], [ "0.5146", "0.8677", "0.882", "0.8677", "0.882", "0.8813", "0.5146", "0.8813" ], [ "0.1503", "0.882", "0.1621", "0.882", "0.1621", "0.8936", "0.1503", "0.8936" ], [ "0.2477", "0.8813", "0.2688", "0.8813", "0.2688", "0.8955", "0.2477", "0.8955" ], [ "0.511", "0.882", "0.882", "0.882", "0.882", "0.8955", "0.511", "0.8955" ], [ "0.513", "0.8965", "0.885", "0.8965", "0.885", "0.91", "0.513", "0.91" ] ]

[ "Table 1: Statistic of various source", "Corpus Source", "#Sentence", "#Entities", "PM Speech text", "4534", "6060", "Assamese play(Wiki)", "403", "595", "9769", "16067", "8798", "Location text", "Person text", "9776", "518", "2465", "Table 2: Dataset Statistics", "Organisation text", "Dataset", "Train", "Dev", "Test", "# sentences", "21475", "767", "1798", "81422", "8292", "8909", "29854", "# tokens", "# entities", "1326", "3783", "Table 3: Frequency distribution of the NER taggset in the dataset", "Token count", "Class Name", "Train", "Dev", "Test", "LOC", "16688", "247", "752", "PER", "10224", "370", "728", "327", "2071", "712", "ORG", "MISC", "1574", "276", "184", "NUM", "656", "106", "48", "0", "51568", "6966", "5126", "CTAGTO", "PER", "441", "aliases.", "ORG includes public organisations (schools, col-", "PER", "frao", "LOC", "QITCA", "leges, universities, charities), companies (Banks,", "stock, markets, company name, brands), govern-", "0", "ment bodies ( ministries, institutions, courts, po", "1", "0", "litical unions of countries such as UNESCO).", "MISC It includes a broad category such as na-", "The first column of the dataset is the word itself", "and the second one is the NER tag.", "tionalities, languages, political ideologies, religions,", "events (conferences, seminar, festivals, book fair,", "4.1.", "Resolving annotation conflict", "expo, sports competitions, forums, parties, con-", "certs) etc. It also categorises words of which one", "We resolve the, annotation conflict manually", "part is a location, person or organisation, or other", "by following semantics and grammatical rules.", "words derived from a word which is a location, or-", "Homonym ambiguity is prevalent in Assamese.", "ganisation, or a person.", "Homonym ambiguity is resolved by looking at the", "NUM includes numbers, money, percentage and", "context of the sentence. e.g. C /zun/ (Moon)", "quantity etc.", "would typically be a MISC, but when it is also of-", "The tag “O\" is used for the remaining tokens.", "ten used in the name of a person in Assamese. In", "Wemeasure the inter-annotator agreementusing", "that case, it would be PER. Similarly, when an", "Cohen's kappa (α). We observed inter-annotator", "institution's name comprises a place name, it can", "agreement on 0.85 using k.", "act both ORG and LOC. e.g. f G /it", "The annotated dataset is prepared in column for-", " guwahati/ is an educational institute which gener-", "mat in which each line represent a word, and each", "ally implies as an organization (ORG); however, in", "column represents one level of linguistic annota-", "some instance, it can be referred to as a location", "An empty line separate sentences in the", "(LOC) also. Therefore, it is a context-dependent", "tion.", "dataset. e.g.", "whether to annotate as ORG or LOC.", "5.", "Experiments", "LOC", "fa", "", "GAGA", "LOC", "We choose a state-of-the-art neural sequence tag-", "", "0", "ger framework based on standard BiLSTM-CRF", "1", "0", "architecture (Yu et al., 2020; Akbik et al., 2019;", "Huang et al., 2015) in our evaluation. It allows" ]

{ "arxiv_id": "2207.03422", "language": "en", "page_idx": 2 }

[ [ "0.1184", "0.076", "0.4902", "0.076", "0.4902", "0.08905", "0.1184", "0.08905" ], [ "0.511", "0.076", "0.885", "0.076", "0.885", "0.08905", "0.511", "0.08905" ], [ "0.1184", "0.0903", "0.4902", "0.0903", "0.4902", "0.10394", "0.1184", "0.10394" ], [ "0.511", "0.0885", "0.854", "0.0885", "0.854", "0.10156", "0.511", "0.10156" ], [ "0.12006", "0.10394", "0.4902", "0.10394", "0.4902", "0.11755", "0.12006", "0.11755" ], [ "0.5146", "0.10394", "0.706", "0.10394", "0.706", "0.11755", "0.5146", "0.11755" ], [ "0.706", "0.10394", "0.818", "0.1057", "0.818", "0.11755", "0.706", "0.1152" ], [ "0.8154", "0.1051", "0.882", "0.1051", "0.882", "0.1164", "0.8154", "0.1164" ], [ "0.12006", "0.11816", "0.4902", "0.11816", "0.4902", "0.1318", "0.12006", "0.1318" ], [ "0.5146", "0.11816", "0.882", "0.11816", "0.882", "0.1318", "0.5146", "0.1318" ], [ "0.12006", "0.1318", "0.4902", "0.1318", "0.4902", "0.1455", "0.12006", "0.1455" ], [ "0.5146", "0.1342", "0.882", "0.1342", "0.882", "0.1443", "0.5146", "0.1443" ], [ "0.12006", "0.1467", "0.4902", "0.1467", "0.4902", "0.1603", "0.12006", "0.1603" ], [ "0.513", "0.1467", "0.885", "0.1467", "0.885", "0.1603", "0.513", "0.1603" ], [ "0.1184", "0.1609", "0.4937", "0.1609", "0.4937", "0.1746", "0.1184", "0.1746" ], [ "0.511", "0.1609", "0.8833", "0.1609", "0.8833", "0.1746", "0.511", "0.1746" ], [ "0.1184", "0.177", "0.2107", "0.177", "0.2107", "0.1882", "0.1184", "0.1882" ], [ "0.513", "0.1746", "0.885", "0.1746", "0.885", "0.1882", "0.513", "0.1882" ], [ "0.513", "0.1895", "0.8833", "0.1895", "0.8833", "0.2031", "0.513", "0.2031" ], [ "0.1696", "0.2037", "0.4114", "0.2037", "0.4114", "0.2185", "0.1696", "0.2185" ], [ "0.513", "0.2037", "0.677", "0.2037", "0.677", "0.2173", "0.513", "0.2173" ], [ "0.1704", "0.2197", "0.2644", "0.2197", "0.2644", "0.231", "0.1704", "0.231" ], [ "0.12006", "0.2048", "0.1628", "0.2048", "0.1628", "0.2162", "0.12006", "0.2162" ], [ "0.11755", "0.2369", "0.4895", "0.2382", "0.4895", "0.253", "0.11755", "0.2517" ], [ "0.513", "0.2322", "0.8833", "0.2322", "0.8833", "0.2452", "0.513", "0.2452" ], [ "0.513", "0.2465", "0.885", "0.2465", "0.885", "0.26", "0.513", "0.26" ], [ "0.12177", "0.2546", "0.4878", "0.2546", "0.4878", "0.2654", "0.12177", "0.2654" ], [ "0.513", "0.2588", "0.885", "0.26", "0.885", "0.275", "0.513", "0.2737" ], [ "0.1184", "0.2673", "0.4902", "0.2673", "0.4902", "0.2808", "0.1184", "0.2808" ], [ "0.11755", "0.2808", "0.4878", "0.2808", "0.4878", "0.2944", "0.11755", "0.2944" ], [ "0.5146", "0.2766", "0.882", "0.2766", "0.882", "0.2874", "0.5146", "0.2874" ], [ "0.513", "0.2893", "0.8833", "0.2893", "0.8833", "0.3027", "0.513", "0.3027" ], [ "0.11755", "0.2957", "0.4895", "0.2957", "0.4895", "0.3088", "0.11755", "0.3088" ], [ "0.511", "0.304", "0.885", "0.3027", "0.885", "0.3164", "0.511", "0.3176" ], [ "0.12006", "0.31", "0.4895", "0.31", "0.4895", "0.3237", "0.12006", "0.3237" ], [ "0.511", "0.3164", "0.8486", "0.3164", "0.8486", "0.33", "0.511", "0.33" ], [ "0.12006", "0.3237", "0.4895", "0.3237", "0.4895", "0.3374", "0.12006", "0.3374" ], [ "0.513", "0.3308", "0.635", "0.3308", "0.635", "0.3445", "0.513", "0.3445" ], [ "0.6265", "0.332", "0.8833", "0.332", "0.8833", "0.3457", "0.6265", "0.3457" ], [ "0.11755", "0.3384", "0.4895", "0.3374", "0.4895", "0.3516", "0.11755", "0.3528" ], [ "0.513", "0.3467", "0.885", "0.3467", "0.885", "0.3599", "0.513", "0.3599" ], [ "0.1184", "0.3528", "0.4878", "0.3528", "0.4878", "0.3665", "0.1184", "0.3665" ], [ "0.513", "0.3599", "0.885", "0.3599", "0.885", "0.3735", "0.513", "0.3735" ], [ "0.1184", "0.3677", "0.4878", "0.3677", "0.4878", "0.3813", "0.1184", "0.3813" ], [ "0.11755", "0.3801", "0.4895", "0.3813", "0.4895", "0.3943", "0.11755", "0.3938" ], [ "0.5146", "0.3748", "0.8193", "0.3748", "0.8193", "0.3884", "0.5146", "0.3884" ], [ "0.513", "0.3872", "0.885", "0.3872", "0.885", "0.401", "0.513", "0.401" ], [ "0.1184", "0.3943", "0.4895", "0.3955", "0.4895", "0.4092", "0.1184", "0.408" ], [ "0.5146", "0.4038", "0.881", "0.4038", "0.881", "0.4146", "0.5146", "0.4146" ], [ "0.1184", "0.4104", "0.4902", "0.4104", "0.4902", "0.4233", "0.1184", "0.4233" ], [ "0.511", "0.4175", "0.8833", "0.4175", "0.8833", "0.4312", "0.511", "0.4312" ], [ "0.12006", "0.4233", "0.4902", "0.4233", "0.4902", "0.437", "0.12006", "0.437" ], [ "0.511", "0.43", "0.8833", "0.4312", "0.8833", "0.4453", "0.511", "0.4448" ], [ "0.1184", "0.4382", "0.4902", "0.4382", "0.4902", "0.452", "0.1184", "0.452" ], [ "0.5146", "0.4453", "0.885", "0.4453", "0.885", "0.459", "0.5146", "0.459" ], [ "0.1184", "0.452", "0.4895", "0.452", "0.4895", "0.4656", "0.1184", "0.4656" ], [ "0.1184", "0.466", "0.4878", "0.466", "0.4878", "0.4797", "0.1184", "0.4797" ], [ "0.5146", "0.4614", "0.615", "0.4614", "0.615", "0.4727", "0.5146", "0.4727" ], [ "0.11755", "0.4797", "0.4895", "0.4797", "0.4895", "0.4934", "0.11755", "0.4934" ], [ "0.513", "0.4744", "0.882", "0.4744", "0.882", "0.488", "0.513", "0.488" ], [ "0.511", "0.4868", "0.8833", "0.488", "0.8833", "0.503", "0.511", "0.502" ], [ "0.11755", "0.4946", "0.486", "0.4946", "0.486", "0.5083", "0.11755", "0.5083" ], [ "0.513", "0.503", "0.8833", "0.503", "0.8833", "0.5166", "0.513", "0.5166" ], [ "0.1184", "0.509", "0.4844", "0.509", "0.4844", "0.5225", "0.1184", "0.5225" ], [ "0.511", "0.5156", "0.8833", "0.5166", "0.8833", "0.531", "0.511", "0.53" ], [ "0.1184", "0.524", "0.4902", "0.524", "0.4902", "0.5376", "0.1184", "0.5376" ], [ "0.5146", "0.531", "0.882", "0.531", "0.882", "0.5444", "0.5146", "0.5444" ], [ "0.1184", "0.5376", "0.2107", "0.539", "0.2107", "0.5503", "0.1184", "0.549" ], [ "0.513", "0.5444", "0.8833", "0.5444", "0.8833", "0.558", "0.513", "0.558" ], [ "0.12006", "0.5527", "0.4895", "0.5527", "0.4895", "0.5664", "0.12006", "0.5664" ], [ "0.513", "0.56", "0.6465", "0.56", "0.6465", "0.5737", "0.513", "0.5737" ], [ "0.1184", "0.568", "0.4895", "0.568", "0.4895", "0.5806", "0.1184", "0.5806" ], [ "0.12006", "0.583", "0.4878", "0.583", "0.4878", "0.5933", "0.12006", "0.5933" ], [ "0.513", "0.585", "0.7256", "0.585", "0.7256", "0.5996", "0.513", "0.5996" ], [ "0.1184", "0.5967", "0.4878", "0.5967", "0.4878", "0.607", "0.1184", "0.607" ], [ "0.5146", "0.6025", "0.885", "0.6025", "0.885", "0.616", "0.5146", "0.616" ], [ "0.1184", "0.61", "0.4902", "0.61", "0.4902", "0.6235", "0.1184", "0.6235" ], [ "0.5156", "0.619", "0.881", "0.619", "0.881", "0.63", "0.5156", "0.63" ], [ "0.12006", "0.626", "0.4878", "0.626", "0.4878", "0.6357", "0.12006", "0.6357" ], [ "0.5146", "0.632", "0.882", "0.632", "0.882", "0.6455", "0.5146", "0.6455" ], [ "0.12006", "0.638", "0.492", "0.638", "0.492", "0.653", "0.12006", "0.653" ], [ "0.5156", "0.648", "0.881", "0.648", "0.881", "0.6577", "0.5156", "0.6577" ], [ "0.12177", "0.654", "0.3628", "0.654", "0.3628", "0.6646", "0.12177", "0.6646" ], [ "0.513", "0.66", "0.8833", "0.66", "0.8833", "0.6733", "0.513", "0.6733" ], [ "0.1184", "0.6665", "0.4895", "0.665", "0.4895", "0.6797", "0.1184", "0.681" ], [ "0.5146", "0.675", "0.885", "0.675", "0.885", "0.6885", "0.5146", "0.6885" ], [ "0.12006", "0.681", "0.4895", "0.681", "0.4895", "0.695", "0.12006", "0.695" ], [ "0.513", "0.6895", "0.885", "0.6895", "0.885", "0.703", "0.513", "0.703" ], [ "0.12177", "0.697", "0.486", "0.697", "0.486", "0.707", "0.12177", "0.707" ], [ "0.513", "0.703", "0.8833", "0.703", "0.8833", "0.7163", "0.513", "0.7163" ], [ "0.11755", "0.709", "0.4878", "0.709", "0.4878", "0.7227", "0.11755", "0.7227" ], [ "0.11755", "0.7227", "0.4878", "0.7227", "0.4878", "0.7363", "0.11755", "0.7363" ], [ "0.513", "0.7173", "0.882", "0.7173", "0.882", "0.731", "0.513", "0.731" ], [ "0.513", "0.7324", "0.8833", "0.7324", "0.8833", "0.745", "0.513", "0.745" ], [ "0.12006", "0.7393", "0.4878", "0.7393", "0.4878", "0.75", "0.12006", "0.75" ], [ "0.12006", "0.752", "0.3064", "0.752", "0.3064", "0.7656", "0.12006", "0.7656" ], [ "0.5146", "0.7476", "0.882", "0.7476", "0.882", "0.7583", "0.5146", "0.7583" ], [ "0.511", "0.759", "0.8833", "0.759", "0.8833", "0.7725", "0.511", "0.7725" ], [ "0.2083", "0.7666", "0.3342", "0.7666", "0.3342", "0.78", "0.2083", "0.78" ], [ "0.346", "0.7666", "0.4902", "0.7666", "0.4902", "0.78", "0.346", "0.78" ], [ "0.511", "0.7725", "0.8833", "0.7725", "0.8833", "0.7876", "0.511", "0.7876" ], [ "0.12006", "0.78", "0.4902", "0.78", "0.4902", "0.7935", "0.12006", "0.7935" ], [ "0.12177", "0.767", "0.215", "0.767", "0.215", "0.7793", "0.12177", "0.7793" ], [ "0.511", "0.7876", "0.76", "0.788", "0.76", "0.8027", "0.511", "0.802" ], [ "0.11755", "0.7944", "0.4902", "0.7944", "0.4902", "0.808", "0.11755", "0.808" ], [ "0.11755", "0.8086", "0.4895", "0.8086", "0.4895", "0.8223", "0.11755", "0.8223" ], [ "0.6406", "0.818", "0.7573", "0.818", "0.7573", "0.832", "0.6406", "0.832" ], [ "0.1184", "0.8237", "0.3896", "0.8237", "0.3896", "0.8374", "0.1184", "0.8374" ], [ "0.5146", "0.8384", "0.8833", "0.8384", "0.8833", "0.8516", "0.5146", "0.8516" ], [ "0.12177", "0.854", "0.2334", "0.854", "0.2334", "0.8643", "0.12177", "0.8643" ], [ "0.251", "0.854", "0.3284", "0.854", "0.3284", "0.8643", "0.251", "0.8643" ], [ "0.3494", "0.854", "0.4895", "0.854", "0.4895", "0.8643", "0.3494", "0.8643" ], [ "0.513", "0.854", "0.881", "0.8525", "0.881", "0.8643", "0.513", "0.865" ], [ "0.1184", "0.866", "0.4902", "0.8677", "0.4902", "0.8813", "0.1184", "0.88" ], [ "0.251", "0.8555", "0.2551", "0.8555", "0.2551", "0.861", "0.251", "0.861" ], [ "0.513", "0.8677", "0.8833", "0.8677", "0.8833", "0.8813", "0.513", "0.8813" ], [ "0.1184", "0.8813", "0.492", "0.8813", "0.492", "0.8955", "0.1184", "0.8955" ], [ "0.511", "0.882", "0.882", "0.882", "0.882", "0.8955", "0.511", "0.8955" ], [ "0.1184", "0.8955", "0.492", "0.8955", "0.492", "0.909", "0.1184", "0.909" ], [ "0.513", "0.8955", "0.885", "0.8955", "0.885", "0.909", "0.513", "0.909" ] ]

[ "to apply various state-of-the-art language models", "input strings are randomly masked and tries to", "to train sequence tagging models such as Named", " predict the masked word based on its context.", "Entity Recognition (NER), and POS tagging, es-", "Embeddingsfrom", "Language", "Models", "pecially for high resource languages such as En-", "(ELMo) (Peters et al., 2018) word represen-", "glish, German and Dutch (Akbik et al., 2018; Pe-", "tationmodels complexfeatures of words usessuch", "ters et al., 2018). Using the architecture, we eval-", "as syntax and semantics. An embedding vector", "uate our dataset using different pre-trained word", "of a word varies according to the context of the", "embeddings.", "sentences (i.e. to model polysemy). ELMo word", "vector representations are calculated based on the", "Word Embeddings used in", "entire input sentence.", "evaluation", "5.1.", "Word embedding is a crucial component in ma-", "Flair Embeddings (Akbik et al., 2018) is a con-", "textual string embeddings for any string of char-", "chinelearning-basedNLP models.Thereal-valued", "acters in a sentential context.The embedding", "vector representation of words has the ability to", "capture both semantic and syntactic meanings", "method based on recent advances inneural lan-", "guage modelling (LM) (Sutskever et al., 2014) that", "of words in a sentence, which led to significant", "provides languages to be modelled as distributions", "advances in recent natural language processing", " over sequences of characters instead of words.", "(NLP) tasks such as sequence classification, POS", "Multilingual", "Representations for Indian", "tagging, Named Entity Recognition (NER), Senti-", "Languages (MuRIL) (Khanuja et al., 2021) is", "ment Analysis, Machine Translation (MT), Ques-", "based on BERT base architecture pre-trained on", "tion Answering (QA). There are various word em-", "beddings to embed the words in sentences in mul-", "17 Indian languages, including Assamese.", "XLM-R (Conneau et al., 2019) uses self-", "tiple ways.. These embeddings are usually trained", "supervised training techniques in cross-lingual un-", "and evaluated on high-resource languages, using", "derstanding. In cross-lingual understanding, a", "a large collection of unlabeled corpora to build", "model is trained for a task in one language and", "feature-rich word embeddings.Eventually, it be-", "then used with other languages without additional", "comes an integral part of neural models in NLP", "applications and is found to be achieved state-of-", "training data.", "the-art performance in the downstream task. We", "Most of these word embeddings are trained on As-", " samese Wikipedia dataset (approx 8 million words)", "used eight different pre-trained word embeddings", "which size is limited for quality training. So, the", "in our experiment of developing Assamese NER", "performance in NER tagging is lower than most", "In the next part, we briefly describe all these word", "recently proposed MuRIL (Khanuja et al., 2021)", "embeddings.", "which are trained on comparatively large size of", "Word2Vec (Mikolov et al., 2013; Kakwani et", "training datasets.", "al., 2020): It is capable of capturing the context", "of a word in a sentence, semantic and syntactic", "5.2.Hyperparameters", "similarity between the words.In Word2vec, word", "We train our models on Nvidia Tesla P100 403", "embeddings can be obtained by utilizing either", "GPU (3584 Cuda Cores).We choose similar hy-", "of the two architectures,continuous bag-of-words", "perparameters in all the eight experiments in ta-", "(CBOW) or continuous skip-gram. We use the", "ble4toevaluatetheperformanceofdifferentpre-", "CBOW model in our experiment", "trained Assamese word embeddings over the As-", "FastText embedding (Bojanowski et al., 2017)", "NER dataset. We kept the hidden layer size of 1024", "are belongs to the type of sub-word embeddings", "with six hidden layers and a maximum sequence", "where it is trained on character n-grams of words", "length of 128. To account for memory constraints,", "rather than whole words. FastText Embeddings", "can give word vectors for out of vocabulary(OOV)", "we use a mini-batch size of 32. Apart from hid-", "den size, hidden layer, sequence length and batch", "words by using the sub-word informationfrom the", "previously trained model.", "size, we use the default values for the remaining", "hyperparameters as in (Akbik et al., 2019).We", "Embeddings", "(Heinzerlingand", "train the model for 50 epochs. We use learning", "Strube, 2018) are pre-computed on sub-word", "Byte-Pair", "rate annealing for early stopping.", "level. They can embed by splitting words into", "subwords or character sequences, looking up the", "6. Results", "pre-computed subwords embeddings.", "We present the performance results of different ex-", "Bidirectional", "Encoder", "Representations", "periments using our annotated dataset in this sec.", "from Transformers (BERT) (Devlin et al.,", "", "tion. To conduct the experiments, we configure", "2018) uses masked language models (MLM). In", "eight setups using different pre-trained word em-", "masked language model, one or more words of the", "bedding models discussed in section 5.1. We use" ]

{ "arxiv_id": "2207.03422", "language": "en", "page_idx": 3 }

[ [ "0.1748", "0.03284", "0.4387", "0.03284", "0.4387", "0.04672", "0.1748", "0.04672" ], [ "0.1772", "0.1016", "0.822", "0.1042", "0.822", "0.1187", "0.1772", "0.11615" ], [ "0.1748", "0.127", "0.636", "0.1276", "0.636", "0.1459", "0.1748", "0.1445" ], [ "0.1854", "0.1704", "0.5522", "0.1704", "0.5522", "0.1843", "0.1854", "0.1843" ], [ "0.1854", "0.1849", "0.701", "0.1849", "0.701", "0.1989", "0.1854", "0.1989" ], [ "0.1854", "0.202", "0.5317", "0.202", "0.5317", "0.2122", "0.1854", "0.2122" ], [ "0.4534", "0.25", "0.546", "0.25", "0.546", "0.2644", "0.4534", "0.2644" ], [ "0.2336", "0.2822", "0.7646", "0.2822", "0.7646", "0.2954", "0.2336", "0.2954" ], [ "0.0237", "0.289", "0.058", "0.289", "0.058", "0.414", "0.0237", "0.414" ], [ "0.2361", "0.2974", "0.763", "0.2974", "0.763", "0.308", "0.2361", "0.308" ], [ "0.2336", "0.3093", "0.7656", "0.3093", "0.7656", "0.3225", "0.2336", "0.3225" ], [ "0.232", "0.3225", "0.763", "0.324", "0.763", "0.3372", "0.232", "0.336" ], [ "0.2345", "0.3372", "0.761", "0.3384", "0.761", "0.3499", "0.2345", "0.3484" ], [ "0.2336", "0.351", "0.767", "0.351", "0.767", "0.3643", "0.2336", "0.3643" ], [ "0.232", "0.3643", "0.706", "0.3655", "0.706", "0.379", "0.232", "0.3774" ], [ "0.02124", "0.4028", "0.05637", "0.4016", "0.05963", "0.4817", "0.0237", "0.483" ], [ "0.1748", "0.3977", "0.824", "0.3977", "0.824", "0.411", "0.1748", "0.411" ], [ "0.1748", "0.4124", "0.824", "0.4124", "0.824", "0.4255", "0.1748", "0.4255" ], [ "0.1732", "0.4248", "0.824", "0.4255", "0.824", "0.44", "0.1732", "0.4395" ], [ "0.1748", "0.4395", "0.4224", "0.4395", "0.4224", "0.4526", "0.1748", "0.4526" ], [ "0.3635", "0.4602", "0.4207", "0.4622", "0.4192", "0.4766", "0.3628", "0.4749" ], [ "0.4902", "0.4622", "0.513", "0.4622", "0.513", "0.4734", "0.4902", "0.4734" ], [ "0.5728", "0.4634", "0.634", "0.4634", "0.634", "0.4749", "0.5728", "0.4749" ], [ "0.8022", "0.4622", "0.8267", "0.4622", "0.8267", "0.476", "0.8022", "0.476" ], [ "0.4265", "0.4702", "0.5728", "0.4697", "0.5728", "0.488", "0.4265", "0.4893" ], [ "0.02287", "0.4873", "0.05963", "0.4873", "0.05963", "0.704", "0.02287", "0.704" ], [ "0.1748", "0.4956", "0.826", "0.4956", "0.826", "0.509", "0.1748", "0.509" ], [ "0.1732", "0.5083", "0.523", "0.509", "0.523", "0.5234", "0.1732", "0.523" ], [ "0.1765", "0.53", "0.652", "0.53", "0.652", "0.5444", "0.1765", "0.5444" ], [ "0.3057", "0.553", "0.3882", "0.553", "0.3882", "0.5674", "0.3057", "0.5674" ], [ "0.5312", "0.5537", "0.694", "0.5537", "0.694", "0.5684", "0.5312", "0.5684" ], [ "0.8022", "0.553", "0.826", "0.553", "0.826", "0.5674", "0.8022", "0.5674" ], [ "0.3628", "0.561", "0.5312", "0.56", "0.5312", "0.5776", "0.3628", "0.579" ], [ "0.452", "0.556", "0.468", "0.556", "0.468", "0.5645", "0.452", "0.5645" ], [ "0.1732", "0.5854", "0.824", "0.586", "0.824", "0.601", "0.1732", "0.5996" ], [ "0.1732", "0.601", "0.3218", "0.601", "0.3218", "0.6143", "0.1732", "0.6143" ], [ "0.3284", "0.6226", "0.6733", "0.6226", "0.6733", "0.637", "0.3284", "0.637" ], [ "0.8022", "0.6235", "0.826", "0.6235", "0.826", "0.637", "0.8022", "0.637" ], [ "0.327", "0.6333", "0.3816", "0.6353", "0.3816", "0.646", "0.327", "0.643" ], [ "0.58", "0.638", "0.5923", "0.638", "0.5923", "0.6416", "0.58", "0.6416" ], [ "0.1748", "0.655", "0.8228", "0.655", "0.8228", "0.6685", "0.1748", "0.6685" ], [ "0.1724", "0.6675", "0.826", "0.6665", "0.826", "0.6846", "0.1724", "0.685" ], [ "0.1765", "0.685", "0.6216", "0.685", "0.6216", "0.696", "0.1765", "0.696" ], [ "0.1765", "0.704", "0.8228", "0.704", "0.8228", "0.7173", "0.1765", "0.7173" ], [ "0.1748", "0.7183", "0.6323", "0.7183", "0.6323", "0.732", "0.1748", "0.732" ], [ "0.3652", "0.7407", "0.4534", "0.7407", "0.4534", "0.755", "0.3652", "0.755" ], [ "0.451", "0.7407", "0.6313", "0.7407", "0.6313", "0.755", "0.451", "0.755" ], [ "0.8022", "0.7407", "0.826", "0.7407", "0.826", "0.755", "0.8022", "0.755" ], [ "0.452", "0.7524", "0.5073", "0.7524", "0.5073", "0.763", "0.452", "0.763" ], [ "0.1732", "0.768", "0.824", "0.771", "0.824", "0.785", "0.1732", "0.7827" ], [ "0.1732", "0.785", "0.6567", "0.785", "0.6567", "0.799", "0.1732", "0.799" ], [ "0.407", "0.8066", "0.6763", "0.8066", "0.6763", "0.8213", "0.407", "0.8213" ], [ "0.8022", "0.8066", "0.826", "0.8066", "0.826", "0.8213", "0.8022", "0.8213" ], [ "0.4028", "0.8193", "0.4592", "0.8193", "0.4592", "0.83", "0.4028", "0.83" ], [ "0.3218", "0.8096", "0.4028", "0.8096", "0.4028", "0.821", "0.3218", "0.821" ], [ "0.1748", "0.837", "0.8228", "0.837", "0.8228", "0.8516", "0.1748", "0.8516" ], [ "0.1748", "0.8516", "0.822", "0.8516", "0.822", "0.8657", "0.1748", "0.8657" ], [ "0.1748", "0.8726", "0.8228", "0.8726", "0.8228", "0.886", "0.1748", "0.886" ], [ "0.1748", "0.887", "0.824", "0.887", "0.824", "0.901", "0.1748", "0.901" ], [ "0.1978", "0.911", "0.4885", "0.911", "0.4885", "0.925", "0.1978", "0.925" ], [ "0.4983", "0.9478", "0.509", "0.952", "0.5024", "0.961", "0.492", "0.957" ] ]

[ "Published as a Tiny Paper at ICLR 2023", "ON A RELATION BETWEEN THE RATE-DISTORTION", "FUNCTION AND OPTIMAL TRANSPORT", "Eric Lei, Hamed Hassani and Shirin Saeedi Bidokhti", "Dept. of Electrical and Systems Engineering, University of Pennsylvania, USA", "elei,hassani,saeedi@seas.upenn.edu", "ABSTRACT", "We discuss a relationship between rate-distortion and optimal transport (OT) the-", "", "ory,even though they seem to be unrelated atfirstglance.In particular,we show", "that a function defined via an extremal entropic OT distance is equivalent to the", "rate-distortion function.We numerically verify this result as well as previous re-", "sults that connect theMongeand Kantorovichproblems to optimal scalarquanti", "zation. Thus, we unify solving scalar quantization and rate-distortion functions in", "an alternative fashion by using their respective optimal transport solvers.", "[LI's]", "Rate-Distortion. Let X ~ Px be the source supported on X. Let be the reproduction space, and", "p : X × → R>o be a distortion measure. The asymptotic limit on the minimum number of bits", "required to represent X with average distortion at most D is given by the rate-distortion function", "(Cover & Thomas (2006)), defined as", "R(D) :=", "inf", "I(X;Y)", "(1)", "Pyix :Px,yle(X,Y)<≤D", "0", "Any rate-distortion pair (R, D) satisfying R > R(D) is achievable by some lossy source code, and", "no code can achieve a rate-distortion less than R(D).", "R(D) has the following alternate form (Cover & Thomas, 2006, Ch. 10),", "R(D) = inf", "DkL(Px,yIPx Qy).", "(2)", "Qy PyIx:EPx.le(X,Y)≤D", "int", "Dueto the convex and strictly decreasing properties (Cover & Thomas (2006)) of R(D),it suffices", "to fix 入> O and solve", "inf inf DkL(Px.Y|/Px Qy) +> E [e(X,Y).", "(3)", "QyPyix", "", "A solution to (3) corresponds to a point on R(D) corresponding to X. The Blahut-Arimoto (BA)", "algorithm (Blahut (1972); Arimoto (1972) solves (2) by alternating steps on Pyix and Qy until", "convergence.Sweeping overgivestheentirerate-distortion curve.", "Optimal Transport.We consider optimal transport (OT) under theKantorovich formulation,which", "finds the minimum distortion coupling between measures μ and v,", "W(μ,v) :=", "infEx,Y~[p(X,Y)].", "(4)", "TEII(μ,V)", "Under certain conditions, the optimal coupling is induced by a fixed mapping, known as the Monge", "map. The Kantorovich problem is often regularized with an entropy term,", "infE[p(X,Y)I +eDkL(lμ@v),", "(5)", "TEII(μ,V)", "Se(μ,v) :=", "which is known as entropy-regularized optimal transport, with e > 0. For discrete measures μ, v, (5)", "can be solved efficiently using Sinkhorn's algorithm (Knopp & Sinkhorn (1967); Sinkhorn (1964).", "Related Work. A connection between source coding and optimal transport was made in a talk", "given by Gray (2013), who discusses how scalar quantizers can be found through an extremal", "'A joint distribution that marginalizes to μ and v.", "1" ]

{ "arxiv_id": "2307.00246", "language": "en", "page_idx": 0 }

[ [ "0.1748", "0.03284", "0.4387", "0.03284", "0.4387", "0.04672", "0.1748", "0.04672" ], [ "0.2059", "0.1042", "0.2239", "0.1042", "0.2239", "0.11237", "0.2059", "0.11237" ], [ "0.7305", "0.1111", "0.7656", "0.1111", "0.7656", "0.1225", "0.7305", "0.1225" ], [ "0.2051", "0.1288", "0.2255", "0.1288", "0.2255", "0.1414", "0.2051", "0.1414" ], [ "0.7344", "0.1256", "0.7974", "0.1256", "0.7974", "0.137", "0.7344", "0.137" ], [ "0.524", "0.132", "0.55", "0.1307", "0.5522", "0.1818", "0.527", "0.1831" ], [ "0.4224", "0.1414", "0.4575", "0.1414", "0.4575", "0.1572", "0.4224", "0.1572" ], [ "0.1896", "0.1572", "0.2239", "0.1547", "0.2255", "0.1674", "0.1912", "0.1692" ], [ "0.4224", "0.1559", "0.4575", "0.1559", "0.4575", "0.1718", "0.4224", "0.1718" ], [ "0.406", "0.1755", "0.423", "0.1755", "0.423", "0.1799", "0.406", "0.1799" ], [ "0.4207", "0.1718", "0.46", "0.1718", "0.46", "0.1875", "0.4207", "0.1875" ], [ "0.2035", "0.1843", "0.2239", "0.1843", "0.2239", "0.1963", "0.2035", "0.1963" ], [ "0.524", "0.1849", "0.536", "0.1755", "0.5474", "0.1843", "0.535", "0.1932" ], [ "0.2035", "0.2122", "0.2263", "0.2122", "0.2263", "0.2235", "0.2035", "0.2235" ], [ "0.546", "0.2102", "0.5596", "0.214", "0.5537", "0.2235", "0.5415", "0.2197" ], [ "0.2239", "0.2235", "0.255", "0.2235", "0.255", "0.2349", "0.2239", "0.2349" ], [ "0.3218", "0.2235", "0.3513", "0.2235", "0.3513", "0.2349", "0.3218", "0.2349" ], [ "0.3708", "0.2235", "0.4004", "0.2235", "0.4004", "0.2349", "0.3708", "0.2349" ], [ "0.4192", "0.2235", "0.4495", "0.2235", "0.4495", "0.2349", "0.4192", "0.2349" ], [ "0.55", "0.226", "0.5605", "0.2203", "0.568", "0.2299", "0.556", "0.2349" ], [ "0.675", "0.2343", "0.6895", "0.2343", "0.6895", "0.2443", "0.675", "0.2443" ], [ "0.2737", "0.226", "0.2998", "0.226", "0.2998", "0.2343", "0.2737", "0.2343" ], [ "0.3406", "0.2361", "0.3545", "0.2361", "0.3545", "0.2443", "0.3406", "0.2443" ], [ "0.1748", "0.2583", "0.492", "0.2583", "0.492", "0.2715", "0.1748", "0.2715" ], [ "0.509", "0.2583", "0.8228", "0.2583", "0.8228", "0.2715", "0.509", "0.2715" ], [ "0.1732", "0.2727", "0.4927", "0.2727", "0.4927", "0.2874", "0.1732", "0.2874" ], [ "0.5063", "0.2727", "0.7983", "0.2727", "0.7983", "0.286", "0.5063", "0.286" ], [ "0.1748", "0.315", "0.822", "0.315", "0.822", "0.3289", "0.1748", "0.3289" ], [ "0.1748", "0.33", "0.8228", "0.33", "0.8228", "0.3435", "0.1748", "0.3435" ], [ "0.1748", "0.3435", "0.824", "0.3435", "0.824", "0.3567", "0.1748", "0.3567" ], [ "0.1748", "0.356", "0.822", "0.356", "0.822", "0.3694", "0.1748", "0.3694" ], [ "0.1732", "0.3718", "0.824", "0.3718", "0.824", "0.385", "0.1732", "0.385" ], [ "0.1748", "0.3838", "0.774", "0.3838", "0.774", "0.3977", "0.1748", "0.3977" ], [ "0.1748", "0.406", "0.826", "0.406", "0.826", "0.4192", "0.1748", "0.4192" ], [ "0.1748", "0.4204", "0.7876", "0.4204", "0.7876", "0.4343", "0.1748", "0.4343" ], [ "0.3447", "0.443", "0.4395", "0.443", "0.4395", "0.4583", "0.3447", "0.4583" ], [ "0.4468", "0.443", "0.4927", "0.443", "0.4927", "0.455", "0.4468", "0.455" ], [ "0.513", "0.4443", "0.6543", "0.4443", "0.6543", "0.459", "0.513", "0.459" ], [ "0.8022", "0.4443", "0.826", "0.4443", "0.826", "0.4583", "0.8022", "0.4583" ], [ "0.4043", "0.4539", "0.5215", "0.4539", "0.5215", "0.4685", "0.4043", "0.4685" ], [ "0.4307", "0.4666", "0.5186", "0.4666", "0.5186", "0.4766", "0.4307", "0.4766" ], [ "0.1732", "0.486", "0.819", "0.486", "0.819", "0.5", "0.1732", "0.5" ], [ "0.1748", "0.5005", "0.8228", "0.5005", "0.8228", "0.5146", "0.1748", "0.5146" ], [ "0.1748", "0.5146", "0.824", "0.5146", "0.824", "0.528", "0.1748", "0.528" ], [ "0.1732", "0.5264", "0.824", "0.529", "0.824", "0.5435", "0.1732", "0.5415" ], [ "0.1765", "0.5425", "0.384", "0.5425", "0.384", "0.556", "0.1765", "0.556" ], [ "0.1748", "0.56", "0.7534", "0.56", "0.7534", "0.5747", "0.1748", "0.5747" ], [ "0.4478", "0.586", "0.55", "0.586", "0.55", "0.5977", "0.4478", "0.5977" ], [ "0.8022", "0.586", "0.826", "0.586", "0.826", "0.5996", "0.8022", "0.5996" ], [ "0.1732", "0.616", "0.824", "0.616", "0.824", "0.631", "0.1732", "0.631" ], [ "0.1748", "0.631", "0.824", "0.631", "0.824", "0.6445", "0.1748", "0.6445" ], [ "0.1748", "0.646", "0.824", "0.646", "0.824", "0.659", "0.1748", "0.659" ], [ "0.1732", "0.659", "0.687", "0.659", "0.687", "0.6724", "0.1732", "0.6724" ], [ "0.1748", "0.6797", "0.822", "0.6797", "0.822", "0.6943", "0.1748", "0.6943" ], [ "0.1748", "0.6943", "0.824", "0.6943", "0.824", "0.7085", "0.1748", "0.7085" ], [ "0.1748", "0.707", "0.822", "0.707", "0.822", "0.7217", "0.1748", "0.7217" ], [ "0.1732", "0.72", "0.8228", "0.7227", "0.8228", "0.7373", "0.1732", "0.735" ], [ "0.1765", "0.7373", "0.8228", "0.7373", "0.8228", "0.7476", "0.1765", "0.7476" ], [ "0.1732", "0.7485", "0.824", "0.75", "0.824", "0.7646", "0.1732", "0.763" ], [ "0.1748", "0.763", "0.8228", "0.763", "0.8228", "0.7764", "0.1748", "0.7764" ], [ "0.1748", "0.7764", "0.824", "0.7764", "0.824", "0.7915", "0.1748", "0.7915" ], [ "0.1748", "0.7915", "0.8228", "0.7915", "0.8228", "0.8047", "0.1748", "0.8047" ], [ "0.1732", "0.8047", "0.826", "0.8047", "0.826", "0.8193", "0.1732", "0.8193" ], [ "0.1748", "0.8193", "0.2646", "0.8193", "0.2646", "0.8306", "0.1748", "0.8306" ], [ "0.1748", "0.839", "0.824", "0.839", "0.824", "0.854", "0.1748", "0.854" ], [ "0.1708", "0.8516", "0.8286", "0.8486", "0.8286", "0.869", "0.1708", "0.8726" ], [ "0.1732", "0.867", "0.824", "0.867", "0.824", "0.882", "0.1732", "0.882" ], [ "0.1732", "0.881", "0.8267", "0.881", "0.8267", "0.8984", "0.1732", "0.8984" ], [ "0.1732", "0.896", "0.826", "0.8965", "0.826", "0.911", "0.1732", "0.91" ], [ "0.1732", "0.911", "0.6406", "0.911", "0.6406", "0.925", "0.1732", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a Tiny Paper at ICLR 2023", "2.0", "EMD", "1.5", "Lloyd-Max", "(w)0=", "S(D)", "R 1.0", "R(D)", "-", "H(X)", "0.5", "R", "0.0", "0", "0.00", "0.50", "0.75", "1.00", "0", "D", "0.25", "D", "Figure l: Equivalence of S(D) and R(D) on a", "Figure2:Equivalenceof extremalEMDand", "5-atom discrete source with p(c, y) = (α -y)?", "Lloyd-Max for M-level scalar quantization.", "Monge/Kantorovich problem, and alludes to a similar connection for Shannon's rate-distortion func-", "tion. Here, we concretely provide R(D)'s connection with entropic OT and discuss how their respec-", "tive computational methods (Blahut-Arimoto and Sinkhorn-Knopp) can compute R(D). In a similar", "vein, we empirically verify Gray (2013)'s results and show that Lloyd-Max and Earth Mover's dis-", "tance can both compute optimal scalar quantizers.A similar result relating rate-distortion with", "entropic OT was also reported in Wu et al. (2022) which was unbeknownst to us at the time.", "Main Result. We first show that entropic OT can be used to upper bound R(D). First, observe that", "the inner minimization problem in (3) looks similar to the entropic OT problem. Let us define", "S(D) := inf ", ".inf", "DKL(元Px@QY),", "(6)", "QY Tell(Px.Qy):", "E[p(X,Y)]≤D", "which we call the Sinkhorn-distortion function, and is an extremal entropic OT distance w.r.t. Px", "Similar to R(D), we can trace out S(D) by sweeping over > O, and solving the inner minimization", "(5), and then optimizing over all Qy, which is a convex problem in Qy (Feydy et al. (2019)). It is", "clear that R(D) ≤ S(D) by comparing (6) and (2). Next, we show that without further assumptions,", "R(D) and S(D) are equivalent.", "Theorem 1. For any source Px and distortion function p : X × y-→ R>o, it holds that", "R(D) = S(D)", "(7)", "See Sec. A.1 for the proof. We numerically verify the equivalence in Fig. 1 on a discrete source", "with 5 atoms under squared-error distortion. For R(D), we use Blahut-Arimoto, and for S(D), we", "solve the convex problem using SQP solvers (Kraft (1988)) with Q → S.(Px, Q) as the objective", "function, showing that the two different objectives result in the same function.", "Discussion. Observe that the joint Px,Y = Px Pyix defined in (2) marginalizes to Px but not nec-", "essarily Qy, whereas the coupling in (6) marginalizes to both. This result says that the additional", "Qy marginalization constraint in S(D) plays no role when both objectives are infimized over Qy", "In computing R(D), this provides an alternative to Blahut-Arimoto: solve (6) directly over Qy,", "using Sinkhorniterationsasa subroutinewhen evaluatingthe objectivefunction(orits gradient).A", "symmetrized variant of the Sinkhorn-distortion function is often used to solve generative modeling", "tasks with Sinkhorn divergences (Genevay et al. (2018); Salimans et al. (2018); Shen et al. (2020)),", "where one wishes to find some Qy ~ Px by solving minQy S.(Px, Qy). However, if one leaves", "the objective un-symmetrized, the optimal Qy and coupling π* are actually R(D)-achieving distri-", "butions with A = 1/e, producing equivalent solutions to exact rate-distortion neural estimators (Lei", "et al. (2022)).", "We also verify that in discretesettings,the extremal non-entropic oT function", "minQy;|Qrl<M W(Px,Qy), where |Qyl is the size of Qy's alphabet, is equivalent to", "optimal scalar quantization of Px as shown in Gray (2013).In Fig. 2, we solve the", "minQy:lQyl<M W(Px,Qy) on a 10-atom source using a linear program to compute the", "Earth Mover's distance (EMD) W(,) and pass the function to a SQP solver as before. The", "achieved rate-distortion is equivalent to that of Lloyd-Max (M-means).", "2" ]

{ "arxiv_id": "2307.00246", "language": "en", "page_idx": 1 }

[ [ "0.1748", "0.03284", "0.4387", "0.03284", "0.4387", "0.04736", "0.1748", "0.04736" ], [ "0.1748", "0.1042", "0.304", "0.1048", "0.304", "0.1193", "0.1748", "0.1187" ], [ "0.1732", "0.1288", "0.621", "0.1276", "0.621", "0.1427", "0.1732", "0.1433" ], [ "0.1748", "0.1674", "0.2876", "0.1674", "0.2876", "0.1818", "0.1748", "0.1818" ], [ "0.1732", "0.1913", "0.822", "0.1913", "0.822", "0.2058", "0.1732", "0.2058" ], [ "0.1896", "0.2058", "0.822", "0.2058", "0.822", "0.2197", "0.1896", "0.2197" ], [ "0.1936", "0.2203", "0.255", "0.2203", "0.255", "0.2317", "0.1936", "0.2317" ], [ "0.1732", "0.2443", "0.8228", "0.2456", "0.8228", "0.26", "0.1732", "0.2588" ], [ "0.1912", "0.26", "0.709", "0.26", "0.709", "0.2734", "0.1912", "0.2734" ], [ "0.1765", "0.2874", "0.8203", "0.2874", "0.8203", "0.3005", "0.1765", "0.3005" ], [ "0.1912", "0.3018", "0.7793", "0.3018", "0.7793", "0.315", "0.1912", "0.315" ], [ "0.1748", "0.3276", "0.8228", "0.3276", "0.8228", "0.3423", "0.1748", "0.3423" ], [ "0.1912", "0.3423", "0.8203", "0.3423", "0.8203", "0.356", "0.1912", "0.356" ], [ "0.1896", "0.3547", "0.8203", "0.3547", "0.8203", "0.3694", "0.1896", "0.3694" ], [ "0.1912", "0.3694", "0.282", "0.3694", "0.282", "0.3838", "0.1912", "0.3838" ], [ "0.1748", "0.3965", "0.822", "0.3965", "0.822", "0.411", "0.1748", "0.411" ], [ "0.1936", "0.4136", "0.822", "0.4136", "0.822", "0.4236", "0.1936", "0.4236" ], [ "0.1912", "0.4248", "0.826", "0.4255", "0.826", "0.4395", "0.1912", "0.4382" ], [ "0.1896", "0.4382", "0.6562", "0.4382", "0.6562", "0.4526", "0.1896", "0.4526" ], [ "0.1732", "0.4653", "0.8228", "0.4653", "0.8228", "0.4792", "0.1732", "0.4792" ], [ "0.1896", "0.4797", "0.822", "0.4797", "0.822", "0.4937", "0.1896", "0.4937" ], [ "0.1969", "0.4944", "0.3545", "0.4944", "0.3545", "0.5083", "0.1969", "0.5083" ], [ "0.1748", "0.521", "0.819", "0.521", "0.819", "0.534", "0.1748", "0.534" ], [ "0.1896", "0.534", "0.8228", "0.534", "0.8228", "0.549", "0.1896", "0.549" ], [ "0.1896", "0.549", "0.357", "0.549", "0.357", "0.5625", "0.1896", "0.5625" ], [ "0.1748", "0.5747", "0.822", "0.5747", "0.822", "0.589", "0.1748", "0.589" ], [ "0.1929", "0.589", "0.822", "0.589", "0.822", "0.603", "0.1929", "0.603" ], [ "0.1896", "0.6016", "0.7876", "0.603", "0.7876", "0.617", "0.1896", "0.616" ], [ "0.1732", "0.629", "0.822", "0.6274", "0.822", "0.6426", "0.1732", "0.643" ], [ "0.1912", "0.643", "0.822", "0.643", "0.822", "0.657", "0.1912", "0.657" ], [ "0.1887", "0.657", "0.6855", "0.656", "0.6855", "0.6704", "0.1887", "0.672" ], [ "0.1748", "0.6846", "0.822", "0.6846", "0.822", "0.6978", "0.1748", "0.6978" ], [ "0.1912", "0.6987", "0.822", "0.6987", "0.822", "0.7134", "0.1912", "0.7134" ], [ "0.1912", "0.712", "0.7656", "0.712", "0.7656", "0.726", "0.1912", "0.726" ], [ "0.1748", "0.7393", "0.8228", "0.7393", "0.8228", "0.754", "0.1748", "0.754" ], [ "0.1887", "0.7524", "0.524", "0.752", "0.524", "0.7666", "0.1887", "0.7676" ], [ "0.1748", "0.78", "0.822", "0.78", "0.822", "0.7944", "0.1748", "0.7944" ], [ "0.1887", "0.7935", "0.8096", "0.7925", "0.8096", "0.8066", "0.1887", "0.808" ], [ "0.1748", "0.821", "0.822", "0.821", "0.822", "0.834", "0.1748", "0.834" ], [ "0.1896", "0.8354", "0.822", "0.8354", "0.822", "0.8486", "0.1896", "0.8486" ], [ "0.1896", "0.8477", "0.71", "0.8486", "0.71", "0.863", "0.1896", "0.8623" ], [ "0.1748", "0.875", "0.822", "0.875", "0.822", "0.8896", "0.1748", "0.8896" ], [ "0.1912", "0.89", "0.824", "0.89", "0.824", "0.904", "0.1912", "0.904" ], [ "0.1912", "0.904", "0.357", "0.904", "0.357", "0.9175", "0.1912", "0.9175" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a Tiny Paper at ICLR 2023", "URM STATEMENT", "All authors meet the URM criteria of ICLR 2023 Tiny Papers Track.", "REFERENCES", "Suguru Arimoto. An algorithm for computing the capacity of arbitrary discrete memoryless chan-", "nels. IEEE Transactions on Information Theory, 18(1):14-20, 1972. doi: 10.1109/TIT.1972.", "1054753.", "Richard Blahut. Computation of channel capacity and rate-distortion functions. IEEE Transactions", "on Information Theory, 18(4):460-473, 1972. doi: 10.1109/TIT.1972.1054855.", "Thomas M. Cover and Joy A. Thomas. Elements of Information Theory (Wiley Series in Telecom", "munications and Signal Processing). Wiley-Interscience, USA, 2006. ISBN 0471241954.", "Jean Feydy, Thibault Sejourné, Francois-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and", "Gabriel Peyre. Interpolating between optimal transport and mmd using sinkhorn divergences", "In The 22nd International Conference on Artificial Intelligence and Statistics, pp.2681-2690", "PMLR, 2019.", "Aude Genevay, Gabriel Peyre, and Marco Cuturi. Learning generative models with sinkhorn di-", "vergences.InAmos StorkeyandFernandoPerez-Cruz(eds.),Proceedingsofthe Twenty-Firs", "International Conference onArtificial Intelligence and Statistics,volume84of Proceedingsof", "Machine Learning Research, Pp. 1608-1617. PMLR, 09-11 Apr 2018.", "Robert M. Gray. Transportation distance, shannon information, and source coding. GRETSI 2013", "Symposium on Signal and Image Processing, 2013. URL https://ee.stanford.edu/", "gray/gretsi.pdf.", "Paul Knopp and Richard Sinkhorn. Concerning nonnegative matrices and doubly stochastic ma", "trices. Pacific Journal of Mathematics, 21(2):343 - 348, 1967. doi: pjm/1102992505. URL", "https://doi.org/.", "D. Kraft. A Software Package for Sequential Quadratic Programming. Deutsche Forschungs-", "und Versuchsanstalt fur Luft- und Raumfahrt Koln: Forschungsbericht. Wiss. Berichtswesen d.", "DFVLR,1988.URL https://books.google.com/books?id=4rKaGwAACAAJ.", "Eric Lei, Hamed Hassani, and Shirin Saeedi Bidokhti. Neural estimation of the rate-distortion", "function with applications to operational source coding. IEEE Journal on Selected Areas in In-", "formation Theory,3(4):674-686,2022.doi:10.1109/JSAIT.2023.3273467.", "Gabriel Peyré and Marco Cuturi. Computational optimal transport: With applications to data sci-", "ence. Foundations and Trends? in Machine Learning, 11(5-6):355-607, 2019. ISSN 1935-8237", "doi:10.1561/2200000073.URLhttp://dx.doi.0rg/10.1561/2200000073.", "Tim Salimans, Han Zhang, Alec Radford, and Dimitris Metaxas. Improving gans using optimal", "transport. arXiv preprintarXiv:1803.05573,2018.", "Zebang Shen, Zhenfu Wang, Alejandro Ribeiro, and Hamed Hassani. Sinkhorn natural gradient fo", "generative models. Advances in Neural Information Processing Systems, 33:1646-1656,2020.", "Richard Sinkhorn.ARelationship Between ArbitraryPositive Matrices and Doubly Stochastic", "Matrices. The Annals of Mathematical Statistics, 35(2):876-879, 1964. doi: 10.1214/aoms/", "1177703591.URLhttps://doi.org/10.1214/aoms/1177703591.", "Shitong Wu, Wenhao Ye, Hao Wu, Huihui Wu, Wenyi Zhang, and Bo Bai. A communica", "tion optimal transport approach to the computation of rate distortion functions. arXiv preprint", "arXiv:2212.10098, 2022.", "3" ]

{ "arxiv_id": "2307.00246", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.0341", "0.4387", "0.0341", "0.4387", "0.04736", "0.1748", "0.04736" ], [ "0.1748", "0.1048", "0.1929", "0.1048", "0.1929", "0.1187", "0.1748", "0.1187" ], [ "0.2091", "0.1048", "0.2998", "0.1048", "0.2998", "0.1193", "0.2091", "0.1193" ], [ "0.1748", "0.1364", "0.2091", "0.1364", "0.2091", "0.1483", "0.1748", "0.1483" ], [ "0.2035", "0.1351", "0.2754", "0.1351", "0.2754", "0.1497", "0.2035", "0.1497" ], [ "0.1748", "0.161", "0.755", "0.161", "0.755", "0.1755", "0.1748", "0.1755" ], [ "0.4495", "0.1849", "0.5483", "0.1849", "0.5483", "0.1963", "0.4495", "0.1963" ], [ "0.8022", "0.1843", "0.824", "0.1843", "0.824", "0.1989", "0.8022", "0.1989" ], [ "0.1732", "0.2134", "0.826", "0.2134", "0.826", "0.2311", "0.1732", "0.2311" ], [ "0.1732", "0.2299", "0.2197", "0.2299", "0.2197", "0.2456", "0.1732", "0.2456" ], [ "0.3733", "0.2456", "0.4944", "0.262", "0.4885", "0.2854", "0.3677", "0.2683" ], [ "0.5254", "0.252", "0.591", "0.25", "0.5923", "0.2651", "0.527", "0.2676" ], [ "0.8013", "0.2615", "0.8267", "0.2615", "0.8267", "0.2798", "0.8013", "0.2798" ], [ "0.3882", "0.2715", "0.423", "0.2715", "0.423", "0.2874", "0.3882", "0.2874" ], [ "0.4958", "0.2683", "0.625", "0.2651", "0.626", "0.2886", "0.4968", "0.2917" ], [ "0.509", "0.2986", "0.6055", "0.2986", "0.6055", "0.3164", "0.509", "0.3164" ], [ "0.8013", "0.2986", "0.8267", "0.2986", "0.8267", "0.3164", "0.8013", "0.3164" ], [ "0.1765", "0.33", "0.8228", "0.33", "0.8228", "0.3447", "0.1765", "0.3447" ], [ "0.452", "0.2998", "0.4983", "0.2974", "0.5", "0.3132", "0.4534", "0.315" ], [ "0.1732", "0.3435", "0.8203", "0.3435", "0.8203", "0.3567", "0.1732", "0.3567" ], [ "0.1732", "0.3567", "0.824", "0.3567", "0.824", "0.375", "0.1732", "0.375" ], [ "0.1748", "0.375", "0.8164", "0.375", "0.8164", "0.3882", "0.1748", "0.3882" ], [ "0.3906", "0.3933", "0.4192", "0.3933", "0.4192", "0.411", "0.3906", "0.411" ], [ "0.4314", "0.3997", "0.63", "0.3997", "0.63", "0.4216", "0.4314", "0.4216" ], [ "0.7935", "0.4028", "0.826", "0.4028", "0.826", "0.4192", "0.7935", "0.4192" ], [ "0.3718", "0.4124", "0.437", "0.4124", "0.437", "0.428", "0.3718", "0.428" ], [ "0.1748", "0.4343", "0.8228", "0.4343", "0.8228", "0.449", "0.1748", "0.449" ], [ "0.1724", "0.4463", "0.8267", "0.4463", "0.8267", "0.4673", "0.1724", "0.4673" ], [ "0.1732", "0.464", "0.259", "0.464", "0.259", "0.4792", "0.1732", "0.4792" ], [ "0.3218", "0.4702", "0.3962", "0.4766", "0.392", "0.4968", "0.3186", "0.4905" ], [ "0.509", "0.4749", "0.5254", "0.4778", "0.5195", "0.4937", "0.5034", "0.4905" ], [ "0.585", "0.4792", "0.6807", "0.483", "0.6787", "0.502", "0.584", "0.4976" ], [ "0.3896", "0.486", "0.3962", "0.486", "0.3962", "0.4912", "0.3896", "0.4912" ], [ "0.3945", "0.486", "0.4478", "0.486", "0.4478", "0.5005", "0.3945", "0.5005" ], [ "0.7935", "0.4849", "0.8267", "0.4849", "0.8267", "0.502", "0.7935", "0.502" ], [ "0.3245", "0.4937", "0.3865", "0.4956", "0.3857", "0.511", "0.3228", "0.509" ], [ "0.451", "0.4924", "0.576", "0.4893", "0.5767", "0.511", "0.452", "0.5146" ], [ "0.1748", "0.5166", "0.824", "0.5166", "0.824", "0.531", "0.1748", "0.531" ], [ "0.1732", "0.529", "0.3635", "0.529", "0.3635", "0.5425", "0.1732", "0.5425" ], [ "0.808", "0.53", "0.826", "0.53", "0.826", "0.5444", "0.808", "0.5444" ], [ "0.4944", "0.9497", "0.5063", "0.9497", "0.5063", "0.962", "0.4944", "0.962" ] ]

[ "Published as a Tiny Paper at ICLR 2023", "A", "APPENDIX", "A.1", "PROOFS", "Theorem 1. For any source Px and distortion function p : X × -→ R>o, it holds that", "R(D) = S(D)", "(7)", "Proof. From (Cover & Thomas, 2006, Ch. 9), the optimizers Qy, Pyix of (3) for a fixed 入 > 0", "satisfy", "dPyx=(r,y)", "e-Ap(a,y)", "(8)", "dQy", "J,e-Ap(z,0) dQy", ". dPyixdPx,", "(9)", "simultaneously, which achieves a unique point on R(D) corresponding to . To show that S(D)", "QY=", "achieves the same objective as R(D) on the same Px and distortion measure, it suffices to show", "that the R(D)-optimal Qy and Py/x are feasible for S(D), since R(D) ≤ S(D). From (Peyre &", "Cuturi, 2019, Ch. 4, Prop. 4.3), the optimal coupling π* in entropic OT is unique and has the form", "dr*", "(,y) = u(r)e-Xp(r,0)(y),", "(10)", "dPxdQy", "where u(α), v(y) are dual variables that ensure * is a valid coupling. The R(D)-optimal joint", "distribution Px Pyix, which is guaranteed to be a coupling between Px and Qy due to (9), indeed", "has the form", "dPxPyix", "1", "e-^p(a,g).. 1,", "", "(c,y) =", "(11)", "dPxdQ", "Je-Ap(r,)dQ", "where the first term only depends on and the last term only depends on y. Since R(D) is a lower", "bound of S(D), we are done", "口", "4" ]

{ "arxiv_id": "2307.00246", "language": "en", "page_idx": 3 }

[ [ "0.1453", "0.0831", "0.858", "0.0831", "0.858", "0.0974", "0.1453", "0.0974" ], [ "0.2678", "0.1277", "0.4072", "0.1277", "0.4072", "0.1426", "0.2678", "0.1426" ], [ "0.5884", "0.1289", "0.748", "0.1289", "0.748", "0.1426", "0.5884", "0.1426" ], [ "0.2393", "0.1455", "0.4358", "0.1455", "0.4358", "0.1592", "0.2393", "0.1592" ], [ "0.5825", "0.1455", "0.756", "0.1455", "0.756", "0.1592", "0.5825", "0.1592" ], [ "0.1855", "0.1621", "0.486", "0.1621", "0.486", "0.177", "0.1855", "0.177" ], [ "0.6206", "0.1633", "0.7173", "0.1633", "0.7173", "0.1746", "0.6206", "0.1746" ], [ "0.2551", "0.1776", "0.4165", "0.1787", "0.4165", "0.1936", "0.2551", "0.1924" ], [ "0.5483", "0.1758", "0.7866", "0.1776", "0.7866", "0.1953", "0.5483", "0.1936" ], [ "0.2644", "0.2434", "0.3435", "0.2434", "0.3435", "0.2583", "0.2644", "0.2583" ], [ "0.531", "0.2446", "0.881", "0.2446", "0.881", "0.2583", "0.531", "0.2583" ], [ "0.513", "0.26", "0.882", "0.26", "0.882", "0.2737", "0.513", "0.2737" ], [ "0.1561", "0.2715", "0.4492", "0.2715", "0.4492", "0.2852", "0.1561", "0.2852" ], [ "0.513", "0.2756", "0.881", "0.2756", "0.881", "0.2893", "0.513", "0.2893" ], [ "0.157", "0.2874", "0.451", "0.2874", "0.451", "0.2976", "0.157", "0.2976" ], [ "0.513", "0.2922", "0.8833", "0.2915", "0.8833", "0.306", "0.513", "0.3071" ], [ "0.01932", "0.3027", "0.06635", "0.3027", "0.0697", "0.437", "0.02184", "0.437" ], [ "0.1561", "0.2998", "0.451", "0.2986", "0.451", "0.3123", "0.1561", "0.3135" ], [ "0.1545", "0.3142", "0.451", "0.3142", "0.451", "0.328", "0.1545", "0.328" ], [ "0.5146", "0.3088", "0.885", "0.3088", "0.885", "0.3225", "0.5146", "0.3225" ], [ "0.1545", "0.328", "0.4492", "0.3289", "0.4492", "0.3425", "0.1545", "0.3416" ], [ "0.5156", "0.3267", "0.882", "0.3267", "0.882", "0.3374", "0.5156", "0.3374" ], [ "0.1561", "0.3416", "0.451", "0.3416", "0.451", "0.3552", "0.1561", "0.3552" ], [ "0.513", "0.3403", "0.882", "0.3403", "0.882", "0.354", "0.513", "0.354" ], [ "0.1561", "0.3557", "0.4475", "0.3557", "0.4475", "0.3694", "0.1561", "0.3694" ], [ "0.5146", "0.3582", "0.885", "0.3582", "0.885", "0.3718", "0.5146", "0.3718" ], [ "0.1561", "0.3718", "0.451", "0.3718", "0.451", "0.3855", "0.1561", "0.3855" ], [ "0.513", "0.3735", "0.885", "0.3735", "0.885", "0.3872", "0.513", "0.3872" ], [ "0.1561", "0.3855", "0.4534", "0.3855", "0.4534", "0.3984", "0.1561", "0.3984" ], [ "0.5156", "0.3914", "0.8833", "0.3914", "0.8833", "0.402", "0.5156", "0.402" ], [ "0.1561", "0.3997", "0.4517", "0.3997", "0.4517", "0.4133", "0.1561", "0.4133" ], [ "0.513", "0.4038", "0.8833", "0.405", "0.8833", "0.4192", "0.513", "0.4187" ], [ "0.1545", "0.4146", "0.4534", "0.4146", "0.4534", "0.4275", "0.1545", "0.4275" ], [ "0.513", "0.4216", "0.8867", "0.4216", "0.8867", "0.4353", "0.513", "0.4353" ], [ "0.1561", "0.4275", "0.451", "0.4275", "0.451", "0.4412", "0.1561", "0.4412" ], [ "0.02184", "0.437", "0.0638", "0.437", "0.0638", "0.5195", "0.02184", "0.5195" ], [ "0.1561", "0.4412", "0.451", "0.4412", "0.451", "0.456", "0.1561", "0.456" ], [ "0.511", "0.437", "0.8833", "0.437", "0.8833", "0.452", "0.511", "0.452" ], [ "0.1561", "0.456", "0.4517", "0.456", "0.4517", "0.4697", "0.1561", "0.4697" ], [ "0.511", "0.4531", "0.882", "0.4531", "0.882", "0.4673", "0.511", "0.4673" ], [ "0.1545", "0.4702", "0.451", "0.4697", "0.451", "0.4827", "0.1545", "0.484" ], [ "0.513", "0.4697", "0.8833", "0.4697", "0.8833", "0.484", "0.513", "0.484" ], [ "0.1561", "0.485", "0.451", "0.485", "0.451", "0.4988", "0.1561", "0.4988" ], [ "0.511", "0.484", "0.8833", "0.485", "0.8833", "0.4995", "0.511", "0.4988" ], [ "0.1561", "0.4988", "0.451", "0.4988", "0.451", "0.5127", "0.1561", "0.5127" ], [ "0.511", "0.502", "0.879", "0.502", "0.879", "0.5156", "0.511", "0.5156" ], [ "0.1561", "0.513", "0.451", "0.513", "0.451", "0.528", "0.1561", "0.528" ], [ "0.02184", "0.5195", "0.06213", "0.5195", "0.06213", "0.73", "0.02184", "0.73" ], [ "0.513", "0.5186", "0.8833", "0.5186", "0.8833", "0.532", "0.513", "0.532" ], [ "0.157", "0.5293", "0.4492", "0.5293", "0.4492", "0.539", "0.157", "0.539" ], [ "0.513", "0.533", "0.881", "0.533", "0.881", "0.5474", "0.513", "0.5474" ], [ "0.1561", "0.5415", "0.4517", "0.5415", "0.4517", "0.5547", "0.1561", "0.5547" ], [ "0.1561", "0.5557", "0.451", "0.5557", "0.451", "0.5693", "0.1561", "0.5693" ], [ "0.5156", "0.5503", "0.8833", "0.5503", "0.8833", "0.564", "0.5156", "0.564" ], [ "0.1561", "0.571", "0.4517", "0.571", "0.4517", "0.5845", "0.1561", "0.5845" ], [ "0.513", "0.5654", "0.8833", "0.5664", "0.8833", "0.5806", "0.513", "0.58" ], [ "0.1561", "0.5845", "0.451", "0.5845", "0.451", "0.597", "0.1561", "0.597" ], [ "0.5146", "0.583", "0.882", "0.583", "0.882", "0.5967", "0.5146", "0.5967" ], [ "0.1561", "0.5967", "0.451", "0.5967", "0.451", "0.61", "0.1561", "0.61" ], [ "0.513", "0.5986", "0.8833", "0.5986", "0.8833", "0.6123", "0.513", "0.6123" ], [ "0.1561", "0.6123", "0.4517", "0.6123", "0.4517", "0.626", "0.1561", "0.626" ], [ "0.513", "0.615", "0.8135", "0.615", "0.8135", "0.629", "0.513", "0.629" ], [ "0.1561", "0.627", "0.451", "0.627", "0.451", "0.64", "0.1561", "0.64" ], [ "0.531", "0.633", "0.8833", "0.6343", "0.8833", "0.6484", "0.531", "0.648" ], [ "0.1545", "0.641", "0.4358", "0.641", "0.4358", "0.655", "0.1545", "0.655" ], [ "0.5146", "0.652", "0.881", "0.652", "0.881", "0.662", "0.5146", "0.662" ], [ "0.513", "0.6665", "0.8833", "0.6665", "0.8833", "0.6797", "0.513", "0.6797" ], [ "0.12427", "0.6816", "0.1301", "0.6816", "0.1301", "0.687", "0.12427", "0.687" ], [ "0.1377", "0.677", "0.2612", "0.677", "0.2612", "0.6904", "0.1377", "0.6904" ], [ "0.5146", "0.684", "0.881", "0.684", "0.881", "0.695", "0.5146", "0.695" ], [ "0.1184", "0.7017", "0.4895", "0.7017", "0.4895", "0.7153", "0.1184", "0.7153" ], [ "0.511", "0.6987", "0.879", "0.697", "0.879", "0.7114", "0.511", "0.7134" ], [ "0.12006", "0.7188", "0.4902", "0.7188", "0.4902", "0.7324", "0.12006", "0.7324" ], [ "0.513", "0.7144", "0.8833", "0.7144", "0.8833", "0.728", "0.513", "0.728" ], [ "0.12006", "0.734", "0.4878", "0.734", "0.4878", "0.7476", "0.12006", "0.7476" ], [ "0.511", "0.73", "0.882", "0.73", "0.882", "0.7437", "0.511", "0.7437" ], [ "0.1184", "0.7505", "0.4895", "0.7505", "0.4895", "0.764", "0.1184", "0.764" ], [ "0.513", "0.7476", "0.882", "0.7476", "0.882", "0.761", "0.513", "0.761" ], [ "0.12177", "0.767", "0.4895", "0.767", "0.4895", "0.778", "0.12177", "0.778" ], [ "0.511", "0.763", "0.882", "0.763", "0.882", "0.777", "0.511", "0.777" ], [ "0.12006", "0.782", "0.4902", "0.782", "0.4902", "0.796", "0.12006", "0.796" ], [ "0.51", "0.7793", "0.8833", "0.778", "0.8833", "0.792", "0.51", "0.7935" ], [ "0.1184", "0.799", "0.4902", "0.799", "0.4902", "0.8125", "0.1184", "0.8125" ], [ "0.511", "0.7944", "0.8833", "0.7935", "0.8833", "0.808", "0.511", "0.8086" ], [ "0.1184", "0.814", "0.492", "0.814", "0.492", "0.8276", "0.1184", "0.8276" ], [ "0.511", "0.811", "0.882", "0.811", "0.882", "0.826", "0.511", "0.826" ], [ "0.1184", "0.8306", "0.4895", "0.8306", "0.4895", "0.8457", "0.1184", "0.8457" ], [ "0.513", "0.8276", "0.75", "0.826", "0.75", "0.8403", "0.513", "0.843" ], [ "0.1184", "0.847", "0.4902", "0.847", "0.4902", "0.861", "0.1184", "0.861" ], [ "0.531", "0.8457", "0.8833", "0.8467", "0.8833", "0.861", "0.531", "0.86" ], [ "0.1184", "0.8633", "0.4902", "0.8633", "0.4902", "0.8774", "0.1184", "0.8774" ], [ "0.513", "0.8633", "0.8833", "0.8633", "0.8833", "0.877", "0.513", "0.877" ], [ "0.11755", "0.8774", "0.4902", "0.879", "0.4902", "0.8936", "0.11755", "0.8926" ], [ "0.511", "0.8774", "0.8833", "0.879", "0.8833", "0.8936", "0.511", "0.8926" ], [ "0.12006", "0.8955", "0.4492", "0.8955", "0.4492", "0.909", "0.12006", "0.909" ], [ "0.513", "0.8955", "0.8833", "0.8955", "0.8833", "0.909", "0.513", "0.909" ] ]

[ "The Parallelism Tradeoff: Limitations of Log-Precision Transformers", "William Merrill", "Ashish Sabharwal", "Center for Data Science", "Allen Institute for AI", "New York University, New York, NY", "Seattle, WA", "willm@nyu.edu", "ashishs@allenai.org", "Abstract", "Early theoretical work on transformers estab-", "lished their Turing completeness, albeit with as-", "Despite their omnipresence in modern NLP", "sumptions like infinite precision and arbitrarily", "characterizing thecomputationalpower of", "powerful feedforward subnets (Pérez et al., 2019;", " ", "transformer neural nets remains an interest-", "ing open question. We prove that trans-", "Dehghani et al., 2019).On the other hand, a", "formers whose arithmetic precision is log-", "strandofmorerecentworkusestechniques from", "arithmic in the number of input tokens (and", "circuit complexity theory to derive strong limita-", "whose feedforward nets are computable us", "tions on the types of problems transformers can", "ing space linear in their input) can be sim-", "solve given restrictions on the form of attention", "ulated by constant-depth logspace-uniform", "allowedinthetransformer.Specifically,Hahn", "threshold circuits.This provides insight", "(2020) and Hao et al. (2022) showed transform-", "on the power of transformers using known", "ers restricted to hard attention are very limited:", "results in complexity theory.For exam-", "['s] ", "ple, if L ≠ P (i.e., not all poly-time", "they can only solve problems in a weak complex-", "problems can be solved using logarithmic", "ity class (non-uniform Ac) that doesn't even con-", "space),then transformers cannot even accu-", "tain basic problems like majority of n bits. Merrill", "rately solve linear equalities or check mem-", "et al. (2022) extended this to a more general class", "bership in an arbitrary context-free gram-", "of “\"saturated attention\" transformers with a float", "mar with empty productions. Our result in-", "00221", "ing point datatype, and showed a larger class of", "tuitively emerges from the transformer ar-", "problems (non-uniform Tc) as an upper bound.", "chitecture's high parallelizability.We thus", "speculatively introduce the idea of a fun-", "This motivates analyzing a setting that strikes a", "damental parallelism tradeoff: any model", "middle ground: Can we characterize transformers", "architecture as parallelizable as the trans-", "whose precision and feedforward nets' computa-", "former will obey limitations similar to it.", "tional power are realistically bounded, but where", "Since parallelism is key to training models", "attention is also realistically expressive?", "at massive scale, this suggests a potential in-", "An important practical limitation of these prior", "herent weakness of the scaling paradigm.", "resultsisthe“non-uniform\"natureofthecon-", "sidered circuit classes, which makes these classes", "1", "Introduction", "non-realizable and the findings difficult to inter-", "This work aims to characterize the computational", "pret. This is because non-uniform Aco and Tco", "model implicit in transformer neural networks", "while highly limited in computation, also con-", "(Vaswani et al., 2017), which form the basis of re-", "tain some problems that are not even decidable.", "cent breakthroughs in large language models such", "i.e., for which there doesn't exist any exact algo-", "as BERT (Devlin et al., 2019). T5 (Raffel et al.,", "rithm. Thus, non-uniform classes cannot be di-", "2020), and GPT-3 (Brown et al., 2020). What", "rectly compared with standard algorithmic com-", "computational primitives can the transformer's", "plexity classes such as P, NP, etc. This motivates", "components implement, and what problems can", "our second key question: Can we derive uniform", "the full system solve in aggregate? These ques-", "upper bounds on transformers?", "tions are important for interpreting transformers in", "We show that one can achieve both of these", "a principled way, understanding potential limita-", "goals by making the modest assumption that all", "tions of their reasoning capabilities, and building", "values in the transformer have O(log n) preci-", "trust in deployed transformer-based systems.", "sion (where n is the number of input tokens)," ]

{ "arxiv_id": "2207.00729", "language": "en", "page_idx": 0 }

[ [ "0.1377", "0.0748", "0.4895", "0.0748", "0.4895", "0.08905", "0.1377", "0.08905" ], [ "0.513", "0.0766", "0.882", "0.0766", "0.882", "0.0903", "0.513", "0.0903" ], [ "0.1184", "0.09143", "0.4878", "0.09265", "0.4878", "0.1069", "0.1184", "0.1057" ], [ "0.513", "0.09265", "0.8833", "0.09143", "0.8833", "0.1057", "0.513", "0.1069" ], [ "0.12006", "0.1081", "0.4902", "0.1081", "0.4902", "0.1223", "0.12006", "0.1223" ], [ "0.513", "0.10986", "0.879", "0.10986", "0.879", "0.12054", "0.513", "0.12054" ], [ "0.12006", "0.1247", "0.4902", "0.1247", "0.4902", "0.1383", "0.12006", "0.1383" ], [ "0.513", "0.1247", "0.754", "0.1247", "0.754", "0.1383", "0.513", "0.1383" ], [ "0.12006", "0.1401", "0.4902", "0.1401", "0.4902", "0.1538", "0.12006", "0.1538" ], [ "0.513", "0.1437", "0.8833", "0.1443", "0.8833", "0.1592", "0.513", "0.158" ], [ "0.12006", "0.1567", "0.4895", "0.1567", "0.4895", "0.1704", "0.12006", "0.1704" ], [ "0.5146", "0.1621", "0.8833", "0.1621", "0.8833", "0.1758", "0.5146", "0.1758" ], [ "0.12006", "0.1746", "0.4878", "0.1746", "0.4878", "0.1853", "0.12006", "0.1853" ], [ "0.5146", "0.1776", "0.7676", "0.1776", "0.7676", "0.1912", "0.5146", "0.1912" ], [ "0.1184", "0.1895", "0.4895", "0.1895", "0.4895", "0.2031", "0.1184", "0.2031" ], [ "0.11755", "0.2037", "0.455", "0.2048", "0.455", "0.2197", "0.11755", "0.2185" ], [ "0.5156", "0.2031", "0.8833", "0.2031", "0.8833", "0.2162", "0.5156", "0.2162" ], [ "0.513", "0.2173", "0.882", "0.2185", "0.882", "0.2328", "0.513", "0.2322" ], [ "0.1184", "0.2351", "0.3376", "0.2351", "0.3376", "0.2494", "0.1184", "0.2494" ], [ "0.5146", "0.2351", "0.885", "0.2351", "0.885", "0.2488", "0.5146", "0.2488" ], [ "0.5146", "0.2517", "0.846", "0.2517", "0.846", "0.2654", "0.5146", "0.2654" ], [ "0.12006", "0.2583", "0.4895", "0.2583", "0.4895", "0.2725", "0.12006", "0.2725" ], [ "0.1184", "0.275", "0.4902", "0.275", "0.4902", "0.288", "0.1184", "0.288" ], [ "0.513", "0.2756", "0.882", "0.2756", "0.882", "0.2893", "0.513", "0.2893" ], [ "0.1184", "0.2915", "0.4895", "0.2915", "0.4895", "0.3047", "0.1184", "0.3047" ], [ "0.513", "0.2915", "0.882", "0.2922", "0.882", "0.3071", "0.513", "0.306" ], [ "0.12006", "0.3071", "0.4878", "0.3071", "0.4878", "0.3206", "0.12006", "0.3206" ], [ "0.513", "0.308", "0.8833", "0.308", "0.8833", "0.3218", "0.513", "0.3218" ], [ "0.1184", "0.3237", "0.4902", "0.3237", "0.4902", "0.3374", "0.1184", "0.3374" ], [ "0.513", "0.3247", "0.8833", "0.3247", "0.8833", "0.339", "0.513", "0.339" ], [ "0.1184", "0.339", "0.4895", "0.339", "0.4895", "0.3528", "0.1184", "0.3528" ], [ "0.513", "0.3416", "0.882", "0.3416", "0.882", "0.3552", "0.513", "0.3552" ], [ "0.11755", "0.3552", "0.4895", "0.354", "0.4895", "0.3682", "0.11755", "0.3694" ], [ "0.511", "0.357", "0.706", "0.3557", "0.706", "0.3706", "0.511", "0.3718" ], [ "0.11755", "0.3706", "0.4878", "0.3706", "0.4878", "0.3843", "0.11755", "0.3843" ], [ "0.5146", "0.3777", "0.881", "0.3777", "0.881", "0.3926", "0.5146", "0.3926" ], [ "0.1184", "0.3872", "0.4844", "0.3872", "0.4844", "0.401", "0.1184", "0.401" ], [ "0.513", "0.3914", "0.885", "0.3926", "0.885", "0.4104", "0.513", "0.4092" ], [ "0.1184", "0.4163", "0.4895", "0.4163", "0.4895", "0.43", "0.1184", "0.43" ], [ "0.5156", "0.411", "0.6187", "0.411", "0.6187", "0.4229", "0.5156", "0.4229" ], [ "0.1184", "0.4312", "0.4895", "0.4316", "0.4895", "0.4465", "0.1184", "0.4453" ], [ "0.529", "0.4287", "0.882", "0.4312", "0.882", "0.4453", "0.529", "0.4436" ], [ "0.12006", "0.449", "0.4902", "0.449", "0.4902", "0.462", "0.12006", "0.462" ], [ "0.511", "0.4465", "0.8833", "0.4453", "0.8833", "0.4602", "0.511", "0.4614" ], [ "0.1184", "0.4644", "0.4902", "0.4631", "0.4902", "0.478", "0.1184", "0.4785" ], [ "0.513", "0.4644", "0.712", "0.4644", "0.712", "0.478", "0.513", "0.478" ], [ "0.1184", "0.481", "0.4895", "0.481", "0.4895", "0.4946", "0.1184", "0.4946" ], [ "0.513", "0.4827", "0.8833", "0.4827", "0.8833", "0.4976", "0.513", "0.4976" ], [ "0.1184", "0.4963", "0.4902", "0.4963", "0.4902", "0.511", "0.1184", "0.511" ], [ "0.513", "0.5005", "0.879", "0.5005", "0.879", "0.514", "0.513", "0.514" ], [ "0.1184", "0.5127", "0.4895", "0.5127", "0.4895", "0.5254", "0.1184", "0.5254" ], [ "0.513", "0.5166", "0.8447", "0.5156", "0.8447", "0.53", "0.513", "0.531" ], [ "0.1184", "0.5293", "0.4895", "0.5293", "0.4895", "0.5435", "0.1184", "0.5435" ], [ "0.531", "0.5376", "0.8833", "0.5376", "0.8833", "0.552", "0.531", "0.552" ], [ "0.12006", "0.546", "0.3552", "0.546", "0.3552", "0.5586", "0.12006", "0.5586" ], [ "0.513", "0.5527", "0.8867", "0.5527", "0.8867", "0.571", "0.513", "0.571" ], [ "0.11755", "0.5747", "0.3315", "0.576", "0.3315", "0.5903", "0.11755", "0.589" ], [ "0.513", "0.5693", "0.8833", "0.5684", "0.8833", "0.583", "0.513", "0.5845" ], [ "0.11755", "0.604", "0.4895", "0.604", "0.4895", "0.6177", "0.11755", "0.6177" ], [ "0.511", "0.605", "0.882", "0.6055", "0.882", "0.6206", "0.511", "0.619" ], [ "0.11755", "0.6206", "0.4902", "0.619", "0.4902", "0.6343", "0.11755", "0.6353" ], [ "0.5156", "0.6226", "0.882", "0.6226", "0.882", "0.637", "0.5156", "0.637" ], [ "0.1184", "0.637", "0.4902", "0.637", "0.4902", "0.651", "0.1184", "0.651" ], [ "0.513", "0.64", "0.8003", "0.64", "0.8003", "0.655", "0.513", "0.655" ], [ "0.11755", "0.652", "0.4844", "0.653", "0.4844", "0.6675", "0.11755", "0.6665" ], [ "0.5283", "0.6567", "0.885", "0.6577", "0.885", "0.676", "0.5283", "0.675" ], [ "0.12006", "0.6694", "0.4878", "0.6694", "0.4878", "0.683", "0.12006", "0.683" ], [ "0.511", "0.675", "0.882", "0.675", "0.882", "0.6895", "0.511", "0.6895" ], [ "0.1184", "0.684", "0.486", "0.684", "0.486", "0.6978", "0.1184", "0.6978" ], [ "0.5146", "0.6914", "0.8833", "0.6914", "0.8833", "0.706", "0.5146", "0.706" ], [ "0.12006", "0.7017", "0.4895", "0.7017", "0.4895", "0.7153", "0.12006", "0.7153" ], [ "0.513", "0.708", "0.7246", "0.708", "0.7246", "0.7217", "0.513", "0.7217" ], [ "0.11755", "0.7163", "0.4744", "0.7163", "0.4744", "0.73", "0.11755", "0.73" ], [ "0.1394", "0.7354", "0.486", "0.7354", "0.486", "0.745", "0.1394", "0.745" ], [ "0.5156", "0.734", "0.882", "0.734", "0.882", "0.7476", "0.5156", "0.7476" ], [ "0.12006", "0.7505", "0.4895", "0.7505", "0.4895", "0.764", "0.12006", "0.764" ], [ "0.513", "0.75", "0.882", "0.75", "0.882", "0.763", "0.513", "0.763" ], [ "0.1184", "0.7666", "0.4895", "0.7666", "0.4895", "0.78", "0.1184", "0.78" ], [ "0.511", "0.7656", "0.8833", "0.764", "0.8833", "0.7793", "0.511", "0.78" ], [ "0.1184", "0.783", "0.4895", "0.783", "0.4895", "0.7974", "0.1184", "0.7974" ], [ "0.513", "0.783", "0.8833", "0.783", "0.8833", "0.7964", "0.513", "0.7964" ], [ "0.12006", "0.8", "0.4895", "0.8", "0.4895", "0.8135", "0.12006", "0.8135" ], [ "0.513", "0.799", "0.882", "0.799", "0.882", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.8154", "0.4902", "0.8154", "0.4902", "0.829", "0.1184", "0.829" ], [ "0.513", "0.814", "0.6733", "0.814", "0.6733", "0.8276", "0.513", "0.8276" ], [ "0.532", "0.8306", "0.885", "0.8306", "0.885", "0.844", "0.532", "0.844" ], [ "0.1427", "0.8433", "0.4902", "0.8433", "0.4902", "0.857", "0.1427", "0.857" ], [ "0.511", "0.847", "0.8833", "0.847", "0.8833", "0.861", "0.511", "0.861" ], [ "0.1184", "0.857", "0.4895", "0.857", "0.4895", "0.8706", "0.1184", "0.8706" ], [ "0.513", "0.8633", "0.885", "0.8633", "0.885", "0.877", "0.513", "0.877" ], [ "0.11755", "0.8716", "0.43", "0.8716", "0.43", "0.882", "0.11755", "0.882" ], [ "0.1427", "0.883", "0.4937", "0.883", "0.4937", "0.8965", "0.1427", "0.8965" ], [ "0.511", "0.8774", "0.885", "0.879", "0.885", "0.8936", "0.511", "0.8926" ], [ "0.12006", "0.898", "0.4265", "0.898", "0.4265", "0.9077", "0.12006", "0.9077" ], [ "0.5146", "0.8945", "0.8867", "0.8945", "0.8867", "0.909", "0.5146", "0.909" ] ]

[ "Our formal model is based on a binary classi-", "directed edges represent the flow of function out", "fication view of transformers. However, our re-", "puts into inputs of other functions. One or more", "sults apply directly to multi-class classification as", "nodes in the circuit are marked such that theii", "well and can be extended to generation problems", "value is the output of the circuit.", "by viewing, for instance, next word prediction in", "Definition 1. For a set of functions G, a G-circuit", "NLP as a multi-class classification problem. How-", "is a directed acyclic computation graph where the", "ever,if thetransformer decoderis allowed to con-", "internal nodes have labels from g.", "dition on its previous output in a generation prob-", "lem, then this would violate our formal setup.", "Complexity Measures. The size of a circuit is", "the total number of gates in it, including negation.", "2.1 Potential Applications", "The depth of a circuit is the length of the longest", "path from any input node to any output node.", "Extracting Circuits from Transformers.El-", "hage et al. (2021) propose extracting circuits5 that", "Circuit Families.A circuit family generalizes a", "capture the computational structure of transform-", "circuit to take variable-length binary strings as in-", "ers. Our results suggest threshold circuit fami-", "put. Formally, a circuit family is a sequence of", "lies are a good formalism for expressing mecha-", "circuits Cn : [0,1]n -→ [0,1] for n E N. A", "nisms extracted from transformers. Constructively", "circuit family implicitly recognizes a formal lan-", "converting transformers to threshold circuits is be-", "guage defined as follows:", "yond the scope of the current paper, although we", "Definition 2. A circuit family Cn recognizes L C", "hope to explore this in more detail in future work.", "[0, 1]* if, for all E [0, 1]*, Cj](c) = 1 if and", "Testing Separation Candidates in Complexity", "only if r E L.", "Theory.Thm. 2 also motivates a paradigm for", "We now define classes of languages by con-", "quickly testing complexity theory conjectures. If", "straining the complexity of the circuit families", "a problem is believed to separate TC and NCl, a", "needed to recognize them:", "transformer can be trained on problem instances.", "Definition 3. Let non-uniform AcO be the set of", "If the transformer generalizes perfectly to harder", "L C {0, 1]* such that L is recognizable by a poly-", "instances than it was trained on, this gives an em-", "size, constant-depth {, A, V}-circuit family.", "pirical hint that the problem is in Tco, providing", "For k E N, a threshold gate <k takes m input", "evidence against the conjecture.", "bits and returns whether 1 &i ≤ k. We define", "3Circuit Computation", ">k analogously. For example, <3(110011) = 0.", "Let f0, 1↓* be the set of finite binary strings. For", "Definition 4. Let Tco be the set of L C", "α E [0, 1]*, let cl be its length. We refer to", "[O, 1]* such that L is recognizable by a poly-size,", "a function from f0,11* to {0,11* as a boolean", "constant-depth [<k, ≥k]heN-circuit.", "function. Boolean functions can implement arith", "The gates , A, and V are all just special cases", "metic operations if we define a semantics for bi-", "of thresholds, so we can imagine Tco circuits to", "nary strings as numbers. We will treat the inter-", "have access to these as well. Thus, Tco circuits", "mediate values in a transformer as binary strings,", "can implement Aco circuits.", "and the internal operations as boolean functions.", "Circuits are a model of computation for com-", "Circuit Serialization. We identify a circuit with", "puting boolean functions of fixed-length binary", "its serialization in a formal language that identi-", "strings.6 Formally, a circuit is a directed acyclic", "fies each node's label and adjacency list.We will", "computation graph. The leaf nodes represent bi-", "adopt a specific grammar for concreteness, but our", "nary variables and their negations. The internal", "construction can be adapted to other string repre-", "nodes represent functions in some set G, and the", "sentations of circuits.", "We define a circuit serialization as a traversal", "Their sense of “circuit\" is not exactly the formal sense", "of a circuit ordered by some topological sort. In", "we use in this paper, though the goal of capturing transform-", "this serialization, leaf nodes (variables) are repre-", "ersimplicit computational mechanismisthe same.", "‘For a mini-tutorial on circuit complexity theory and its", "sented by the string X. An internal node (gate) is", "relevance to transformers, see Merrill et al. (2022).", "represented in Polish notation by the function it" ]

{ "arxiv_id": "2207.00729", "language": "en", "page_idx": 3 }

[ [ "0.1184", "0.0766", "0.4902", "0.0748", "0.4902", "0.0885", "0.1184", "0.0903" ], [ "0.513", "0.0766", "0.882", "0.0766", "0.882", "0.0903", "0.513", "0.0903" ], [ "0.1184", "0.09265", "0.492", "0.09143", "0.492", "0.1057", "0.1184", "0.1069" ], [ "0.511", "0.09143", "0.8833", "0.09265", "0.8833", "0.1069", "0.511", "0.1057" ], [ "0.12006", "0.10925", "0.4895", "0.10925", "0.4895", "0.1223", "0.12006", "0.1223" ], [ "0.511", "0.1069", "0.8833", "0.1081", "0.8833", "0.1223", "0.511", "0.1217" ], [ "0.1184", "0.1247", "0.4902", "0.1247", "0.4902", "0.1383", "0.1184", "0.1383" ], [ "0.513", "0.1247", "0.8833", "0.1247", "0.8833", "0.1383", "0.513", "0.1383" ], [ "0.1184", "0.1401", "0.4475", "0.1401", "0.4475", "0.1538", "0.1184", "0.1538" ], [ "0.5146", "0.1426", "0.8833", "0.1426", "0.8833", "0.1526", "0.5146", "0.1526" ], [ "0.1368", "0.155", "0.4895", "0.155", "0.4895", "0.1687", "0.1368", "0.1687" ], [ "0.5146", "0.1561", "0.8833", "0.1561", "0.8833", "0.1704", "0.5146", "0.1704" ], [ "0.1184", "0.1692", "0.4895", "0.1692", "0.4895", "0.1829", "0.1184", "0.1829" ], [ "0.511", "0.1716", "0.881", "0.1716", "0.881", "0.1853", "0.511", "0.1853" ], [ "0.1184", "0.1829", "0.4902", "0.1829", "0.4902", "0.1965", "0.1184", "0.1965" ], [ "0.513", "0.187", "0.8833", "0.1882", "0.8833", "0.2031", "0.513", "0.2019" ], [ "0.12006", "0.1996", "0.336", "0.1996", "0.336", "0.2102", "0.12006", "0.2102" ], [ "0.511", "0.2048", "0.882", "0.2048", "0.882", "0.2185", "0.511", "0.2185" ], [ "0.1545", "0.2214", "0.4417", "0.2214", "0.4417", "0.2351", "0.1545", "0.2351" ], [ "0.5146", "0.2203", "0.8867", "0.2203", "0.8867", "0.234", "0.5146", "0.234" ], [ "0.1545", "0.2393", "0.4492", "0.2393", "0.4492", "0.2542", "0.1545", "0.2542" ], [ "0.511", "0.2363", "0.882", "0.2369", "0.882", "0.2517", "0.511", "0.2505" ], [ "0.1587", "0.2559", "0.351", "0.2559", "0.351", "0.2737", "0.1587", "0.2737" ], [ "0.5146", "0.2542", "0.854", "0.2542", "0.854", "0.2673", "0.5146", "0.2673" ], [ "0.529", "0.2683", "0.881", "0.2683", "0.881", "0.282", "0.529", "0.282" ], [ "0.1823", "0.2756", "0.3167", "0.2756", "0.3167", "0.2903", "0.1823", "0.2903" ], [ "0.513", "0.2852", "0.8833", "0.2852", "0.8833", "0.2986", "0.513", "0.2986" ], [ "0.1234", "0.3018", "0.4878", "0.3018", "0.4878", "0.3154", "0.1234", "0.3154" ], [ "0.5146", "0.3018", "0.882", "0.3018", "0.882", "0.3154", "0.5146", "0.3154" ], [ "0.12006", "0.3184", "0.4878", "0.3184", "0.4878", "0.332", "0.12006", "0.332" ], [ "0.5156", "0.3176", "0.8833", "0.3176", "0.8833", "0.3308", "0.5156", "0.3308" ], [ "0.1184", "0.3333", "0.4895", "0.3342", "0.4895", "0.3486", "0.1184", "0.3474" ], [ "0.511", "0.332", "0.882", "0.332", "0.882", "0.3467", "0.511", "0.3467" ], [ "0.11755", "0.3499", "0.4382", "0.3486", "0.4382", "0.3635", "0.11755", "0.364" ], [ "0.513", "0.3486", "0.833", "0.3499", "0.833", "0.364", "0.513", "0.3635" ], [ "0.1621", "0.3706", "0.451", "0.3706", "0.451", "0.3843", "0.1621", "0.3843" ], [ "0.513", "0.3682", "0.8833", "0.3706", "0.8833", "0.3855", "0.513", "0.383" ], [ "0.1184", "0.3896", "0.4895", "0.3884", "0.4895", "0.4026", "0.1184", "0.4038" ], [ "0.5146", "0.3872", "0.8833", "0.3872", "0.8833", "0.401", "0.5146", "0.401" ], [ "0.1184", "0.4062", "0.4902", "0.4062", "0.4902", "0.4192", "0.1184", "0.4192" ], [ "0.511", "0.4026", "0.8833", "0.402", "0.8833", "0.4163", "0.511", "0.4175" ], [ "0.511", "0.4163", "0.8833", "0.4175", "0.8833", "0.4353", "0.511", "0.434" ], [ "0.11755", "0.4229", "0.4878", "0.4229", "0.4878", "0.4358", "0.11755", "0.4358" ], [ "0.11755", "0.4395", "0.4878", "0.4395", "0.4878", "0.4531", "0.11755", "0.4531" ], [ "0.511", "0.434", "0.748", "0.4353", "0.748", "0.4495", "0.511", "0.449" ], [ "0.11755", "0.4536", "0.4895", "0.4548", "0.4895", "0.4697", "0.11755", "0.4685" ], [ "0.531", "0.456", "0.8833", "0.456", "0.8833", "0.4702", "0.531", "0.4702" ], [ "0.11755", "0.4697", "0.3198", "0.4702", "0.3198", "0.485", "0.11755", "0.484" ], [ "0.513", "0.4727", "0.8833", "0.4727", "0.8833", "0.4863", "0.513", "0.4863" ], [ "0.1368", "0.485", "0.4878", "0.485", "0.4878", "0.4988", "0.1368", "0.4988" ], [ "0.511", "0.4868", "0.8833", "0.488", "0.8833", "0.503", "0.511", "0.502" ], [ "0.1184", "0.4995", "0.4895", "0.4995", "0.4895", "0.513", "0.1184", "0.513" ], [ "0.511", "0.505", "0.882", "0.505", "0.882", "0.5186", "0.511", "0.5186" ], [ "0.1184", "0.514", "0.465", "0.514", "0.465", "0.528", "0.1184", "0.528" ], [ "0.513", "0.521", "0.885", "0.5195", "0.885", "0.5337", "0.513", "0.535" ], [ "0.1621", "0.5376", "0.4458", "0.5376", "0.4458", "0.552", "0.1621", "0.552" ], [ "0.513", "0.536", "0.8833", "0.535", "0.8833", "0.55", "0.513", "0.5503" ], [ "0.1604", "0.554", "0.4282", "0.554", "0.4282", "0.572", "0.1604", "0.572" ], [ "0.513", "0.5527", "0.8555", "0.5527", "0.8555", "0.5664", "0.513", "0.5664" ], [ "0.532", "0.5693", "0.882", "0.5693", "0.882", "0.583", "0.532", "0.583" ], [ "0.168", "0.576", "0.3552", "0.576", "0.3552", "0.5903", "0.168", "0.5903" ], [ "0.5146", "0.585", "0.8833", "0.585", "0.8833", "0.5986", "0.5146", "0.5986" ], [ "0.1873", "0.5933", "0.3132", "0.5933", "0.3132", "0.607", "0.1873", "0.607" ], [ "0.5156", "0.6016", "0.881", "0.6016", "0.881", "0.615", "0.5156", "0.615" ], [ "0.513", "0.615", "0.882", "0.6133", "0.882", "0.6313", "0.513", "0.633" ], [ "0.1368", "0.6353", "0.4902", "0.6353", "0.4902", "0.6494", "0.1368", "0.6494" ], [ "0.511", "0.632", "0.882", "0.632", "0.882", "0.6455", "0.511", "0.6455" ], [ "0.1184", "0.652", "0.4878", "0.652", "0.4878", "0.665", "0.1184", "0.665" ], [ "0.513", "0.6494", "0.882", "0.6494", "0.882", "0.663", "0.513", "0.663" ], [ "0.1184", "0.6665", "0.4895", "0.6675", "0.4895", "0.6816", "0.1184", "0.681" ], [ "0.513", "0.6665", "0.8833", "0.665", "0.8833", "0.6787", "0.513", "0.6797" ], [ "0.1184", "0.683", "0.4895", "0.683", "0.4895", "0.697", "0.1184", "0.697" ], [ "0.513", "0.681", "0.885", "0.681", "0.885", "0.6953", "0.513", "0.6953" ], [ "0.12006", "0.7007", "0.4895", "0.7007", "0.4895", "0.7144", "0.12006", "0.7144" ], [ "0.511", "0.6978", "0.8135", "0.6987", "0.8135", "0.7134", "0.511", "0.712" ], [ "0.12006", "0.7188", "0.4878", "0.7188", "0.4878", "0.7285", "0.12006", "0.7285" ], [ "0.513", "0.7246", "0.833", "0.7256", "0.833", "0.7407", "0.513", "0.7393" ], [ "0.12427", "0.733", "0.4895", "0.733", "0.4895", "0.7476", "0.12427", "0.7476" ], [ "0.1184", "0.7476", "0.482", "0.749", "0.482", "0.763", "0.1184", "0.7627" ], [ "0.5146", "0.752", "0.882", "0.752", "0.882", "0.7656", "0.5146", "0.7656" ], [ "0.1621", "0.7686", "0.4282", "0.7686", "0.4282", "0.782", "0.1621", "0.782" ], [ "0.5146", "0.7686", "0.879", "0.7686", "0.879", "0.7793", "0.5146", "0.7793" ], [ "0.12006", "0.788", "0.4895", "0.788", "0.4895", "0.8027", "0.12006", "0.8027" ], [ "0.5146", "0.785", "0.882", "0.785", "0.882", "0.799", "0.5146", "0.799" ], [ "0.11755", "0.8047", "0.4895", "0.8047", "0.4895", "0.8184", "0.11755", "0.8184" ], [ "0.513", "0.8003", "0.882", "0.8003", "0.882", "0.814", "0.513", "0.814" ], [ "0.1184", "0.821", "0.4702", "0.821", "0.4702", "0.835", "0.1184", "0.835" ], [ "0.513", "0.8164", "0.882", "0.8154", "0.882", "0.83", "0.513", "0.8306" ], [ "0.513", "0.833", "0.705", "0.833", "0.705", "0.8467", "0.513", "0.8467" ], [ "0.1427", "0.8433", "0.4878", "0.8433", "0.4878", "0.857", "0.1427", "0.857" ], [ "0.1184", "0.857", "0.4895", "0.857", "0.4895", "0.8706", "0.1184", "0.8706" ], [ "0.532", "0.854", "0.882", "0.8555", "0.882", "0.8706", "0.532", "0.868" ], [ "0.12006", "0.8716", "0.3628", "0.8716", "0.3628", "0.882", "0.12006", "0.882" ], [ "0.511", "0.868", "0.881", "0.869", "0.881", "0.883", "0.511", "0.882" ], [ "0.1411", "0.882", "0.4902", "0.882", "0.4902", "0.8965", "0.1411", "0.8965" ], [ "0.511", "0.883", "0.8833", "0.883", "0.8833", "0.8965", "0.511", "0.8965" ], [ "0.11755", "0.8965", "0.2477", "0.8965", "0.2477", "0.91", "0.11755", "0.91" ], [ "0.513", "0.898", "0.6675", "0.898", "0.6675", "0.9077", "0.513", "0.9077" ] ]

[ "computes (AND, OR, or NOT) followed by a list of", "Uniformity.The circuit families we have de", "pointers to its arguments. Each argument &1j of", "fined above are non-uniform, meaning that we do", "gate i encodes (in a unary)a zero-indexed pointer", "not enforce that the circuits processing different", "to the j-th gate in the circuit, where j < i. The", "input sizes must be related in any way. In degen-", "final node is interpreted as the circuit output.", "erate cases,non-uniformcircuitfamiliescan solve", "To serialize {A, V}-circuits, we use the follow-", "undecidable problems because they have infinite", "ing grammar, where the i parameter is passed", "description length, making them a physically unre-", "through Gate[il nonterminals to track the index of", "alizable model of computation. Complexity theo-", "the gate in left-to-right order:", "rists have thus introduced uniform circuit families.", "Circuit -→ Gate[1] Gate[2] ... Gatelg]", "Uniform circuit families are a realizable model of", "Gate[i] → X [ NOT Arg[] Op Arg[]", "computation with relations to classes in computa-", "Arg[i] → &1i s.t. j<i", "tional complexity and formal language theory.", "Intuitively, in a uniform circuit family, the cir-", "Op → ANDI OR", "cuits for different input sizes must be “somewhat", "In the Arglil rule, we enforce that j < i so that ar-", "similar\" to each other. We formalize this (cf.", "guments must be pointers to already defined gates.", "Arora and Barak,2009)by saying that there exists", "As an example of this serialization language, the", "a resource-constrained Turing machine that maps", "circuit for i V-2 V 3 is represented as?", "the input 1n to a serialization of circuit Cn.", "X X X NOT &1 OR & &111 &11", "Definition 5. A language L is (S(n), I(n))-space", "By convention (cf. s3), negations in Aco circuits", "uniformly computable by a circuit model M iff", "are usually taken to occur at the beginning of the", "there exists a Turing machine that, for all n ≥ O,", "uses S(n) space to map 1n to an M-circuit recog-", "circuit, rather than after A or V nodes.8 Our seri-", "alization grammar does not enforce this property,", "nizing L on inputs of size I(n).", "but of course any circuit with this property can be", "This notion of uniformity is more general than", "serialized by our grammar.", "the standard notion in that the input size I(n) is a", "It is a bit more complicated to serialize thresh-", "function of the problem complexity n. The reason", "old circuits. Formally, a threshold circuit serial-", "for this is that we will apply uniformity to subcom-", "ization is generated by the following grammar:", "putations with different input sizes I(n) within a", "Circuit → Gate[1] Gate[2] ... Gate[g]", "larger computation of input size n. The standard", "Gate[] → X [ Dir 1om-k Arg[]m", "notion of uniformity corresponds to I(n) = n.", "Furthermore, we will refer to a circuit fam-", "Arglil → &ij s.t.j<i", "ily as uniform if it is uniformly computable with", "Dir →<= |>=", "S(n) = O(log n) (cf. Arora and Barak, 2009)", "We can define uniform versions of ACO and TCo", "In the rewrite rule for Gate[il, m E N is the", "by adopting the previous definitions exactly, but", "arity of the gate, and k < m is its threshold.", "also enforcing uniformity. For the rest of the pa-", "The span 1 after Dir can be interpreted semanti-", "per we will clarify whether we mean the uniform", "cally as a unary encoding of the parameter k for a", "or non-uniform variant of Tc0 when unclear from", "threshold gate, padded by o's to the number of to-", "context, since both classes will come up.", "tal arguments of gatei.For simplicity,weimagine", "4Bounded-Precision Transformers", " gates are represented as unary <o gates. Thus,", "the circuit >1(c1, -2) would be represented as", "A transformer (Vaswani et al., 2017) is a neural", "X X <= 00 &1 >= 10 & &11", "network architecture made up of a constant num-", "We say a threshold circuit serialization is in pre-", "ber of transformer layers. A transformer layer is", "fix form if all inputs (x) come before all threshold", "a module that computes self-attention over a se-", "gates (<= or >-), as is the case in this example.", "quencefollowed by an elementwise transforma-", "tion of the output vectors.", "7 Spaces here (and in the grammar) are added for readabil-", "ity. We will ignore these spaces when passing circuit serial-", "'Consider the unary language 1n such that Turing ma-", "izations as inputs to a transformer in s7.", "chine n (under some arbitrary enumeration) halts. This prob", "8 we can apply De Morgan's laws to force any ACO circuit", "lem is in non-uniform Aco since we can hard-code the right", "to have this property.", "answerforeachn in Cn." ]

{ "arxiv_id": "2207.00729", "language": "en", "page_idx": 4 }

[ [ "0.2217", "0.08075", "0.781", "0.082", "0.781", "0.0998", "0.2217", "0.0986" ], [ "0.2275", "0.1312", "0.775", "0.1312", "0.775", "0.1484", "0.2275", "0.1484" ], [ "0.2527", "0.152", "0.756", "0.1538", "0.756", "0.1716", "0.2527", "0.1692" ], [ "0.2544", "0.1704", "0.7524", "0.1733", "0.7524", "0.1912", "0.2544", "0.1882" ], [ "0.2284", "0.1936", "0.781", "0.1948", "0.781", "0.209", "0.2284", "0.2079" ], [ "0.2678", "0.2666", "0.3418", "0.2666", "0.3418", "0.2778", "0.2678", "0.2778" ], [ "0.5996", "0.2625", "0.7983", "0.2625", "0.7983", "0.2756", "0.5996", "0.2756" ], [ "0.544", "0.2798", "0.642", "0.2798", "0.642", "0.2944", "0.544", "0.2944" ], [ "0.7173", "0.279", "0.839", "0.279", "0.839", "0.2922", "0.7173", "0.2922" ], [ "0.147", "0.2922", "0.4634", "0.2922", "0.4634", "0.306", "0.147", "0.306" ], [ "0.5425", "0.2944", "0.659", "0.2935", "0.659", "0.308", "0.5425", "0.3088" ], [ "0.7153", "0.2944", "0.8613", "0.2944", "0.8613", "0.308", "0.7153", "0.308" ], [ "0.02016", "0.3027", "0.06046", "0.3027", "0.06046", "0.4358", "0.02016", "0.4358" ], [ "0.147", "0.306", "0.461", "0.306", "0.461", "0.3196", "0.147", "0.3196" ], [ "0.529", "0.31", "0.663", "0.31", "0.663", "0.3237", "0.529", "0.3237" ], [ "0.7173", "0.31", "0.8096", "0.31", "0.8096", "0.3218", "0.7173", "0.3218" ], [ "0.147", "0.3218", "0.461", "0.3218", "0.461", "0.335", "0.147", "0.335" ], [ "0.546", "0.3267", "0.623", "0.3267", "0.623", "0.3384", "0.546", "0.3384" ], [ "0.7188", "0.3247", "0.8193", "0.3247", "0.8193", "0.3362", "0.7188", "0.3362" ], [ "0.1486", "0.3374", "0.4592", "0.3374", "0.4592", "0.3474", "0.1486", "0.3474" ], [ "0.147", "0.3499", "0.461", "0.3499", "0.461", "0.3635", "0.147", "0.3635" ], [ "0.596", "0.3474", "0.8003", "0.3474", "0.8003", "0.361", "0.596", "0.361" ], [ "0.544", "0.3623", "0.6753", "0.3623", "0.6753", "0.3723", "0.544", "0.3723" ], [ "0.669", "0.3623", "0.8364", "0.3623", "0.8364", "0.3723", "0.669", "0.3723" ], [ "0.546", "0.3723", "0.686", "0.3723", "0.686", "0.383", "0.546", "0.383" ], [ "0.147", "0.364", "0.461", "0.364", "0.461", "0.3777", "0.147", "0.3777" ], [ "0.147", "0.3772", "0.4592", "0.3777", "0.4592", "0.3914", "0.147", "0.3901" ], [ "0.544", "0.3872", "0.827", "0.3872", "0.827", "0.401", "0.544", "0.401" ], [ "0.8555", "0.3855", "0.8657", "0.3855", "0.8657", "0.4092", "0.8555", "0.4092" ], [ "0.147", "0.3926", "0.461", "0.3926", "0.461", "0.4062", "0.147", "0.4062" ], [ "0.546", "0.398", "0.7", "0.398", "0.7", "0.411", "0.546", "0.411" ], [ "0.147", "0.4062", "0.4626", "0.4062", "0.4626", "0.4192", "0.147", "0.4192" ], [ "0.147", "0.4204", "0.461", "0.4204", "0.461", "0.434", "0.147", "0.434" ], [ "0.532", "0.4175", "0.8657", "0.4175", "0.8657", "0.4312", "0.532", "0.4312" ], [ "0.147", "0.4353", "0.461", "0.4353", "0.461", "0.449", "0.147", "0.449" ], [ "0.6304", "0.4353", "0.7866", "0.4353", "0.7866", "0.449", "0.6304", "0.449" ], [ "0.6035", "0.449", "0.784", "0.4507", "0.784", "0.4644", "0.6035", "0.462" ], [ "0.02184", "0.4382", "0.05878", "0.437", "0.0638", "0.5166", "0.02603", "0.517" ], [ "0.1486", "0.4507", "0.4592", "0.4507", "0.4592", "0.4614", "0.1486", "0.4614" ], [ "0.147", "0.4631", "0.4626", "0.4631", "0.4626", "0.4768", "0.147", "0.4768" ], [ "0.6094", "0.4631", "0.79", "0.4631", "0.79", "0.4768", "0.6094", "0.4768" ], [ "0.147", "0.478", "0.461", "0.478", "0.461", "0.4912", "0.147", "0.4912" ], [ "0.608", "0.478", "0.7305", "0.478", "0.7305", "0.4912", "0.608", "0.4912" ], [ "0.1486", "0.4934", "0.461", "0.4934", "0.461", "0.5044", "0.1486", "0.5044" ], [ "0.5728", "0.4995", "0.858", "0.4995", "0.858", "0.513", "0.5728", "0.513" ], [ "0.1453", "0.506", "0.4626", "0.506", "0.4626", "0.521", "0.1453", "0.521" ], [ "0.02184", "0.521", "0.06213", "0.521", "0.06213", "0.73", "0.02184", "0.73" ], [ "0.534", "0.5156", "0.8345", "0.5156", "0.8345", "0.53", "0.534", "0.53" ], [ "0.147", "0.5225", "0.2585", "0.5225", "0.2585", "0.5337", "0.147", "0.5337" ], [ "0.571", "0.5337", "0.8623", "0.5337", "0.8623", "0.5474", "0.571", "0.5474" ], [ "0.1436", "0.549", "0.2612", "0.549", "0.2612", "0.5625", "0.1436", "0.5625" ], [ "0.534", "0.546", "0.8364", "0.546", "0.8364", "0.5557", "0.534", "0.5557" ], [ "0.12427", "0.552", "0.1301", "0.552", "0.1301", "0.558", "0.12427", "0.558" ], [ "0.1335", "0.552", "0.1503", "0.552", "0.1503", "0.5586", "0.1335", "0.5586" ], [ "0.5356", "0.557", "0.827", "0.557", "0.827", "0.568", "0.5356", "0.568" ], [ "0.1184", "0.5723", "0.4878", "0.5723", "0.4878", "0.5874", "0.1184", "0.5874" ], [ "0.5474", "0.579", "0.8486", "0.579", "0.8486", "0.593", "0.5474", "0.593" ], [ "0.1184", "0.5903", "0.4902", "0.5903", "0.4902", "0.604", "0.1184", "0.604" ], [ "0.553", "0.5996", "0.6455", "0.5996", "0.6455", "0.611", "0.553", "0.611" ], [ "0.7446", "0.594", "0.839", "0.597", "0.837", "0.615", "0.744", "0.6123" ], [ "0.1184", "0.6055", "0.4895", "0.6055", "0.4895", "0.619", "0.1184", "0.619" ], [ "0.6533", "0.6025", "0.686", "0.6025", "0.686", "0.611", "0.6533", "0.611" ], [ "0.1184", "0.6226", "0.4902", "0.6226", "0.4902", "0.6357", "0.1184", "0.6357" ], [ "0.617", "0.6216", "0.76", "0.6216", "0.76", "0.6353", "0.617", "0.6353" ], [ "0.11755", "0.637", "0.4902", "0.6396", "0.4902", "0.653", "0.11755", "0.651" ], [ "0.617", "0.6343", "0.7676", "0.6343", "0.7676", "0.648", "0.617", "0.648" ], [ "0.6187", "0.6465", "0.764", "0.6465", "0.764", "0.6567", "0.6187", "0.6567" ], [ "0.11755", "0.655", "0.4895", "0.653", "0.4895", "0.6675", "0.11755", "0.6694" ], [ "0.6187", "0.6577", "0.7793", "0.6577", "0.7793", "0.6685", "0.6187", "0.6685" ], [ "0.11755", "0.6704", "0.4895", "0.6694", "0.4895", "0.684", "0.11755", "0.685" ], [ "0.1184", "0.687", "0.4902", "0.687", "0.4902", "0.7007", "0.1184", "0.7007" ], [ "0.5146", "0.683", "0.8833", "0.683", "0.8833", "0.6934", "0.5146", "0.6934" ], [ "0.513", "0.6953", "0.8833", "0.695", "0.8833", "0.708", "0.513", "0.709" ], [ "0.1184", "0.703", "0.4878", "0.703", "0.4878", "0.7163", "0.1184", "0.7163" ], [ "0.513", "0.7104", "0.8657", "0.7104", "0.8657", "0.7236", "0.513", "0.7236" ], [ "0.1184", "0.7197", "0.4895", "0.7197", "0.4895", "0.733", "0.1184", "0.733" ], [ "0.1184", "0.7354", "0.4895", "0.734", "0.4895", "0.749", "0.1184", "0.75" ], [ "0.12006", "0.752", "0.4895", "0.752", "0.4895", "0.7656", "0.12006", "0.7656" ], [ "0.513", "0.7505", "0.885", "0.7505", "0.885", "0.764", "0.513", "0.764" ], [ "0.1184", "0.7666", "0.471", "0.767", "0.471", "0.782", "0.1184", "0.781" ], [ "0.5146", "0.7666", "0.885", "0.7666", "0.885", "0.78", "0.5146", "0.78" ], [ "0.1368", "0.7837", "0.4895", "0.7837", "0.4895", "0.7974", "0.1368", "0.7974" ], [ "0.5146", "0.782", "0.8833", "0.782", "0.8833", "0.796", "0.5146", "0.796" ], [ "0.11755", "0.8", "0.4878", "0.8", "0.4878", "0.8135", "0.11755", "0.8135" ], [ "0.513", "0.799", "0.8833", "0.799", "0.8833", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.8164", "0.492", "0.8164", "0.492", "0.83", "0.1184", "0.83" ], [ "0.513", "0.814", "0.885", "0.814", "0.885", "0.8276", "0.513", "0.8276" ], [ "0.12006", "0.832", "0.4902", "0.832", "0.4902", "0.8457", "0.12006", "0.8457" ], [ "0.513", "0.8306", "0.885", "0.8306", "0.885", "0.844", "0.513", "0.844" ], [ "0.11755", "0.847", "0.4895", "0.8467", "0.4895", "0.861", "0.11755", "0.8623" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.513", "0.8633", "0.8833", "0.8633", "0.8833", "0.877", "0.513", "0.877" ], [ "0.1486", "0.8716", "0.4182", "0.8716", "0.4182", "0.882", "0.1486", "0.882" ], [ "0.1411", "0.882", "0.4954", "0.883", "0.4954", "0.898", "0.1411", "0.8965" ], [ "0.511", "0.88", "0.8833", "0.88", "0.8833", "0.8936", "0.511", "0.8936" ], [ "0.1184", "0.898", "0.2778", "0.898", "0.2778", "0.9077", "0.1184", "0.9077" ], [ "0.513", "0.8955", "0.876", "0.8955", "0.876", "0.909", "0.513", "0.909" ] ]

[ "OCR Post Correction for Endangered Language Texts", "Shruti Rijhwani, Antonios Anastasopoulos,2, Graham Neubig", "'Language Technologies Institute, Carnegie Mellon University", "2Department of Computer Science, George Mason University", "fsrijhwan,gneubigi@cs.cmu.edu, antonis@gmu.edu", "Abstract", "(a) Ainu (left) - Japanese (right)", "kira-an patek", "n", "There is little to no data available to build nat-", "aeyairamshitne)", "苦", "Z A 1", "ural language processing models for most en-", "7 hushkotoiwano(2)", "", "dangered languages. However, textual data", "iki-an aine", "", "in these languages often exists in formats that", "are not machine-readable, such as paper books", "(b) Griko (top) - Italian (bottom)", "\"'Iati Cini lane plon Orria", "ppiriss'emena ce ivo en itela.\"", "\"Ce tis tin ipire son bosko?", "and scanned images. In this work, we address", "the task of extracting textfrom these resources.", "\"Perche qulla era pi bella di me ed io non volevo.\"", "", "We create a benchmark dataset of transcrip-", "\"E chi I'ha portata nel bosco?\"", "tions for scanned books in three critically en-", "dangered languages and present a systematic", "(c) Yakkha (top) - Nepali (middle)- English (bottom)", "analysis of how general-purpose OCR tools", "", "", "[T's]", "are not robust to the data-scarce setting of en-", "dangered languages. We develop an OCR post-", "", "correction method tailored to ease training in", "", "this data-scarce setting, reducing the recogni-", "", "tion error rate by 34% on average across the", "1ATA", "", "three languages.", "Although my mother and sister scolded me, they loved", "Introduction", "meverymuch.Later,they convinced me of their loveand", "", "", "concern and said, “\"Children should not be disobedient.\"", "Natural language processing (NLP) systems exist", "(d) Handwrtten Shangaji - typed English glosses", "for a small fraction of the world's over 6,o00 liv-", "", "\"tikwa", "ing languages, the primary reason being the lack", "Wa", "of resources required to train and evaluate models.", "nxiuzi wa nantikwa", "Technological advances are concentrated on lan-", "mu-xuzi o-a nantikwa", "3-sauce 3-Conn la.cashew", "guages that have readily available data, and most", "A sauce ofgreen cashew nuts.", "other languages are left behind (Joshi et al., 2020)", "This is particularly notable in the case of endan-", "Figure I:Examples of scanned documents in endan-", "gered languages accompanied by translations from the", "gered languages, i.e., languages that are in danger", "same scanned book (a, b, c) or linguistic archive (d).", "of becoming extinct due to dwindling numbers of", "native speakers and the younger generations shift-", "ing to using other languages. For most endangered", "houses within the communities that speak endan-", "languages, finding any data at all is challenging", "gered languages, such as educational or cultural ma-", "In many cases, natural language text in these", "terials. Additionally, linguists documenting these", "languages does exist. However, it is locked away", "languages also create data such as word lists and", "in formats that are not machine-readable pa-", "interlinear glosses, often in the form of handwrit-", "per books, scanned images, and unstructured web", "ten notes. Examples from such scanned documents", "pages. These include books from local publishing", "are shown in Figure 1. Digitizing the textual data", "from these sources will not only enable NLP for", "t:Work done at CarnegieMellon University.", "'Code and data are available at https : / /shrutirij.", " endangered languages but also aid linguistic docu-", "github.io/ocr-el/.", "mentation, preservation, and accessibility efforts." ]

{ "arxiv_id": "2011.05402", "language": "en", "page_idx": 0 }

[ [ "0.1351", "0.076", "0.4878", "0.076", "0.4878", "0.08905", "0.1351", "0.08905" ], [ "0.5283", "0.076", "0.885", "0.076", "0.885", "0.0903", "0.5283", "0.0903" ], [ "0.12006", "0.09265", "0.4895", "0.09265", "0.4895", "0.1069", "0.12006", "0.1069" ], [ "0.5264", "0.0932", "0.7524", "0.09143", "0.7524", "0.1051", "0.5264", "0.1069" ], [ "0.12006", "0.10925", "0.4895", "0.10925", "0.4895", "0.1223", "0.12006", "0.1223" ], [ "0.513", "0.11816", "0.687", "0.12054", "0.687", "0.1354", "0.513", "0.133" ], [ "0.1184", "0.1247", "0.4902", "0.1247", "0.4902", "0.1395", "0.1184", "0.1395" ], [ "0.12006", "0.1414", "0.4937", "0.1414", "0.4937", "0.155", "0.12006", "0.155" ], [ "0.513", "0.1443", "0.882", "0.1443", "0.882", "0.158", "0.513", "0.158" ], [ "0.1184", "0.1567", "0.492", "0.1567", "0.492", "0.1704", "0.1184", "0.1704" ], [ "0.511", "0.1603", "0.882", "0.1603", "0.882", "0.1733", "0.511", "0.1733" ], [ "0.1184", "0.1729", "0.4902", "0.1729", "0.4902", "0.1859", "0.1184", "0.1859" ], [ "0.513", "0.1776", "0.881", "0.1776", "0.881", "0.1882", "0.513", "0.1882" ], [ "0.11755", "0.1882", "0.4902", "0.187", "0.4902", "0.2019", "0.11755", "0.2031" ], [ "0.513", "0.1936", "0.8833", "0.1936", "0.8833", "0.2073", "0.513", "0.2073" ], [ "0.12006", "0.2048", "0.492", "0.2048", "0.492", "0.2185", "0.12006", "0.2185" ], [ "0.51", "0.209", "0.879", "0.209", "0.879", "0.2227", "0.51", "0.2227" ], [ "0.12006", "0.2214", "0.4895", "0.2214", "0.4895", "0.2351", "0.12006", "0.2351" ], [ "0.11755", "0.2369", "0.4895", "0.2369", "0.4895", "0.2505", "0.11755", "0.2505" ], [ "0.513", "0.2351", "0.657", "0.2351", "0.657", "0.2488", "0.513", "0.2488" ], [ "0.1184", "0.2542", "0.4902", "0.2542", "0.4902", "0.2673", "0.1184", "0.2673" ], [ "0.513", "0.2559", "0.8833", "0.2559", "0.8833", "0.2695", "0.513", "0.2695" ], [ "0.1184", "0.2683", "0.492", "0.2683", "0.492", "0.282", "0.1184", "0.282" ], [ "0.511", "0.2708", "0.8833", "0.2715", "0.8833", "0.2861", "0.511", "0.2852" ], [ "0.12006", "0.2861", "0.4568", "0.2861", "0.4568", "0.2998", "0.12006", "0.2998" ], [ "0.511", "0.2874", "0.8867", "0.288", "0.8867", "0.3027", "0.511", "0.3018" ], [ "0.1377", "0.3018", "0.4895", "0.3018", "0.4895", "0.3154", "0.1377", "0.3154" ], [ "0.513", "0.3047", "0.885", "0.3047", "0.885", "0.3184", "0.513", "0.3184" ], [ "0.1184", "0.3184", "0.492", "0.3184", "0.492", "0.332", "0.1184", "0.332" ], [ "0.511", "0.3196", "0.8833", "0.3196", "0.8833", "0.3333", "0.511", "0.3333" ], [ "0.12006", "0.3342", "0.4902", "0.3342", "0.4902", "0.3474", "0.12006", "0.3474" ], [ "0.511", "0.3362", "0.882", "0.3362", "0.882", "0.3499", "0.511", "0.3499" ], [ "0.12006", "0.3499", "0.492", "0.3499", "0.492", "0.3635", "0.12006", "0.3635" ], [ "0.513", "0.3528", "0.885", "0.3528", "0.885", "0.3665", "0.513", "0.3665" ], [ "0.1184", "0.3652", "0.4895", "0.364", "0.4895", "0.379", "0.1184", "0.3801" ], [ "0.511", "0.3682", "0.8833", "0.3677", "0.8833", "0.3818", "0.511", "0.383" ], [ "0.12006", "0.3818", "0.492", "0.3818", "0.492", "0.3955", "0.12006", "0.3955" ], [ "0.513", "0.3855", "0.8833", "0.3855", "0.8833", "0.3984", "0.513", "0.3984" ], [ "0.1184", "0.3984", "0.4895", "0.3984", "0.4895", "0.412", "0.1184", "0.412" ], [ "0.513", "0.401", "0.8833", "0.401", "0.8833", "0.415", "0.513", "0.415" ], [ "0.11755", "0.4133", "0.4902", "0.4146", "0.4902", "0.4287", "0.11755", "0.4275" ], [ "0.511", "0.415", "0.8364", "0.4175", "0.8364", "0.4316", "0.511", "0.43" ], [ "0.1184", "0.4312", "0.4902", "0.4312", "0.4902", "0.4448", "0.1184", "0.4448" ], [ "0.513", "0.4407", "0.8867", "0.4412", "0.8867", "0.456", "0.513", "0.4548" ], [ "0.12006", "0.4465", "0.4902", "0.4465", "0.4902", "0.4614", "0.12006", "0.4614" ], [ "0.1184", "0.4631", "0.4048", "0.4631", "0.4048", "0.4768", "0.1184", "0.4768" ], [ "0.5146", "0.4578", "0.885", "0.4578", "0.885", "0.4714", "0.5146", "0.4714" ], [ "0.1368", "0.4785", "0.4895", "0.4785", "0.4895", "0.4922", "0.1368", "0.4922" ], [ "0.513", "0.4744", "0.8833", "0.4744", "0.8833", "0.488", "0.513", "0.488" ], [ "0.1184", "0.4946", "0.492", "0.4946", "0.492", "0.5083", "0.1184", "0.5083" ], [ "0.513", "0.4905", "0.8833", "0.4905", "0.8833", "0.5044", "0.513", "0.5044" ], [ "0.11755", "0.5103", "0.4902", "0.509", "0.4902", "0.524", "0.11755", "0.525" ], [ "0.511", "0.506", "0.8833", "0.505", "0.8833", "0.5195", "0.511", "0.521" ], [ "0.511", "0.521", "0.885", "0.5215", "0.885", "0.536", "0.511", "0.535" ], [ "0.1184", "0.527", "0.4902", "0.527", "0.4902", "0.5405", "0.1184", "0.5405" ], [ "0.11755", "0.542", "0.4895", "0.542", "0.4895", "0.5557", "0.11755", "0.5557" ], [ "0.5146", "0.538", "0.885", "0.538", "0.885", "0.552", "0.5146", "0.552" ], [ "0.1184", "0.558", "0.4895", "0.558", "0.4895", "0.572", "0.1184", "0.572" ], [ "0.5146", "0.554", "0.885", "0.554", "0.885", "0.568", "0.5146", "0.568" ], [ "0.1184", "0.576", "0.4895", "0.5737", "0.4895", "0.5874", "0.1184", "0.589" ], [ "0.5146", "0.571", "0.885", "0.571", "0.885", "0.5845", "0.5146", "0.5845" ], [ "0.1184", "0.5913", "0.4902", "0.5913", "0.4902", "0.6055", "0.1184", "0.6055" ], [ "0.513", "0.586", "0.8833", "0.586", "0.8833", "0.5996", "0.513", "0.5996" ], [ "0.12006", "0.607", "0.4895", "0.607", "0.4895", "0.6206", "0.12006", "0.6206" ], [ "0.5146", "0.6025", "0.885", "0.6025", "0.885", "0.6177", "0.5146", "0.6177" ], [ "0.12006", "0.6235", "0.4902", "0.6235", "0.4902", "0.638", "0.12006", "0.638" ], [ "0.513", "0.619", "0.729", "0.619", "0.729", "0.6343", "0.513", "0.6343" ], [ "0.12006", "0.641", "0.4878", "0.641", "0.4878", "0.652", "0.12006", "0.652" ], [ "0.602", "0.6436", "0.791", "0.641", "0.791", "0.662", "0.602", "0.6646" ], [ "0.1184", "0.6562", "0.4895", "0.6562", "0.4895", "0.6694", "0.1184", "0.6694" ], [ "0.663", "0.661", "0.679", "0.661", "0.679", "0.6714", "0.663", "0.6714" ], [ "0.1184", "0.6704", "0.4895", "0.6714", "0.4895", "0.686", "0.1184", "0.685" ], [ "0.11755", "0.686", "0.4626", "0.6885", "0.4626", "0.703", "0.11755", "0.7007" ], [ "0.5146", "0.681", "0.8833", "0.681", "0.8833", "0.6953", "0.5146", "0.6953" ], [ "0.5146", "0.6978", "0.8833", "0.6978", "0.8833", "0.712", "0.5146", "0.712" ], [ "0.1184", "0.709", "0.4844", "0.709", "0.4844", "0.7236", "0.1184", "0.7236" ], [ "0.513", "0.7153", "0.8833", "0.7153", "0.8833", "0.7285", "0.513", "0.7285" ], [ "0.12006", "0.733", "0.4902", "0.733", "0.4902", "0.7476", "0.12006", "0.7476" ], [ "0.511", "0.73", "0.885", "0.73", "0.885", "0.7437", "0.511", "0.7437" ], [ "0.1318", "0.75", "0.4937", "0.75", "0.4937", "0.763", "0.1318", "0.763" ], [ "0.513", "0.7466", "0.8867", "0.7466", "0.8867", "0.761", "0.513", "0.761" ], [ "0.1335", "0.7656", "0.2795", "0.7656", "0.2795", "0.78", "0.1335", "0.78" ], [ "0.511", "0.7627", "0.8833", "0.7627", "0.8833", "0.7754", "0.511", "0.7754" ], [ "0.5146", "0.7793", "0.882", "0.7793", "0.882", "0.792", "0.5146", "0.792" ], [ "0.1184", "0.789", "0.4895", "0.788", "0.4895", "0.8027", "0.1184", "0.804" ], [ "0.513", "0.7935", "0.885", "0.7915", "0.885", "0.8086", "0.513", "0.811" ], [ "0.1335", "0.807", "0.4878", "0.807", "0.4878", "0.821", "0.1335", "0.821" ], [ "0.1318", "0.822", "0.4634", "0.8223", "0.4634", "0.8374", "0.1318", "0.836" ], [ "0.594", "0.821", "0.7983", "0.8184", "0.7983", "0.836", "0.594", "0.8384" ], [ "0.657", "0.8384", "0.6733", "0.8384", "0.6733", "0.847", "0.657", "0.847" ], [ "0.12006", "0.8467", "0.4895", "0.8467", "0.4895", "0.861", "0.12006", "0.861" ], [ "0.1335", "0.8633", "0.492", "0.8633", "0.492", "0.877", "0.1335", "0.877" ], [ "0.5146", "0.858", "0.8286", "0.858", "0.8286", "0.8726", "0.5146", "0.8726" ], [ "0.1318", "0.8774", "0.492", "0.879", "0.492", "0.8936", "0.1318", "0.8926" ], [ "0.513", "0.8774", "0.8833", "0.879", "0.8833", "0.8936", "0.513", "0.8926" ], [ "0.1335", "0.8955", "0.4902", "0.8955", "0.4902", "0.909", "0.1335", "0.909" ], [ "0.513", "0.8955", "0.885", "0.8955", "0.885", "0.91", "0.513", "0.91" ] ]

[ "In this work, we create a benchmark dataset and", "the word error rate by 34% over a state-of-the-art", "propose a suite of methods to extract data from", "general-purpose OCR system.", "these resources, focusing on scanned images of", "2Problem Setting", "paper books containing endangered language text.", "Typically, this sort of digitization requires an opti-", "In this section, we first define the task of OCR", "cal character recognition (OCR) system. However,", "post-correction and introduce how we incorporate", "the large amounts of textual data and transcribed", "translations into the correction model. Nextwe", "images needed to train state-of-the-art OCR models", "discuss the sources from which we obtain scanned", "from scratch are unavailable in the endangered lan-", " documents containing endangered language texts.", "guage setting. Instead, we focus on post-correcting", "the output of an off-the-shelf OCR tool that can", "2.1 Formulation", "handle a variety of scripts. We show that targeted", "Optical Character Recognition OCR tools are", "methods for post-correction can significantly im-", "trained to find the best transcription corresponding", "prove performance on endangered languages.", "to the text in an image. The system typically con-", "Although OCR post-correction is relatively well-", "sists of a recognition model that produces candidate", "studied, most existing methods rely on consider-", "text sequences conditioned on the input image and", "able resources in the target language, including a", "a language model that determines the probability", "substantial amount of textual data to train a lan-", "of these sequences in the target language. We use", "guage model (Schnober et al., 2016; Dong and", "a general-purpose OCR system (detailed in Sec-", "Smith, 2018; Rigaud et al., 2019) or to create syn-", "tion 4) to produce a first pass transcription of the", "thetic data (Krishna et al., 2018). While readily", "endangered language text in the image. Let this be", "available for high-resource languages, these re-", "a sequence of characters a =[c1,..., n].", "sources are severely limited in endangered lan-", "OCR post-correction The goal of post-", "guages, preventing the direct application of existing", "post-correction methods in our setting.", "correction is to reduce recognition errors in the", "As an alternative, we present a method that", "first pass transcription often caused by low", "compounds on previous models for OCR post-", "quality scanning, physical deterioration of the", "correction, making three improvements tailored", "paper book, or diverse layouts and typefaces (Dong", "and Smith, 2018). The focus of our work is on", "to the data-scarce setting. First, we use a multi-", "source model to incorporate information from the", "using post-correction to counterbalance the lack", "high-resource translations that commonly appear in", "of OCR training data in the target endangered", "endangered languagebooks.Thesetranslations are", "languages. The correction model takes as", "usually in the lingua franca of the region (e.g., Fig-", "input and produces the final transcription of the", "ure 1 (a,b,c)) or the documentary linguist's primary", "endangered language document, a sequence of", "language (e.g., Figure 1 (d) from Devos (2019).", "characters y = [yi,..., Yk].", "Next, we introduce structural biases to ease learn-", "y = arg max Pcor(y'lα)", "ing from small amounts of data. Finally, we add", "y'", "pretraining methods to utilize the little unanno-", "tated data that exists in endangered languages.", "Incorporating translationsWe use information", "from high-resource translations of the endangered", "We summarize our main contributions as follows:", "language text. These translations are commonly", ": A benchmark dataset for OCR post-correction", "found within the same paper book or linguis-", "on three critically endangered languages: Ainu,", "tic archive (e.g., Figure I). We use an exist-", "Griko, and Yakkha.", "ing OCR system to obtain a transcription of the", "scanned translation, a sequence of characters t =", ": A systematic analysis of a general-purpose OCR", "[ti, ..., tm]. This is used to condition the model:", "system, demonstrating that it is not robust to the", "data-scarce setting of endangered languages.", "y = arg max pcorr(y'la, t)", "y", ": An OCR post-correction method that adapts the", "standard neural encoder-decoder framework to", "2.2Endangered Language Documents", "the highly under-resourced endangered language", "We explore online archives to determine how many", "setting, reducing both the character error rate and", "scanned documents in endangered languages exist" ]

{ "arxiv_id": "2011.05402", "language": "en", "page_idx": 1 }

[ [ "0.1184", "0.0766", "0.4902", "0.0766", "0.4902", "0.0903", "0.1184", "0.0903" ], [ "0.513", "0.076", "0.8833", "0.076", "0.8833", "0.0903", "0.513", "0.0903" ], [ "0.1184", "0.09265", "0.2957", "0.09265", "0.2957", "0.1057", "0.1184", "0.1057" ], [ "0.511", "0.09265", "0.8833", "0.09143", "0.8833", "0.1057", "0.511", "0.1069" ], [ "0.1368", "0.10925", "0.4937", "0.10925", "0.4937", "0.1223", "0.1368", "0.1223" ], [ "0.511", "0.1081", "0.885", "0.1081", "0.885", "0.1217", "0.511", "0.1217" ], [ "0.1184", "0.1247", "0.4902", "0.1247", "0.4902", "0.1383", "0.1184", "0.1383" ], [ "0.513", "0.1247", "0.885", "0.1247", "0.885", "0.1383", "0.513", "0.1383" ], [ "0.11755", "0.1401", "0.4902", "0.1401", "0.4902", "0.1538", "0.11755", "0.1538" ], [ "0.5146", "0.1426", "0.8833", "0.1426", "0.8833", "0.1526", "0.5146", "0.1526" ], [ "0.12006", "0.158", "0.486", "0.158", "0.486", "0.1687", "0.12006", "0.1687" ], [ "0.511", "0.1561", "0.8833", "0.1561", "0.8833", "0.1692", "0.511", "0.1692" ], [ "0.12006", "0.1733", "0.4902", "0.1733", "0.4902", "0.187", "0.12006", "0.187" ], [ "0.513", "0.1733", "0.8833", "0.1733", "0.8833", "0.187", "0.513", "0.187" ], [ "0.12006", "0.1912", "0.4895", "0.1912", "0.4895", "0.2019", "0.12006", "0.2019" ], [ "0.513", "0.1895", "0.8867", "0.1895", "0.8867", "0.2031", "0.513", "0.2031" ], [ "0.12006", "0.2073", "0.4878", "0.2073", "0.4878", "0.2173", "0.12006", "0.2173" ], [ "0.513", "0.2048", "0.885", "0.2048", "0.885", "0.2185", "0.513", "0.2185" ], [ "0.1184", "0.2214", "0.486", "0.2214", "0.486", "0.2351", "0.1184", "0.2351" ], [ "0.511", "0.2197", "0.885", "0.2203", "0.885", "0.2351", "0.511", "0.234" ], [ "0.1184", "0.2369", "0.492", "0.2369", "0.492", "0.2505", "0.1184", "0.2505" ], [ "0.513", "0.2382", "0.7656", "0.2382", "0.7656", "0.2517", "0.513", "0.2517" ], [ "0.11755", "0.2542", "0.471", "0.2517", "0.471", "0.2654", "0.11755", "0.2673" ], [ "0.532", "0.257", "0.8833", "0.257", "0.8833", "0.2708", "0.532", "0.2708" ], [ "0.511", "0.2715", "0.885", "0.2725", "0.885", "0.2874", "0.511", "0.2861" ], [ "0.12006", "0.2778", "0.4895", "0.2778", "0.4895", "0.2915", "0.12006", "0.2915" ], [ "0.511", "0.288", "0.882", "0.288", "0.882", "0.3018", "0.511", "0.3018" ], [ "0.12006", "0.2944", "0.4895", "0.2944", "0.4895", "0.308", "0.12006", "0.308" ], [ "0.12006", "0.31", "0.4878", "0.31", "0.4878", "0.3237", "0.12006", "0.3237" ], [ "0.511", "0.306", "0.8867", "0.306", "0.8867", "0.3196", "0.511", "0.3196" ], [ "0.12006", "0.326", "0.4895", "0.326", "0.4895", "0.339", "0.12006", "0.339" ], [ "0.513", "0.3218", "0.885", "0.3218", "0.885", "0.335", "0.513", "0.335" ], [ "0.511", "0.3362", "0.8833", "0.3362", "0.8833", "0.3499", "0.511", "0.3499" ], [ "0.12006", "0.3433", "0.4895", "0.3433", "0.4895", "0.354", "0.12006", "0.354" ], [ "0.1184", "0.3582", "0.4902", "0.3582", "0.4902", "0.3718", "0.1184", "0.3718" ], [ "0.513", "0.3528", "0.8833", "0.3516", "0.8833", "0.3665", "0.513", "0.3677" ], [ "0.1184", "0.3735", "0.4878", "0.3735", "0.4878", "0.3872", "0.1184", "0.3872" ], [ "0.5146", "0.3694", "0.8867", "0.3694", "0.8867", "0.383", "0.5146", "0.383" ], [ "0.1184", "0.3896", "0.4895", "0.3896", "0.4895", "0.4026", "0.1184", "0.4026" ], [ "0.5146", "0.386", "0.885", "0.386", "0.885", "0.3997", "0.5146", "0.3997" ], [ "0.12006", "0.4067", "0.4902", "0.4067", "0.4902", "0.4204", "0.12006", "0.4204" ], [ "0.5146", "0.402", "0.8833", "0.402", "0.8833", "0.415", "0.5146", "0.415" ], [ "0.1184", "0.4229", "0.4902", "0.4216", "0.4902", "0.4358", "0.1184", "0.437" ], [ "0.5146", "0.4175", "0.8833", "0.4175", "0.8833", "0.4312", "0.5146", "0.4312" ], [ "0.511", "0.4329", "0.818", "0.434", "0.818", "0.449", "0.511", "0.4478" ], [ "0.12006", "0.4407", "0.4895", "0.4407", "0.4895", "0.4507", "0.12006", "0.4507" ], [ "0.1184", "0.4548", "0.4836", "0.4548", "0.4836", "0.4697", "0.1184", "0.4697" ], [ "0.531", "0.452", "0.8833", "0.4531", "0.8833", "0.4673", "0.531", "0.466" ], [ "0.513", "0.4697", "0.8833", "0.4685", "0.8833", "0.4827", "0.513", "0.484" ], [ "0.1184", "0.4827", "0.3225", "0.4827", "0.3225", "0.4976", "0.1184", "0.4976" ], [ "0.513", "0.485", "0.882", "0.484", "0.882", "0.4988", "0.513", "0.4995" ], [ "0.5146", "0.502", "0.8833", "0.502", "0.8833", "0.5156", "0.5146", "0.5156" ], [ "0.12006", "0.5103", "0.4878", "0.5103", "0.4878", "0.524", "0.12006", "0.524" ], [ "0.513", "0.5186", "0.882", "0.5186", "0.882", "0.532", "0.513", "0.532" ], [ "0.12006", "0.527", "0.4844", "0.527", "0.4844", "0.5405", "0.12006", "0.5405" ], [ "0.513", "0.533", "0.881", "0.533", "0.881", "0.5464", "0.513", "0.5464" ], [ "0.1184", "0.542", "0.4895", "0.542", "0.4895", "0.5557", "0.1184", "0.5557" ], [ "0.5146", "0.5503", "0.882", "0.5503", "0.882", "0.564", "0.5146", "0.564" ], [ "0.11755", "0.557", "0.4895", "0.558", "0.4895", "0.5723", "0.11755", "0.572" ], [ "0.511", "0.5654", "0.882", "0.5654", "0.882", "0.579", "0.511", "0.579" ], [ "0.12006", "0.5747", "0.4895", "0.5747", "0.4895", "0.5884", "0.12006", "0.5884" ], [ "0.513", "0.5806", "0.882", "0.5806", "0.882", "0.594", "0.513", "0.594" ], [ "0.1184", "0.5903", "0.4878", "0.5903", "0.4878", "0.604", "0.1184", "0.604" ], [ "0.511", "0.5986", "0.885", "0.597", "0.885", "0.6123", "0.511", "0.6133" ], [ "0.1184", "0.607", "0.4902", "0.607", "0.4902", "0.6206", "0.1184", "0.6206" ], [ "0.513", "0.6143", "0.8228", "0.6143", "0.8228", "0.629", "0.513", "0.629" ], [ "0.11755", "0.6216", "0.4895", "0.6226", "0.4895", "0.637", "0.11755", "0.6357" ], [ "0.1184", "0.6396", "0.339", "0.6396", "0.339", "0.653", "0.1184", "0.653" ], [ "0.519", "0.648", "0.534", "0.648", "0.534", "0.652", "0.519", "0.652" ], [ "0.531", "0.6426", "0.86", "0.6426", "0.86", "0.6562", "0.531", "0.6562" ], [ "0.1351", "0.654", "0.4895", "0.654", "0.4895", "0.6675", "0.1351", "0.6675" ], [ "0.1184", "0.6714", "0.492", "0.6714", "0.492", "0.685", "0.1184", "0.685" ], [ "0.511", "0.6665", "0.8833", "0.6675", "0.8833", "0.6816", "0.511", "0.681" ], [ "0.1184", "0.687", "0.4895", "0.687", "0.4895", "0.7007", "0.1184", "0.7007" ], [ "0.513", "0.684", "0.8833", "0.684", "0.8833", "0.6987", "0.513", "0.6987" ], [ "0.1184", "0.7036", "0.492", "0.7036", "0.492", "0.7173", "0.1184", "0.7173" ], [ "0.513", "0.7007", "0.8833", "0.7007", "0.8833", "0.7144", "0.513", "0.7144" ], [ "0.1184", "0.7197", "0.492", "0.7188", "0.492", "0.733", "0.1184", "0.734" ], [ "0.511", "0.7153", "0.882", "0.7153", "0.882", "0.7285", "0.511", "0.7285" ], [ "0.1184", "0.7363", "0.4895", "0.7354", "0.4895", "0.749", "0.1184", "0.75" ], [ "0.5146", "0.7324", "0.8833", "0.7324", "0.8833", "0.7466", "0.5146", "0.7466" ], [ "0.1184", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.1184", "0.764" ], [ "0.513", "0.75", "0.7", "0.75", "0.7", "0.763", "0.513", "0.763" ], [ "0.12006", "0.7686", "0.4434", "0.7686", "0.4434", "0.782", "0.12006", "0.782" ], [ "0.5283", "0.763", "0.882", "0.763", "0.882", "0.778", "0.5283", "0.778" ], [ "0.1351", "0.786", "0.4902", "0.7876", "0.4902", "0.802", "0.1351", "0.8003" ], [ "0.513", "0.781", "0.885", "0.781", "0.885", "0.7944", "0.513", "0.7944" ], [ "0.513", "0.7974", "0.885", "0.7974", "0.885", "0.811", "0.513", "0.811" ], [ "0.12006", "0.8047", "0.4895", "0.8047", "0.4895", "0.8184", "0.12006", "0.8184" ], [ "0.5146", "0.8135", "0.885", "0.8135", "0.885", "0.8267", "0.5146", "0.8267" ], [ "0.1411", "0.83", "0.33", "0.83", "0.33", "0.8433", "0.1411", "0.8433" ], [ "0.5146", "0.83", "0.885", "0.83", "0.885", "0.8433", "0.5146", "0.8433" ], [ "0.1394", "0.843", "0.3645", "0.844", "0.3645", "0.8594", "0.1394", "0.857" ], [ "0.513", "0.8457", "0.885", "0.8457", "0.885", "0.86", "0.513", "0.86" ], [ "0.1394", "0.857", "0.3628", "0.858", "0.3628", "0.8716", "0.1394", "0.8706" ], [ "0.1411", "0.8716", "0.4895", "0.8716", "0.4895", "0.8853", "0.1411", "0.8853" ], [ "0.5356", "0.869", "0.885", "0.8706", "0.885", "0.8843", "0.5356", "0.883" ], [ "0.1184", "0.8843", "0.492", "0.8843", "0.492", "0.898", "0.1184", "0.898" ], [ "0.511", "0.882", "0.8364", "0.883", "0.8364", "0.8945", "0.511", "0.8936" ], [ "0.12006", "0.8965", "0.451", "0.8965", "0.451", "0.91", "0.12006", "0.91" ], [ "0.534", "0.8945", "0.827", "0.8955", "0.827", "0.909", "0.534", "0.9077" ] ]

[ "as potential sources for data extraction (as of this", "a language isolate. In our dataset, we use a book", "writing,October 2020).", "of Ainu epic poetry (yukara), with the “Kutune", "The Internet Archive, a general-purpose archive", "Shirka\" yukara (Kindaichi, 1931) in Ainu tran-", "of web content, has scanned books labeled with the", "scribed in Latin script. Each page in the book", "language of their content. We find 11,674 books la-", "hasatwo-column structure-theleftcolumnhas", "beled withlanguages classified as\"endangered\"by", "the Ainu text, and the right has its Japanese trans-", "UNESCO. Additionally, we find that endangered", "lation already aligned at the line-level, removing", "languagelinguisticarchivescontainthousandsof", "the need for manual alignment (see Figure 1 (a)).", "documentsinPDFformattheArchiveofthe", "The book has 338 pages in total. Given the effort", "Indigenous Languages of Latin America (AILLA)", "involved in annotation, we transcribe the Ainu text", "contains 10,000 such documents and the Endan-", "from 33 pages, totaling 816 lines.", "gered LanguagesArchive(ELAR)4has7.000.", "Griko is an endangered Greek dialect spoken in", " southern Italy. The language uses a combination", "How common are translations? As described in", "of the Latin alphabet and the Greek alphabet as its", "the introduction, endangered language documents", "often contain a translation into another (usually", "writing system. The document we use is a book of", "high-resource) language. While it is difficult to es-", "Griko folk tales compiled by Stomeo (1980). The", "book is structured such that in each fold of two", "timatethenumber of documents withtranslations,", "multilingual documents represent the majority in", "pages, the left page has Griko text, and the right", "the archives we examined: AILLA contains 4.383", "page has the corresponding translation in Italian.", "PDFs with bilingual text and 1,246 PDFs with trilin-", "Of the 175 pages in the book, we annotate the", "gual text, while ELAR contains ~5,000 multilin-", "Griko text from 33 pages and manually align it at", "gual documents. The structure of translations in", "the sentence-level to the Italian translation. This", "results in 807 annotated Griko sentences.", "these documents is varied,from dictionaries and", "interlinear glosses to scanned multilingual books.", "Yakkha is an endangered Sino-Tibetan language", "spoken in Nepal. It uses the Devanagari writing", "3Benchmark Dataset", "system.We use scanned images of three chil", "dren's books, each of which has a story written", "From the sources described above, we select docu", "in Yakkha along with its translation in Nepali and", "ments from three critically endangered languages", "English (Schackow,2012).We manually transcribe", "for annotation Ainu, Griko, and Yakkha. These", "the Yakkha text from all three books. We also align", "languages were chosen in an effort to create a ge-", "the Yakkha text to both the Nepali and the English", "ographically, typologically, and orthographically", "OCR at the sentence level with the help of an exist", "diverse benchmark. We focus this initial study", "ing Yakkha dictionary (Schackow, 2015). In total,", "on scanned images of printed books as opposed", "we have 159 annotated Yakkha sentences.", "to handwritten notes, which are a relatively more", "challenging domain for OCR.", "", "OCR Systems: Promises and Pitfalls", "We manually transcribed the text correspond-", "ing to the endangered language content. The text", "As briefly alluded to in the introduction, training an", "corresponding to the translations is not manually", "OCR model for each endangered language is chal-", "transcribed. We also aligned the endangered lan-", "lenging, given the limited available data. Instead.", "guage text to the OCR output on the translations,", "we use the general-purpose OCR system from the", "per the formulation in Section 2.1. We describe the", "Google Vision AI toolkit' to get the first pass OCR", "annotated documents below and example images", "transcription on our data.", "from our dataset are in Figure 1 (a), (b), (c).", "The Google Vision OCR system (Fujli et al.", "Ainu is a severely endangered indigenous lan-", "2017; Ingle et al., 2019) is highly performant and", "supports 60 major languages in 29 scripts. It can", "guage from northern Japan, typically considered", "transcribe a wide range of higher resource lan-", "\"https://archive.org/", "guages with high accuracy, ideal for our proposed", "3https://ailla.utexas.org", "method of incorporating high-resource translations", "*https://elar.soas.ac.uk/", "'UNESCO defines critically endangered languages as", "‘Some transcriptions of Ainu also use the Katakana script.", "those where the youngest speakers are grandparents and older,", "See Howell (1951)for a discussion on Ainufolklore.", "and they speak the language partially and infrequently.", "7https://cloud.google.com/vision" ]

{ "arxiv_id": "2011.05402", "language": "en", "page_idx": 2 }

[ [ "0.5283", "0.076", "0.6675", "0.076", "0.6675", "0.0932", "0.5283", "0.0932" ], [ "0.738", "0.0748", "0.8716", "0.076", "0.8716", "0.0903", "0.738", "0.08905" ], [ "0.3105", "0.079", "0.3435", "0.079", "0.3435", "0.09143", "0.3105", "0.09143" ], [ "0.2065", "0.0766", "0.2747", "0.079", "0.2747", "0.0932", "0.2065", "0.09143" ], [ "0.3594", "0.079", "0.398", "0.079", "0.398", "0.0903", "0.3594", "0.0903" ], [ "0.2091", "0.0998", "0.2468", "0.0998", "0.2468", "0.11224", "0.2091", "0.11224" ], [ "0.3132", "0.0986", "0.345", "0.0986", "0.345", "0.111", "0.3132", "0.111" ], [ "0.3628", "0.0998", "0.398", "0.0998", "0.398", "0.11224", "0.3628", "0.11224" ], [ "0.5474", "0.1057", "0.6543", "0.1057", "0.6543", "0.1247", "0.5474", "0.1247" ], [ "0.6963", "0.1051", "0.729", "0.1069", "0.727", "0.11816", "0.6943", "0.1164" ], [ "0.7617", "0.1069", "0.852", "0.111", "0.8506", "0.1259", "0.76", "0.1217" ], [ "0.2091", "0.114", "0.251", "0.114", "0.251", "0.1259", "0.2091", "0.1259" ], [ "0.3132", "0.1134", "0.345", "0.1134", "0.345", "0.1259", "0.3132", "0.1259" ], [ "0.3552", "0.114", "0.3972", "0.114", "0.3972", "0.1259", "0.3552", "0.1259" ], [ "0.2091", "0.1271", "0.2627", "0.1271", "0.2627", "0.1414", "0.2091", "0.1414" ], [ "0.3123", "0.1277", "0.345", "0.1277", "0.345", "0.1401", "0.3123", "0.1401" ], [ "0.3535", "0.1247", "0.4014", "0.1271", "0.3997", "0.1426", "0.3525", "0.1401" ], [ "0.5146", "0.1426", "0.885", "0.1426", "0.885", "0.1561", "0.5146", "0.1561" ], [ "0.1184", "0.155", "0.4895", "0.155", "0.4895", "0.1687", "0.1184", "0.1687" ], [ "0.5146", "0.1592", "0.8228", "0.1592", "0.8228", "0.1692", "0.5146", "0.1692" ], [ "0.1184", "0.1716", "0.4358", "0.1716", "0.4358", "0.1853", "0.1184", "0.1853" ], [ "0.511", "0.1924", "0.684", "0.1936", "0.684", "0.2114", "0.511", "0.2102" ], [ "0.11755", "0.212", "0.4954", "0.2131", "0.4954", "0.228", "0.11755", "0.2268" ], [ "0.511", "0.2156", "0.8833", "0.2162", "0.8833", "0.231", "0.511", "0.2299" ], [ "0.11755", "0.228", "0.4902", "0.2286", "0.4902", "0.2434", "0.11755", "0.2423" ], [ "0.511", "0.2328", "0.882", "0.2328", "0.882", "0.2465", "0.511", "0.2465" ], [ "0.1184", "0.2452", "0.4902", "0.2452", "0.4902", "0.2588", "0.1184", "0.2588" ], [ "0.511", "0.2488", "0.8833", "0.2477", "0.8833", "0.2625", "0.511", "0.2632" ], [ "0.11755", "0.2612", "0.4902", "0.26", "0.4902", "0.275", "0.11755", "0.2756" ], [ "0.513", "0.2654", "0.8833", "0.2654", "0.8833", "0.279", "0.513", "0.279" ], [ "0.12006", "0.279", "0.4895", "0.279", "0.4895", "0.2893", "0.12006", "0.2893" ], [ "0.513", "0.2808", "0.885", "0.2808", "0.885", "0.2944", "0.513", "0.2944" ], [ "0.1184", "0.2944", "0.4902", "0.2944", "0.4902", "0.308", "0.1184", "0.308" ], [ "0.513", "0.2976", "0.885", "0.2976", "0.885", "0.3113", "0.513", "0.3113" ], [ "0.1184", "0.31", "0.4902", "0.3088", "0.4902", "0.3237", "0.1184", "0.3247" ], [ "0.513", "0.3142", "0.885", "0.3142", "0.885", "0.328", "0.513", "0.328" ], [ "0.12006", "0.326", "0.4895", "0.326", "0.4895", "0.3403", "0.12006", "0.3403" ], [ "0.513", "0.33", "0.885", "0.33", "0.885", "0.3433", "0.513", "0.3433" ], [ "0.1184", "0.3425", "0.4902", "0.3425", "0.4902", "0.3557", "0.1184", "0.3557" ], [ "0.513", "0.3467", "0.8867", "0.3467", "0.8867", "0.3599", "0.513", "0.3599" ], [ "0.11755", "0.3582", "0.492", "0.357", "0.492", "0.3718", "0.11755", "0.3723" ], [ "0.513", "0.3623", "0.8237", "0.3623", "0.8237", "0.376", "0.513", "0.376" ], [ "0.1184", "0.3748", "0.4895", "0.3748", "0.4895", "0.3884", "0.1184", "0.3884" ], [ "0.513", "0.383", "0.8833", "0.3843", "0.8833", "0.3984", "0.513", "0.398" ], [ "0.1184", "0.3901", "0.4265", "0.3901", "0.4265", "0.405", "0.1184", "0.405" ], [ "0.1351", "0.4062", "0.4895", "0.4067", "0.4895", "0.4216", "0.1351", "0.4204" ], [ "0.5283", "0.402", "0.8833", "0.402", "0.8833", "0.415", "0.5283", "0.415" ], [ "0.5264", "0.4175", "0.882", "0.4175", "0.882", "0.4312", "0.5264", "0.4312" ], [ "0.1184", "0.4233", "0.4878", "0.4233", "0.4878", "0.437", "0.1184", "0.437" ], [ "0.5283", "0.4329", "0.8833", "0.4329", "0.8833", "0.4478", "0.5283", "0.4478" ], [ "0.11755", "0.4395", "0.4878", "0.4395", "0.4878", "0.4531", "0.11755", "0.4531" ], [ "0.5283", "0.4495", "0.8833", "0.4495", "0.8833", "0.4631", "0.5283", "0.4631" ], [ "0.11755", "0.456", "0.4878", "0.456", "0.4878", "0.4697", "0.11755", "0.4697" ], [ "0.5283", "0.4656", "0.8833", "0.4656", "0.8833", "0.4785", "0.5283", "0.4785" ], [ "0.11755", "0.4714", "0.4895", "0.4714", "0.4895", "0.485", "0.11755", "0.485" ], [ "0.5283", "0.481", "0.885", "0.481", "0.885", "0.4954", "0.5283", "0.4954" ], [ "0.1184", "0.4868", "0.4895", "0.4868", "0.4895", "0.5005", "0.1184", "0.5005" ], [ "0.529", "0.4976", "0.8833", "0.4976", "0.8833", "0.511", "0.529", "0.511" ], [ "0.12006", "0.5044", "0.486", "0.5044", "0.486", "0.517", "0.12006", "0.517" ], [ "0.5283", "0.514", "0.8833", "0.514", "0.8833", "0.528", "0.5283", "0.528" ], [ "0.1184", "0.532", "0.3105", "0.532", "0.3105", "0.5464", "0.1184", "0.5464" ], [ "0.529", "0.53", "0.885", "0.53", "0.885", "0.5435", "0.529", "0.5435" ], [ "0.5283", "0.546", "0.8833", "0.546", "0.8833", "0.56", "0.5283", "0.56" ], [ "0.12006", "0.5557", "0.4895", "0.5557", "0.4895", "0.5693", "0.12006", "0.5693" ], [ "0.5283", "0.5625", "0.8833", "0.5625", "0.8833", "0.576", "0.5283", "0.576" ], [ "0.1184", "0.5723", "0.4895", "0.5723", "0.4895", "0.586", "0.1184", "0.586" ], [ "0.5264", "0.5767", "0.882", "0.5767", "0.882", "0.5913", "0.5264", "0.5913" ], [ "0.11755", "0.5884", "0.4878", "0.5884", "0.4878", "0.6016", "0.11755", "0.6016" ], [ "0.5264", "0.593", "0.882", "0.593", "0.882", "0.6055", "0.5264", "0.6055" ], [ "0.1184", "0.605", "0.4895", "0.605", "0.4895", "0.618", "0.1184", "0.618" ], [ "0.5283", "0.61", "0.8037", "0.61", "0.8037", "0.6235", "0.5283", "0.6235" ], [ "0.11755", "0.619", "0.4895", "0.6206", "0.4895", "0.6353", "0.11755", "0.6343" ], [ "0.12006", "0.6357", "0.4895", "0.6357", "0.4895", "0.651", "0.12006", "0.651" ], [ "0.5156", "0.633", "0.8833", "0.633", "0.8833", "0.6465", "0.5156", "0.6465" ], [ "0.12006", "0.653", "0.4895", "0.653", "0.4895", "0.6665", "0.12006", "0.6665" ], [ "0.5264", "0.6494", "0.8833", "0.6484", "0.8833", "0.663", "0.5264", "0.6646" ], [ "0.1184", "0.6685", "0.4895", "0.6685", "0.4895", "0.6816", "0.1184", "0.6816" ], [ "0.5283", "0.6665", "0.8833", "0.6665", "0.8833", "0.6797", "0.5283", "0.6797" ], [ "0.12006", "0.685", "0.4895", "0.685", "0.4895", "0.6987", "0.12006", "0.6987" ], [ "0.5264", "0.681", "0.8833", "0.681", "0.8833", "0.695", "0.5264", "0.695" ], [ "0.1184", "0.6997", "0.4055", "0.7007", "0.4055", "0.7153", "0.1184", "0.7144" ], [ "0.5283", "0.6978", "0.885", "0.6978", "0.885", "0.7114", "0.5283", "0.7114" ], [ "0.1377", "0.7173", "0.4878", "0.7173", "0.4878", "0.731", "0.1377", "0.731" ], [ "0.5283", "0.7144", "0.8867", "0.7144", "0.8867", "0.728", "0.5283", "0.728" ], [ "0.12006", "0.734", "0.4895", "0.734", "0.4895", "0.7476", "0.12006", "0.7476" ], [ "0.5283", "0.73", "0.8867", "0.73", "0.8867", "0.7437", "0.5283", "0.7437" ], [ "0.12006", "0.7505", "0.4895", "0.7505", "0.4895", "0.764", "0.12006", "0.764" ], [ "0.5283", "0.7466", "0.8833", "0.7466", "0.8833", "0.7573", "0.5283", "0.7573" ], [ "0.1184", "0.7666", "0.4895", "0.7656", "0.4895", "0.7793", "0.1184", "0.78" ], [ "0.529", "0.7627", "0.8833", "0.7627", "0.8833", "0.7754", "0.529", "0.7754" ], [ "0.12006", "0.782", "0.4895", "0.782", "0.4895", "0.796", "0.12006", "0.796" ], [ "0.5283", "0.777", "0.885", "0.778", "0.885", "0.792", "0.5283", "0.7915" ], [ "0.12006", "0.799", "0.4902", "0.799", "0.4902", "0.8125", "0.12006", "0.8125" ], [ "0.5264", "0.7935", "0.885", "0.7944", "0.885", "0.8086", "0.5264", "0.808" ], [ "0.1184", "0.814", "0.4895", "0.814", "0.4895", "0.8276", "0.1184", "0.8276" ], [ "0.5283", "0.811", "0.858", "0.811", "0.858", "0.8247", "0.5283", "0.8247" ], [ "0.11755", "0.83", "0.4902", "0.8306", "0.4902", "0.8457", "0.11755", "0.844" ], [ "0.5156", "0.8403", "0.531", "0.8403", "0.531", "0.8486", "0.5156", "0.8486" ], [ "0.538", "0.8384", "0.7925", "0.8384", "0.7925", "0.8516", "0.538", "0.8516" ], [ "0.1184", "0.847", "0.4902", "0.847", "0.4902", "0.861", "0.1184", "0.861" ], [ "0.12006", "0.8643", "0.492", "0.8643", "0.492", "0.8745", "0.12006", "0.8745" ], [ "0.513", "0.8623", "0.882", "0.8623", "0.882", "0.877", "0.513", "0.877" ], [ "0.11755", "0.8774", "0.4937", "0.879", "0.4937", "0.8936", "0.11755", "0.8926" ], [ "0.511", "0.879", "0.885", "0.8774", "0.885", "0.8926", "0.511", "0.8936" ], [ "0.11755", "0.8936", "0.471", "0.8955", "0.471", "0.909", "0.11755", "0.9067" ], [ "0.511", "0.8945", "0.885", "0.8955", "0.885", "0.909", "0.511", "0.9077" ] ]

[ "exi i kaddinara", "exiikaddinara", "CER", "Language", "WER", "Ainu", "1.34", "6.27", "", "OCR", "RHS", "Griko", "3.27", "15.63", "Yakkha", "8.90", "31.64", "Figure 2: Examples of errors in Griko (top) and Yakkha", "Table I: Character error rate and word error rate with", "(bottom)whenusingtheGoogleVisionOCR.", "the Google Vision OCR system on our dataset.", "4.2Types of Errors", "into the post-correction model. Moreover, it is par-", "To better understand the challenges posed by the", "ticularly well-suited to our task because it provides", "endangered language setting, we manually inspect", "script-specific OCR models in addition to language-", "all the errors made by the OCR system. While", "specific ones. Per-script models are more robust", "some errors are commonly seen in the OCR task.", "to unknownlanguages because they aretrained", "such as misidentified punctuation or incorrect word", "on data from multiple languages and can act as a", "boundaries, 85% of the total errors occur due to", "general character recognizer without relying on a", "specific characteristics of endangered languages", "single language's model. Since most endangered", "that general-purpose OCR systems do not account", "languages adopt standard scripts (often from the", "for. Broadly, they can be categorized into two types,", "region's dominant language) as their writing sys-", "examples of which are shown in Figure 2:", "tems, the per-script recognition models can provide", " Mixed scripts The existing scripts that most", "a stable starting point for post-correction.", "The metrics we use for evaluating performance", "endangered languages adopt as writing systems", "are often not ideal for comprehensively represent", "are character error rate (CER) and word error rate", "ing the language. For example, the Devanagari", "(WER), representing the ratio of erroneous char-", "script does not have a grapheme for the glottal", "acters or words in the OCR prediction to the total", "stop - as a solution, printed texts in the Yakkha", "number in the annotated transcription. More de-", "language use the IPA symbol ?'(Schackow,", "tails are in Section 6. The CER and WER using the", "2015). Similarly, both Greek and Latin charac-", "Google Vision OCR on our dataset are in Table 1.", "ters are used to write Griko. The Google Vision", "4.1 OCRPerformance", "OCR is trained to detect script at the line-level", "and is not equipped to handle multiple scripts", "Across the three languages, the error rates indicate", "within a single word. As seen in Figure 2, the", "that we have a first pass transcription that is of rea-", "system does not recognize the Greek character x", "sonable quality,giving our post-correction method", "in Griko and the IPA symbol ? in Yakkha. Mixed", "a reliable starting point. We note the particularly", "scripts cause 11% of the OCR errors.", "low CER for the Ainu data, reflecting previous", "work that has evaluated the Google Vision system", ". Uncommon characters and diacritics En-", "to have strong performance on typed Latin script", "dangered languages often use graphemes and di-", "documents (Fujii et al., 2017). However, there re-", "acritics that are part of the standard script but are", "mains considerable room for improvement in both", "not commonly seen in high-resource languages", "CER and WER for all three languages.", "Since these are likely rare in the OCR system's", "Next, we look at the edit distance between the", "training data, they are frequently misidentified,", "predicted and the gold transcriptions, in terms of", "accounting for 74% of the errors. In Figure 2,", "insertion, deletion, and replacement operations. Re-", "we seethatthe OCRsvstem substitutestheun-", "placement accounts for over84%of the errors in", "common diacritic d in Griko. The system also", "the Griko and Ainu datasets, and 55% overall. This", "deletes the Yakkha character , which is a half", "pattern is expected in the OCR task, as the recogni-", "form' alphabet that is infrequent in several other", "tion model uses the image to make predictions and", "Devanagari script languages (such as Hindi).", "is more likely to confuse a character's shape for an-", "5", " OCR Post-Correction Model", "other than to hallucinate or erase pixels. However,", "we observe that the errors in the Yakkha dataset do", "In this section, we describe our proposed OCR", "not follow this pattern. Instead, 87% of the errors", "post-correction model. The base architecture of", "for Yakkha occur because of deleted characters.", "the model is a multi-source sequence-to-sequence" ]

{ "arxiv_id": "2011.05402", "language": "en", "page_idx": 3 }

[ [ "0.2612", "0.0748", "0.3494", "0.0766", "0.3494", "0.09143", "0.2612", "0.08905" ], [ "0.5146", "0.0766", "0.8833", "0.0766", "0.8833", "0.0903", "0.5146", "0.0903" ], [ "0.3032", "0.0932", "0.3594", "0.0956", "0.3586", "0.111", "0.3015", "0.10925" ], [ "0.5146", "0.09265", "0.8833", "0.09265", "0.8833", "0.1069", "0.5146", "0.1069" ], [ "0.513", "0.1081", "0.882", "0.1081", "0.882", "0.1217", "0.513", "0.1217" ], [ "0.2737", "0.114", "0.336", "0.114", "0.336", "0.1289", "0.2737", "0.1289" ], [ "0.5146", "0.1247", "0.8833", "0.1247", "0.8833", "0.1395", "0.5146", "0.1395" ], [ "0.3032", "0.133", "0.3586", "0.133", "0.3586", "0.1478", "0.3032", "0.1478" ], [ "0.5146", "0.1401", "0.8833", "0.1401", "0.8833", "0.155", "0.5146", "0.155" ], [ "0.2737", "0.152", "0.336", "0.152", "0.336", "0.1663", "0.2737", "0.1663" ], [ "0.511", "0.1561", "0.882", "0.1561", "0.882", "0.1692", "0.511", "0.1692" ], [ "0.2065", "0.1687", "0.2585", "0.1687", "0.2585", "0.1799", "0.2065", "0.1799" ], [ "0.3525", "0.1687", "0.4072", "0.1687", "0.4072", "0.1799", "0.3525", "0.1799" ], [ "0.513", "0.1729", "0.885", "0.1729", "0.885", "0.1859", "0.513", "0.1859" ], [ "0.2107", "0.1859", "0.2737", "0.1859", "0.2737", "0.2007", "0.2107", "0.2007" ], [ "0.3342", "0.1859", "0.398", "0.1859", "0.398", "0.2019", "0.3342", "0.2019" ], [ "0.513", "0.1895", "0.7656", "0.1895", "0.7656", "0.2031", "0.513", "0.2031" ], [ "0.2393", "0.2037", "0.297", "0.2037", "0.297", "0.2197", "0.2393", "0.2197" ], [ "0.3645", "0.2048", "0.4207", "0.2048", "0.4207", "0.2197", "0.3645", "0.2197" ], [ "0.6455", "0.2131", "0.7515", "0.2131", "0.7515", "0.2351", "0.6455", "0.2351" ], [ "0.215", "0.2245", "0.2703", "0.2245", "0.2703", "0.2351", "0.215", "0.2351" ], [ "0.339", "0.2239", "0.3955", "0.2239", "0.3955", "0.2382", "0.339", "0.2382" ], [ "0.2393", "0.2411", "0.2778", "0.2411", "0.2778", "0.257", "0.2393", "0.257" ], [ "0.3628", "0.2411", "0.4014", "0.2411", "0.4014", "0.257", "0.3628", "0.257" ], [ "0.511", "0.2446", "0.882", "0.2446", "0.882", "0.2588", "0.511", "0.2588" ], [ "0.3132", "0.2612", "0.424", "0.2612", "0.424", "0.2756", "0.3132", "0.2756" ], [ "0.5146", "0.2612", "0.885", "0.2612", "0.885", "0.275", "0.5146", "0.275" ], [ "0.511", "0.2756", "0.885", "0.2766", "0.885", "0.2915", "0.511", "0.2903" ], [ "0.2007", "0.2625", "0.2854", "0.2625", "0.2854", "0.2766", "0.2007", "0.2766" ], [ "0.2275", "0.279", "0.2612", "0.279", "0.2612", "0.2915", "0.2275", "0.2915" ], [ "0.3376", "0.2766", "0.3955", "0.275", "0.3972", "0.2893", "0.339", "0.2915" ], [ "0.513", "0.2944", "0.721", "0.2944", "0.721", "0.308", "0.513", "0.308" ], [ "0.1184", "0.3113", "0.4895", "0.3113", "0.4895", "0.3247", "0.1184", "0.3247" ], [ "0.613", "0.3184", "0.818", "0.3206", "0.818", "0.3384", "0.613", "0.3362" ], [ "0.1184", "0.326", "0.4895", "0.326", "0.4895", "0.339", "0.1184", "0.339" ], [ "0.8564", "0.3218", "0.8833", "0.3218", "0.8833", "0.335", "0.8564", "0.335" ], [ "0.11755", "0.3384", "0.4878", "0.3384", "0.4878", "0.3516", "0.11755", "0.3516" ], [ "0.5825", "0.339", "0.812", "0.339", "0.812", "0.354", "0.5825", "0.354" ], [ "0.8564", "0.339", "0.8867", "0.339", "0.8867", "0.3557", "0.8564", "0.3557" ], [ "0.1184", "0.3528", "0.4878", "0.3528", "0.4878", "0.3665", "0.1184", "0.3665" ], [ "0.12006", "0.3682", "0.4895", "0.3682", "0.4895", "0.3818", "0.12006", "0.3818" ], [ "0.5146", "0.3682", "0.8833", "0.3682", "0.8833", "0.3818", "0.5146", "0.3818" ], [ "0.11755", "0.3813", "0.4902", "0.3818", "0.4902", "0.3955", "0.11755", "0.3943" ], [ "0.511", "0.383", "0.8833", "0.3855", "0.8833", "0.3997", "0.511", "0.398" ], [ "0.11755", "0.3967", "0.3535", "0.3967", "0.3535", "0.4104", "0.11755", "0.4104" ], [ "0.5146", "0.401", "0.885", "0.401", "0.885", "0.4146", "0.5146", "0.4146" ], [ "0.5146", "0.4175", "0.797", "0.4175", "0.797", "0.4312", "0.5146", "0.4312" ], [ "0.12006", "0.434", "0.4895", "0.434", "0.4895", "0.4478", "0.12006", "0.4478" ], [ "0.5146", "0.4436", "0.833", "0.4436", "0.833", "0.4573", "0.5146", "0.4573" ], [ "0.12006", "0.4507", "0.4895", "0.4507", "0.4895", "0.4644", "0.12006", "0.4644" ], [ "0.12006", "0.4673", "0.4895", "0.4673", "0.4895", "0.481", "0.12006", "0.481" ], [ "0.511", "0.4644", "0.882", "0.4644", "0.882", "0.478", "0.511", "0.478" ], [ "0.1184", "0.4827", "0.4895", "0.4822", "0.4895", "0.4963", "0.1184", "0.4976" ], [ "0.513", "0.481", "0.885", "0.481", "0.885", "0.4946", "0.513", "0.4946" ], [ "0.12006", "0.4995", "0.4902", "0.4995", "0.4902", "0.513", "0.12006", "0.513" ], [ "0.511", "0.4954", "0.885", "0.4963", "0.885", "0.511", "0.511", "0.5103" ], [ "0.1184", "0.5166", "0.4072", "0.5166", "0.4072", "0.53", "0.1184", "0.53" ], [ "0.511", "0.5127", "0.885", "0.511", "0.885", "0.5254", "0.511", "0.527" ], [ "0.513", "0.5293", "0.882", "0.5293", "0.882", "0.542", "0.513", "0.542" ], [ "0.12006", "0.5415", "0.367", "0.5415", "0.367", "0.5547", "0.12006", "0.5547" ], [ "0.5146", "0.5444", "0.8833", "0.5444", "0.8833", "0.558", "0.5146", "0.558" ], [ "0.1184", "0.561", "0.4902", "0.5625", "0.4902", "0.5767", "0.1184", "0.576" ], [ "0.513", "0.561", "0.8833", "0.561", "0.8833", "0.5747", "0.513", "0.5747" ], [ "0.12006", "0.579", "0.4902", "0.579", "0.4902", "0.593", "0.12006", "0.593" ], [ "0.513", "0.5767", "0.885", "0.576", "0.885", "0.5903", "0.513", "0.5913" ], [ "0.1184", "0.5933", "0.4895", "0.594", "0.4895", "0.6094", "0.1184", "0.608" ], [ "0.5146", "0.6016", "0.8833", "0.6016", "0.8833", "0.615", "0.5146", "0.615" ], [ "0.12006", "0.611", "0.4895", "0.611", "0.4895", "0.6245", "0.12006", "0.6245" ], [ "0.511", "0.616", "0.881", "0.616", "0.881", "0.63", "0.511", "0.63" ], [ "0.1184", "0.627", "0.4475", "0.627", "0.4475", "0.64", "0.1184", "0.64" ], [ "0.513", "0.6343", "0.8833", "0.6343", "0.8833", "0.648", "0.513", "0.648" ], [ "0.1377", "0.641", "0.4114", "0.641", "0.4114", "0.655", "0.1377", "0.655" ], [ "0.4055", "0.6445", "0.4895", "0.6445", "0.4895", "0.655", "0.4055", "0.655" ], [ "0.513", "0.6494", "0.8833", "0.6494", "0.8833", "0.663", "0.513", "0.663" ], [ "0.1184", "0.6577", "0.4895", "0.6577", "0.4895", "0.6714", "0.1184", "0.6714" ], [ "0.513", "0.6665", "0.8833", "0.6665", "0.8833", "0.6797", "0.513", "0.6797" ], [ "0.11755", "0.6733", "0.4895", "0.6733", "0.4895", "0.687", "0.11755", "0.687" ], [ "0.513", "0.683", "0.882", "0.683", "0.882", "0.697", "0.513", "0.697" ], [ "0.1184", "0.6914", "0.4895", "0.6914", "0.4895", "0.705", "0.1184", "0.705" ], [ "0.513", "0.6987", "0.8833", "0.6987", "0.8833", "0.712", "0.513", "0.712" ], [ "0.11755", "0.707", "0.4902", "0.706", "0.4902", "0.72", "0.11755", "0.7217" ], [ "0.1184", "0.7217", "0.4895", "0.7227", "0.4895", "0.737", "0.1184", "0.7363" ], [ "0.5156", "0.7163", "0.882", "0.7163", "0.882", "0.727", "0.5156", "0.727" ], [ "0.5146", "0.73", "0.885", "0.73", "0.885", "0.7446", "0.5146", "0.7446" ], [ "0.1184", "0.7393", "0.3105", "0.7393", "0.3105", "0.753", "0.1184", "0.753" ], [ "0.531", "0.745", "0.8833", "0.745", "0.8833", "0.759", "0.531", "0.759" ], [ "0.1368", "0.755", "0.4878", "0.755", "0.4878", "0.7695", "0.1368", "0.7695" ], [ "0.511", "0.761", "0.8833", "0.7627", "0.8833", "0.777", "0.511", "0.7754" ], [ "0.1184", "0.7715", "0.4895", "0.7715", "0.4895", "0.786", "0.1184", "0.786" ], [ "0.513", "0.7793", "0.8833", "0.7793", "0.8833", "0.792", "0.513", "0.792" ], [ "0.1184", "0.7876", "0.4895", "0.7876", "0.4895", "0.8003", "0.1184", "0.8003" ], [ "0.513", "0.7935", "0.8833", "0.7944", "0.8833", "0.8086", "0.513", "0.808" ], [ "0.1184", "0.804", "0.4902", "0.804", "0.4902", "0.818", "0.1184", "0.818" ], [ "0.5156", "0.8125", "0.8833", "0.8125", "0.8833", "0.8223", "0.5156", "0.8223" ], [ "0.12006", "0.8193", "0.3855", "0.8193", "0.3855", "0.833", "0.12006", "0.833" ], [ "0.513", "0.8267", "0.885", "0.8267", "0.885", "0.8403", "0.513", "0.8403" ], [ "0.1704", "0.8433", "0.4382", "0.8433", "0.4382", "0.861", "0.1704", "0.861" ], [ "0.4534", "0.847", "0.4878", "0.847", "0.4878", "0.8594", "0.4534", "0.8594" ], [ "0.511", "0.8433", "0.869", "0.8433", "0.869", "0.857", "0.511", "0.857" ], [ "0.2275", "0.865", "0.3787", "0.8677", "0.3787", "0.8896", "0.2275", "0.887" ], [ "0.7583", "0.8706", "0.7734", "0.8706", "0.7734", "0.88", "0.7583", "0.88" ], [ "0.574", "0.8813", "0.617", "0.8857", "0.613", "0.902", "0.5693", "0.8965" ], [ "0.692", "0.8857", "0.7466", "0.883", "0.748", "0.8984", "0.693", "0.902" ], [ "0.756", "0.887", "0.7676", "0.887", "0.7676", "0.891", "0.756", "0.891" ], [ "0.7866", "0.8853", "0.812", "0.8896", "0.8096", "0.8994", "0.784", "0.8955" ], [ "0.2377", "0.8926", "0.3687", "0.8926", "0.3687", "0.91", "0.2377", "0.91" ], [ "0.669", "0.902", "0.69", "0.902", "0.69", "0.909", "0.669", "0.909" ], [ "0.7446", "0.898", "0.785", "0.8994", "0.784", "0.912", "0.744", "0.91" ] ]

[ "P(yi...yk)", "where Sk-i is the decoder state of the previous time", " softmax", "step and v, W and W2 are trainable parameters.", "The encoder hidden states h\" are weighted by the", "S1...SK", "attention distribution Q to produce the context", "decoder", "vector c. We follow a similar procedure for the", "C....CK", "second encoder to produce ct.We concatenate", "attention", "attention", "the context vectors to combine attention from both", "h'...hw", "hi...hm", "sources (Zoph and Knight, 2016):", "Tencoder", "Tencoder", "Ck = [ck; ch]", "X1.XN", "ti...tm", "ToCR", "TocR", "Ck is used by the decoder LSTM to compute the", "", "next hidden state Sk and subsequently, the proba-", "bility distribution for predicting the next character", "iki-an aine", "Ainu", "Japanese", "Yk of the target sequence y:", "Figure 3: The proposed multi-source architecture with", "Sh. = lstm(sk-1,Ch,yk-1)", "the encoder for an endangered language segment (left)", "(2)", "and an encoder for the translated segment (right). The", "P(yk) = softmax(Wsk + b)", "(3)", "input to the encoders is the first pass OCR over the", "scanned images of each segment. For example, the", "Training and Inference The model is trained for", "OCRonthe scanned images of someAinutext (left)", "each language with the cross-entropy loss (Cce)", "and its Japanese translation (right).", "on the small amount of transcribed data we have.", "Beam search is used at inference time.", "framework (Zoph and Knight, 2016; Libovicky and", "5.2Model and Training Improvements", "Helcl, 2017) that uses an LSTM encoder-decoder", "model with attention (Bahdanau et al., 2015). We", "With the minimal annotated data we have, it is", "propose improvements to training and modeling for", "challenging for the neural network to learn a good", "the multi-source architecture, specifically tailored", "distribution over the target characters. We propose", "to ease learning in data-scarce settings.", "a set of adaptations to the base architecture that", "improves the post-correction performance without", "5.1 Multi-source Architecture", "additional annotation. The adaptations are based", "Our post-correction formulation takes as input the", "on characteristics of the OCR task itself and the", "first pass OCR of the endangered language segment", "performance of the upstream OCR tool (Section 4).", " and the OCR of the translated segment t, to", "Diagonal attention loss As seen in Section 4,", "predict an error-free endangered language text y.", " substitution errors are more frequent in the OCR", "The model architecture is shown in Figure 3.", "task than insertions or deletions; consequently.", "The model consists of two encoders -", "-onethat", "we expect the source and target to have similar", "encodes and one that encodes t. Each encoder is", "lengths. Moreover, post-correction is a monotonic", "a character-level bidirectional LSTM (Hochreiter", "sequence-to-sequence task, and reordering rarely", "and Schmidhuber, 1997) and transforms the input", "occurs (Schnober et al., 2016). Hence, we expect", "sequence of characters to a sequence of hidden", "state vectors: h\" for the endangered language text", "attentionweights tobehigher at characters closeto", "the diagonal for the endangered language encoder.", "and ht for the translation.", "We modify the model such that all the elements", "The model uses an attention mechanism during", "in the attention vector that are not within j steps", "the decoding process to use information from the", "(we use j = 3) of the current time step k are added", "encoder hidden states. We compute the attention", "to the training loss, thereby encouraging elements", "weights over each of the two encoders indepen-", "awayfromthe diagonal tohavelower values.The", "dently. At the decoding time step k:", "diagonal loss summed over all time steps for a", "ex., = v\" tanh (Wist-1 + W2h,)", "(1)", "training instance, where N is the length of ac, is:", "αx = softmax (er)", "N", "Caiag", "Qk,i+", "", "ak,", "c =[E,ok,h]]", "i=1", "i=k+j" ]

{ "arxiv_id": "2011.05402", "language": "en", "page_idx": 4 }

[ [ "0.2059", "0.125", "0.7935", "0.1256", "0.7935", "0.1465", "0.2059", "0.1459" ], [ "0.2932", "0.2026", "0.4143", "0.2026", "0.4143", "0.2172", "0.2932", "0.2172" ], [ "0.603", "0.2039", "0.701", "0.2039", "0.701", "0.2153", "0.603", "0.2153" ], [ "0.2361", "0.2166", "0.4714", "0.2166", "0.4714", "0.2311", "0.2361", "0.2311" ], [ "0.581", "0.2166", "0.7197", "0.2166", "0.7197", "0.2311", "0.581", "0.2311" ], [ "0.2402", "0.2299", "0.4673", "0.2299", "0.4673", "0.2437", "0.2402", "0.2437" ], [ "0.539", "0.2299", "0.7646", "0.2299", "0.7646", "0.2437", "0.539", "0.2437" ], [ "0.2754", "0.2443", "0.4314", "0.2443", "0.4314", "0.2583", "0.2754", "0.2583" ], [ "0.54", "0.2468", "0.763", "0.2468", "0.763", "0.2568", "0.54", "0.2568" ], [ "0.02287", "0.2676", "0.06046", "0.2676", "0.06046", "0.3738", "0.02287", "0.3738" ], [ "0.4395", "0.2778", "0.5596", "0.2778", "0.5596", "0.2922", "0.4395", "0.2922" ], [ "0.38", "0.2922", "0.6167", "0.2922", "0.6167", "0.3062", "0.38", "0.3062" ], [ "0.3857", "0.307", "0.613", "0.3062", "0.613", "0.3208", "0.3857", "0.3213" ], [ "0.3865", "0.3213", "0.614", "0.3213", "0.614", "0.3352", "0.3865", "0.3352" ], [ "0.02533", "0.363", "0.058", "0.363", "0.058", "0.3984", "0.02533", "0.3984" ], [ "0.46", "0.3694", "0.54", "0.3694", "0.54", "0.387", "0.46", "0.387" ], [ "0.02124", "0.404", "0.06046", "0.404", "0.06046", "0.4893", "0.02124", "0.4893" ], [ "0.2336", "0.4028", "0.7656", "0.4028", "0.7656", "0.4167", "0.2336", "0.4167" ], [ "0.2336", "0.4172", "0.767", "0.4172", "0.767", "0.4312", "0.2336", "0.4312" ], [ "0.2336", "0.4312", "0.7656", "0.4312", "0.7656", "0.4443", "0.2336", "0.4443" ], [ "0.2336", "0.4443", "0.7656", "0.4443", "0.7656", "0.4583", "0.2336", "0.4583" ], [ "0.2336", "0.459", "0.7656", "0.459", "0.7656", "0.473", "0.2336", "0.473" ], [ "0.2336", "0.4717", "0.767", "0.4717", "0.767", "0.486", "0.2336", "0.486" ], [ "0.02287", "0.4893", "0.05963", "0.4893", "0.05963", "0.712", "0.02287", "0.712" ], [ "0.2336", "0.486", "0.7656", "0.486", "0.7656", "0.5", "0.2336", "0.5" ], [ "0.2336", "0.5", "0.7656", "0.5", "0.7656", "0.5146", "0.2336", "0.5146" ], [ "0.2336", "0.513", "0.767", "0.513", "0.767", "0.5264", "0.2336", "0.5264" ], [ "0.2336", "0.528", "0.3782", "0.528", "0.3782", "0.5415", "0.2336", "0.5415" ], [ "0.1772", "0.5693", "0.1978", "0.5693", "0.1978", "0.58", "0.1772", "0.58" ], [ "0.1969", "0.5674", "0.312", "0.5674", "0.312", "0.582", "0.1969", "0.582" ], [ "0.1732", "0.5986", "0.824", "0.5986", "0.824", "0.6123", "0.1732", "0.6123" ], [ "0.1748", "0.613", "0.824", "0.613", "0.824", "0.627", "0.1748", "0.627" ], [ "0.1732", "0.6255", "0.8228", "0.6255", "0.8228", "0.6396", "0.1732", "0.6396" ], [ "0.1732", "0.64", "0.824", "0.64", "0.824", "0.654", "0.1732", "0.654" ], [ "0.1748", "0.655", "0.6665", "0.655", "0.6665", "0.6685", "0.1748", "0.6685" ], [ "0.1748", "0.675", "0.824", "0.675", "0.824", "0.688", "0.1748", "0.688" ], [ "0.1732", "0.6875", "0.8228", "0.6875", "0.8228", "0.702", "0.1732", "0.702" ], [ "0.1748", "0.7026", "0.824", "0.7026", "0.824", "0.7163", "0.1748", "0.7163" ], [ "0.1748", "0.7163", "0.824", "0.7163", "0.824", "0.73", "0.1748", "0.73" ], [ "0.1732", "0.73", "0.746", "0.729", "0.746", "0.7437", "0.1732", "0.744" ], [ "0.1748", "0.7505", "0.824", "0.7505", "0.824", "0.7646", "0.1748", "0.7646" ], [ "0.1732", "0.7646", "0.824", "0.7646", "0.824", "0.779", "0.1732", "0.779" ], [ "0.1732", "0.779", "0.824", "0.779", "0.824", "0.7925", "0.1732", "0.7925" ], [ "0.1748", "0.7925", "0.824", "0.7925", "0.824", "0.806", "0.1748", "0.806" ], [ "0.1748", "0.806", "0.824", "0.806", "0.824", "0.8193", "0.1748", "0.8193" ], [ "0.1732", "0.8193", "0.824", "0.8193", "0.824", "0.8335", "0.1732", "0.8335" ], [ "0.1748", "0.834", "0.8267", "0.834", "0.8267", "0.8477", "0.1748", "0.8477" ], [ "0.1748", "0.8467", "0.824", "0.8467", "0.824", "0.86", "0.1748", "0.86" ], [ "0.1732", "0.86", "0.8228", "0.86", "0.8228", "0.8735", "0.1732", "0.8735" ], [ "0.1748", "0.875", "0.824", "0.875", "0.824", "0.888", "0.1748", "0.888" ], [ "0.1748", "0.888", "0.824", "0.888", "0.824", "0.903", "0.1748", "0.903" ], [ "0.1732", "0.9224", "0.778", "0.924", "0.778", "0.938", "0.1732", "0.9375" ] ]

[ "User-Interactive Offline Reinforcement Learning", "Phillip Swazinna", "SteffenUdluft", "Siemens Technology & TU Munich", "Siemens Technology", "Otto-Hahn-Ring 6, 81739 Munich", "Otto-Hahn-Ring 6, 81739 Munich", "swazinna@in.tum.de", "steffen.udluft@siemens.com", "", "Thomas Runkler", "Siemens Technology&TUMunich", "Otto-Hahn-Ring 6, 81739 Munich", "thomas.runkler@siemens.com", "5", "Abstract", "['S]", "Offline reinforcement learning algorithms still lack trust in practice due to the", "risk that the learned policy performs worse than the original policy that generated", "the dataset or behaves in an unexpected way that is unfamiliar to the user. At", "the same time, offline RL algorithms are not able to tune their most important", "hyperparameter - the proximity of the learned policy to the original policy. We", "propose an algorithm that allows the user to tune this hyperparameter at runtime,", "262901521", "thereby addressing both of the above mentioned issues simultaneously. This allows", "users to start with the original behavior and grant successively greater deviation, as", "well as stopping at any time when the policy deteriorates or the behavior is too far", "from the familiar one.", "1", "Introduction", "Recently, offline reinforcement learning (RL) methods have shown that it is possible to learn effective", "policies from a static pre-collected dataset instead of directly interacting with the environment", "(Laroche et al., 2019; Fujimoto et al., 2019; Yu et al., 2020; Swazinna et al., 2021b). Since direct", "interaction is in practice usually very costly, these techniques have alleviated a large obstacle on the", "path of applying reinforcement learning techniques in real world problems.", "A major issue that these algorithms still face is tuning their most important hyperparameter: The", "proximity to the original policy. Virtually all algorithms tackling the offline setting have such a", "hyperparameter, and it is obviously hard to tune, since no interaction with the real environment is", "permitted until final deployment. Practitioners thus risk being overly conservative (resulting in no", "improvement) or overly progressive (risking worse performing policies) in their choice.", "Additionally, one of the arguably largest obstacles on the path to deployment of RL trained policies", "in most industrial control problems is that (offline) RL algorithms ignore the presence of domain", "experts, who can be seen as users of the final product - the policy. Instead, most algorithms today can", "be seen as trying to make human practitioners obsolete. We argue that it is important to provide these", "users with a utility - something that makes them want to use RL solutions. Other research fields, such", "as machine learning for medical diagnoses, have already established the idea that domain experts", "are crucially important to solve the task and complement human users in various ways Babbar et al.", "(2022); Cai et al. (2019); De-Arteaga et al. (2021); Fard & Pineau (2011); Tang et al. (2020). We see", "our work in line with these and other researchers (Shneiderman, 2020; Schmidt et al., 2021), who", "suggest that the next generation of AI systems needs to adopt a user-centered approach and develop", "systems that behave more like an intelligent tool, combining both high levels of human control and", "Accepted at ICLR 2023 - 11th International Conference on Learning Representations, Kigali, Rwanda." ]

{ "arxiv_id": "2205.10629", "language": "en", "page_idx": 0 }

[ [ "0.1732", "0.09155", "0.826", "0.09216", "0.826", "0.1067", "0.1732", "0.1061" ], [ "0.1765", "0.1092", "0.8228", "0.1092", "0.8228", "0.1193", "0.1765", "0.1193" ], [ "0.1732", "0.1206", "0.824", "0.1206", "0.824", "0.1338", "0.1732", "0.1338" ], [ "0.1732", "0.1338", "0.824", "0.1338", "0.824", "0.1477", "0.1732", "0.1477" ], [ "0.1732", "0.1465", "0.824", "0.1477", "0.824", "0.1622", "0.1732", "0.161" ], [ "0.1748", "0.1622", "0.6665", "0.1622", "0.6665", "0.1755", "0.1748", "0.1755" ], [ "0.1732", "0.1812", "0.824", "0.1818", "0.824", "0.1963", "0.1732", "0.1957" ], [ "0.1748", "0.1963", "0.824", "0.1963", "0.824", "0.2102", "0.1748", "0.2102" ], [ "0.1748", "0.2102", "0.824", "0.2102", "0.824", "0.2235", "0.1748", "0.2235" ], [ "0.1765", "0.226", "0.8228", "0.226", "0.8228", "0.2361", "0.1765", "0.2361" ], [ "0.1748", "0.238", "0.8228", "0.238", "0.8228", "0.252", "0.1748", "0.252" ], [ "0.1748", "0.252", "0.824", "0.252", "0.824", "0.2651", "0.1748", "0.2651" ], [ "0.1765", "0.2676", "0.8228", "0.2676", "0.8228", "0.2778", "0.1765", "0.2778" ], [ "0.1748", "0.279", "0.824", "0.279", "0.824", "0.2922", "0.1748", "0.2922" ], [ "0.1748", "0.2922", "0.824", "0.2922", "0.824", "0.3062", "0.1748", "0.3062" ], [ "0.1748", "0.307", "0.826", "0.307", "0.826", "0.3208", "0.1748", "0.3208" ], [ "0.1732", "0.3193", "0.824", "0.3208", "0.824", "0.334", "0.1732", "0.3333" ], [ "0.1748", "0.334", "0.824", "0.334", "0.824", "0.348", "0.1748", "0.348" ], [ "0.1732", "0.3484", "0.826", "0.3484", "0.826", "0.3623", "0.1732", "0.3623" ], [ "0.1765", "0.363", "0.8228", "0.363", "0.8228", "0.3738", "0.1765", "0.3738" ], [ "0.1732", "0.3757", "0.5986", "0.3757", "0.5986", "0.3894", "0.1732", "0.3894" ], [ "0.1765", "0.4204", "0.1912", "0.4204", "0.1912", "0.4312", "0.1765", "0.4312" ], [ "0.1978", "0.4185", "0.3218", "0.4192", "0.3218", "0.4343", "0.1978", "0.433" ], [ "0.1748", "0.4583", "0.826", "0.4583", "0.826", "0.4717", "0.1748", "0.4717" ], [ "0.1748", "0.4717", "0.824", "0.4717", "0.824", "0.4849", "0.1748", "0.4849" ], [ "0.1748", "0.4849", "0.826", "0.4849", "0.826", "0.4988", "0.1748", "0.4988" ], [ "0.1732", "0.5005", "0.8228", "0.5005", "0.8228", "0.511", "0.1732", "0.511" ], [ "0.1732", "0.513", "0.824", "0.513", "0.824", "0.5264", "0.1732", "0.5264" ], [ "0.1748", "0.529", "0.2279", "0.529", "0.2279", "0.539", "0.1748", "0.539" ], [ "0.1732", "0.547", "0.822", "0.547", "0.822", "0.56", "0.1732", "0.56" ], [ "0.1748", "0.56", "0.824", "0.56", "0.824", "0.5737", "0.1748", "0.5737" ], [ "0.1732", "0.5747", "0.824", "0.5747", "0.824", "0.5884", "0.1732", "0.5884" ], [ "0.1748", "0.5884", "0.824", "0.5884", "0.824", "0.6016", "0.1748", "0.6016" ], [ "0.1748", "0.603", "0.826", "0.603", "0.826", "0.616", "0.1748", "0.616" ], [ "0.1765", "0.616", "0.826", "0.616", "0.826", "0.63", "0.1765", "0.63" ], [ "0.1748", "0.63", "0.826", "0.63", "0.826", "0.643", "0.1748", "0.643" ], [ "0.1748", "0.643", "0.826", "0.643", "0.826", "0.657", "0.1748", "0.657" ], [ "0.1748", "0.656", "0.826", "0.656", "0.826", "0.6694", "0.1748", "0.6694" ], [ "0.1748", "0.672", "0.826", "0.672", "0.826", "0.685", "0.1748", "0.685" ], [ "0.1748", "0.685", "0.826", "0.685", "0.826", "0.6987", "0.1748", "0.6987" ], [ "0.1732", "0.6987", "0.826", "0.6978", "0.826", "0.712", "0.1732", "0.7134" ], [ "0.1748", "0.712", "0.824", "0.712", "0.824", "0.726", "0.1748", "0.726" ], [ "0.1732", "0.7246", "0.824", "0.726", "0.824", "0.7407", "0.1732", "0.7393" ], [ "0.1732", "0.7393", "0.824", "0.7393", "0.824", "0.7524", "0.1732", "0.7524" ], [ "0.1732", "0.7524", "0.8267", "0.754", "0.8267", "0.768", "0.1732", "0.7676" ], [ "0.1724", "0.7666", "0.6846", "0.7676", "0.6846", "0.782", "0.1724", "0.781" ], [ "0.1748", "0.786", "0.824", "0.787", "0.824", "0.802", "0.1748", "0.8003" ], [ "0.1748", "0.802", "0.824", "0.802", "0.824", "0.815", "0.1748", "0.815" ], [ "0.1748", "0.8164", "0.824", "0.8164", "0.824", "0.83", "0.1748", "0.83" ], [ "0.1748", "0.83", "0.826", "0.83", "0.826", "0.8433", "0.1748", "0.8433" ], [ "0.1732", "0.8433", "0.824", "0.8433", "0.824", "0.857", "0.1732", "0.857" ], [ "0.1748", "0.858", "0.824", "0.858", "0.824", "0.872", "0.1748", "0.872" ], [ "0.1732", "0.872", "0.824", "0.872", "0.824", "0.8853", "0.1732", "0.8853" ], [ "0.1748", "0.8853", "0.824", "0.8853", "0.824", "0.8984", "0.1748", "0.8984" ], [ "0.1732", "0.8984", "0.824", "0.8984", "0.824", "0.912", "0.1732", "0.912" ], [ "0.492", "0.9346", "0.5073", "0.9346", "0.5073", "0.9507", "0.492", "0.9507" ] ]

[ "high levels of automation. We seek to develop an offine RL method that does just that. Furthermore,", "weseegivingcontroltotheuserasarequirementthatmayinthefuturebemuchmoreenforcedwhen", "regulations regarding AI systems become more strict: The EU's high level expert group on AI has", "already recognized \"human autonomy and oversight\" as a key requirement for trustworthy Al in their", "Ethics Guidelines for Trustworthy AI (Smuha, 2019). In the future, solutions found with RL might", "thus be required by law to exhibit features that enable more human control.", "In this paper, we thus propose a simple method to provide users with more control over how an", "offline RL policy will behave after deployment. The algorithm that we develop trains a conditional", "policy, that can after training adapt the trade-off between proximity to the data generating policy on", "theonehandandestimatedperformanceontheotherCloseproximitytoaknownsolutionnaturally", "facilitates trust, enabling conservative users to choose behavior they are more inclined to confidently", "deploy. That way, users may benefit from the automation provided by offline RL (users don't need", "tohandcraftcontrollers,possiblyeveninteractivelychooseactions)yetstillremainincontrolas", "they can e.g. make the policy move to a more conservative or more liberal trade-off. We show how", "such an algorithm can be designed, as well as compare its performance with a variety of ofline", "RL baselines and show that a user can achieve state of the art performance with it. Furthermore,", "we showthat ourmethod has advantages over simplerapproaches liketrainingmanypolicies with", "diverse hyperparameters. Finally, since we train a policy conditional on one of the most important", "hyperparameters in offline RL, we show how a user could potentially use it to tune this hyperparameter.", "Inmany cases of ourevaluations,this works almostregret-free,sinceweobservethattheperformance", "as a function of the hyperparameter is mostly a smooth function.", "2", "Related Work", "Offline RL Recently, a plethora of methods has been published that learn policies from static datasets.", "Early works, such as FQI and NFQ (Ernst et al., 2005; Riedmiller, 2005), were termed batch instead", "of offline since they didn't explicitly address issue that the data collection cannot be influenced.", "Instead,similarlytootherbatch methods (Depeweg et al.,2016;Heinet al.,2018;Kaiser et al.,2020)", "they assumed a uniform random data collection that made generalization to the real environment", "simpler.", "Among the first to explicitly address the limitations in the offline setting under unknown data", "collection were SPIBB(-DQN) (Laroche et al., 2019) in the discrete and BCQ (Fujimoto et al., 2019)", "in the continuous actions case. Many works with different focuses followed: Some treat discrete", "MDPs and come with provable bounds on the performance at least with a certain probability Thomas", "et al. (2015); Nadjahi et al. (2019), however many more focused on the continuous setting: EMaQ,", "BEAR, BRAC, ABM, various DICE based methods, REM, PEBL, PSEC-TD-O, CQL, IQL, BAIL,", "CRR, COIL, O-RAAC, OPAL, TD3+BC, and RvS (Ghasemipour et al., 2021; Kumar et al., 2019;", "Wu et al., 2019; Siegel et al., 2020; Nachum et al., 2019; Zhang et al., 2020; Agarwal et al., 2020;", "Smit et al., 2021; Pavse et al., 2020: Kumar et al., 2020: Kostrikov et al., 2021: Chen et al., 2019:", "Wang et al., 2020; Liu et al., 2021; Urpi et al., 2021; Ajay et al., 2020; Brandfonbrener et al., 2021;", "Emmons et al., 2021) are just a few of the proposed model-free methods over the last few years.", "Additionally, many model-based as well as hybrid approaches have been proposed, such as MOPO,", "MOReL, MOOSE, COMBO, RAMBO, and WSBC (Yu et al., 2020; Kidambi et al., 2020; Swazinna", "et al., 2021b; Yu et al., 2021; Rigter et al., 2022; Swazinna et al., 2021a). Even approaches that train", "policies purely supervised, by conditioning on performance, have been proposed (Peng et al., 2o19:", "Emmons et al., 2021; Chen et al., 2021). Model based algorithms more often use model uncertainty,", "while model-free methods use a more direct behavior regularization approach", "Offline policy evaluation or offline hyperparameter selection is concerned with evaluating (or at", "least ranking) policies that have been found by an offline RL algorithm, in order to either pick the best", "performing one or to tune hyperparameters. Often, dynamics models are used to evaluate policies", "found in model-free algorithms, however also model-free evaluation methods exist (Hans et al., 2011;", "Paine et al., 2020; Konyushova et al., 2021; Zhang et al., 2021b; Fu et al., 2021). Unfortunately, but", "also intuitively, this problem is rather hard since if any method is found that can more accurately", "assess the policy performance than the mechanism in the offline algorithm used for training, it should", "be used instead of the previously employed method for training. Also, the general dilemma of not", "knowing in which parts of the state-action space we know enough to optimize behavior seems to", "2" ]

{ "arxiv_id": "2205.10629", "language": "en", "page_idx": 1 }

[ [ "0.1748", "0.09155", "0.824", "0.09155", "0.824", "0.1061", "0.1748", "0.1061" ], [ "0.1748", "0.1067", "0.709", "0.1067", "0.709", "0.1206", "0.1748", "0.1206" ], [ "0.1748", "0.127", "0.824", "0.127", "0.824", "0.1401", "0.1748", "0.1401" ], [ "0.1748", "0.1414", "0.8228", "0.1414", "0.8228", "0.1547", "0.1748", "0.1547" ], [ "0.1748", "0.1547", "0.824", "0.1547", "0.824", "0.1686", "0.1748", "0.1686" ], [ "0.1748", "0.1686", "0.826", "0.1686", "0.826", "0.1818", "0.1748", "0.1818" ], [ "0.1748", "0.1831", "0.826", "0.1831", "0.826", "0.1963", "0.1748", "0.1963" ], [ "0.1765", "0.1963", "0.824", "0.1963", "0.824", "0.2102", "0.1765", "0.2102" ], [ "0.1765", "0.2102", "0.824", "0.2102", "0.824", "0.2235", "0.1765", "0.2235" ], [ "0.1732", "0.2229", "0.8228", "0.2235", "0.8228", "0.2374", "0.1732", "0.2361" ], [ "0.1765", "0.2393", "0.822", "0.2393", "0.822", "0.25", "0.1765", "0.25" ], [ "0.1732", "0.252", "0.822", "0.252", "0.822", "0.2651", "0.1732", "0.2651" ], [ "0.1765", "0.2676", "0.822", "0.2676", "0.822", "0.2778", "0.1765", "0.2778" ], [ "0.1765", "0.281", "0.8228", "0.281", "0.8228", "0.2917", "0.1765", "0.2917" ], [ "0.1724", "0.2917", "0.5415", "0.2922", "0.5415", "0.307", "0.1724", "0.3062" ], [ "0.1732", "0.3257", "0.6445", "0.3257", "0.6445", "0.3403", "0.1732", "0.3403" ], [ "0.1748", "0.356", "0.8228", "0.356", "0.8228", "0.3694", "0.1748", "0.3694" ], [ "0.1732", "0.3694", "0.824", "0.3687", "0.824", "0.3818", "0.1732", "0.3833" ], [ "0.1732", "0.3833", "0.8228", "0.3838", "0.8228", "0.3977", "0.1732", "0.3965" ], [ "0.1748", "0.3977", "0.824", "0.3977", "0.824", "0.411", "0.1748", "0.411" ], [ "0.1748", "0.411", "0.824", "0.411", "0.824", "0.4248", "0.1748", "0.4248" ], [ "0.1748", "0.4248", "0.824", "0.4248", "0.824", "0.4395", "0.1748", "0.4395" ], [ "0.1732", "0.4395", "0.2852", "0.4375", "0.2852", "0.4507", "0.1732", "0.4526" ], [ "0.1732", "0.457", "0.8228", "0.4583", "0.8228", "0.473", "0.1732", "0.4717" ], [ "0.1748", "0.473", "0.824", "0.473", "0.824", "0.486", "0.1748", "0.486" ], [ "0.1748", "0.4873", "0.8228", "0.4873", "0.8228", "0.5005", "0.1748", "0.5005" ], [ "0.1748", "0.5005", "0.826", "0.5005", "0.826", "0.5146", "0.1748", "0.5146" ], [ "0.1748", "0.5146", "0.8228", "0.5146", "0.8228", "0.528", "0.1748", "0.528" ], [ "0.1748", "0.528", "0.8228", "0.528", "0.8228", "0.5415", "0.1748", "0.5415" ], [ "0.1748", "0.5425", "0.824", "0.5425", "0.824", "0.556", "0.1748", "0.556" ], [ "0.1748", "0.556", "0.4968", "0.556", "0.4968", "0.5693", "0.1748", "0.5693" ], [ "0.1748", "0.5854", "0.2059", "0.5854", "0.2059", "0.5977", "0.1748", "0.5977" ], [ "0.2018", "0.586", "0.2737", "0.586", "0.2737", "0.5977", "0.2018", "0.5977" ], [ "0.1748", "0.61", "0.8228", "0.61", "0.8228", "0.625", "0.1748", "0.625" ], [ "0.1748", "0.627", "0.8228", "0.627", "0.8228", "0.64", "0.1748", "0.64" ], [ "0.1732", "0.6426", "0.8228", "0.6416", "0.8228", "0.656", "0.1732", "0.657" ], [ "0.1732", "0.657", "0.824", "0.657", "0.824", "0.6704", "0.1732", "0.6704" ], [ "0.1748", "0.6704", "0.824", "0.6704", "0.824", "0.6846", "0.1748", "0.6846" ], [ "0.1732", "0.683", "0.824", "0.6846", "0.824", "0.6987", "0.1732", "0.6978" ], [ "0.1748", "0.6987", "0.6807", "0.6987", "0.6807", "0.712", "0.1748", "0.712" ], [ "0.3865", "0.7217", "0.437", "0.7217", "0.437", "0.7363", "0.3865", "0.7363" ], [ "0.4697", "0.7183", "0.6113", "0.72", "0.6113", "0.738", "0.4697", "0.7363" ], [ "0.8022", "0.7227", "0.826", "0.7227", "0.826", "0.7373", "0.8022", "0.7373" ], [ "0.4429", "0.732", "0.4631", "0.732", "0.4631", "0.7407", "0.4429", "0.7407" ], [ "0.4739", "0.742", "0.5073", "0.7407", "0.5073", "0.7476", "0.4739", "0.75" ], [ "0.1765", "0.754", "0.824", "0.754", "0.824", "0.7676", "0.1765", "0.7676" ], [ "0.1772", "0.7695", "0.822", "0.7695", "0.822", "0.78", "0.1772", "0.78" ], [ "0.1765", "0.782", "0.8228", "0.782", "0.8228", "0.7954", "0.1765", "0.7954" ], [ "0.1732", "0.7944", "0.824", "0.7954", "0.824", "0.81", "0.1732", "0.8096" ], [ "0.1765", "0.81", "0.3145", "0.81", "0.3145", "0.8237", "0.1765", "0.8237" ], [ "0.29", "0.829", "0.3022", "0.829", "0.3022", "0.837", "0.29", "0.837" ], [ "0.3308", "0.83", "0.357", "0.83", "0.357", "0.8486", "0.3308", "0.8486" ], [ "0.3816", "0.832", "0.5254", "0.83", "0.5254", "0.8477", "0.3816", "0.8496" ], [ "0.8013", "0.832", "0.8267", "0.832", "0.8267", "0.8496", "0.8013", "0.8496" ], [ "0.3816", "0.837", "0.3857", "0.837", "0.3857", "0.8423", "0.3816", "0.8423" ], [ "0.2222", "0.8335", "0.2737", "0.8335", "0.2737", "0.8477", "0.2222", "0.8477" ], [ "0.2852", "0.8413", "0.308", "0.8413", "0.308", "0.853", "0.2852", "0.853" ], [ "0.3057", "0.8516", "0.3743", "0.8516", "0.3743", "0.863", "0.3057", "0.863" ], [ "0.2222", "0.88", "0.2712", "0.88", "0.2712", "0.8945", "0.2222", "0.8945" ], [ "0.304", "0.8755", "0.777", "0.877", "0.777", "0.898", "0.304", "0.8965" ], [ "0.2876", "0.89", "0.3057", "0.89", "0.3057", "0.8994", "0.2876", "0.8994" ], [ "0.3057", "0.8965", "0.3594", "0.8965", "0.3594", "0.9077", "0.3057", "0.9077" ], [ "0.492", "0.935", "0.5073", "0.935", "0.5073", "0.9507", "0.492", "0.9507" ] ]

[ "always remain. Works such as Zhang et al. (202la); Lu et al. (2021) become applicable if limited", "online evaluations are allowed, making hyperparameter tuning much more viable.", "OfflineRL with onlineadaptation Otherworksproposean onlinelearningphase thatfollows after", "offline learning has conceded. In the most basic form, Kurenkov & Kolesnikov (2021) introduce", "an online evaluation budget that lets them find the best set of hyperparameters for an offline RL", "algorithm given limited online evaluation resources. In an effort to minimize such a budget, Yang et al.", "(2021) train a set of policies spanning a diverse set of uncertainty-performance trade-offs. Ma et al.", "(2021) propose a conservative adaptive penalty, that penalizes unknown behavior more during the", "beginning and less during the end of training, leading to safer policies during training. In Pong et al", "(2021);Nair et al.(2020);Zhao etal.(2021)methods for effectiveonline learningphases thatfollow", "theofflinelearningphaseareproposed.Incontrasttothesemethods,wearenotaimingforafully", "automated solution. Instead, we want to provide the user with a valuable tool after training, so we do", "notproposeanactualonlinephase,alsosincepractitionersmayfindanyperformancedeterioration", "inacceptable.Tothebestofourknowledge,nopriorofflineRLmethodproducespoliciesthatremain", "adaptable after deployment without any further training.", "3LION: Learning in Interactive Offline eNvironments", "In this work, we address two dilemmas of the offline RL setting: First and foremost, we would like to", "provide the user with a high level control option in order to influence the behavior of the policy,since", "we argue that the user is crucially important for solving the task and not tobe madeobsolete by an", "algorithm. Further we address the issue that in offline RL, the correct hyperparameter controlling the", "trade-off between conservatism and performance is unknown and can hardly be tuned. By training a", "policy conditioned in the proximity hyperparameter, we aim to enable the user to find a good trade-off", "hyperparameter.", "As mentioned, behavior cloning, will most likely yield the most trustworthy solution due to its", "familiarity, however the solution is of very limited use since it does not outperform the previous one", "Offline RL on the other hand is problematic since we cannot simply evaluate policy candidates on the", "real system and offline policy evaluation is still an open problem (Hans et al., 2011; Paine et al., 2020;", "Konyushova et al., 2021; Zhang et al., 2021b; Fu et al., 2021). In the following, we thus propose a", "solution that moves the hyperparameter choice from training to deployment time, enabling the user to", "interactively find the desired trade-off between BC and offline optimization. A user may then slowly", "move from conservative towards better solutions.", "3.1", "Training", "During training time, we optimize three components: A model of the original policy β(s), an", "ensemble of transition dynamics models (fu, (s, a)li E O, :.., N - 1), as well as the user adaptive", "policy Te(s, A). The dynamics models (fi) as well as the original policy β are trained in isolation", "before the actual policy training starts.Both and β are always simplefeedforward neural networks", "which map states directly to actions in a deterministic fashion (practitioners likely favor deterministic", "policies over stochastic ones due to trust issues). β is trained to simply imitate the behavior present in", "the dataset by minimizing the mean squared distance to the observed actions:", "L(Φ) =", " [at -βs(st)]?", "(1)", "N", "at~", "Depending on the environment, the transition models are either also feedforward networks or simple", "recurrentnetworks withasinglerecurrentlayer.Therecurrentnetworksbuildtheirhiddenstate", "over G steps and are then trained to predict a window of size F into the future (similarly to (Hein", "et al., 2017b), while the feedforward dynamics simply predict single step transitions.Both use mean", "squared erroras loss:", "1", "7", "[st+1-fi,(st,at)]", "(2)", "", "L(a) :", "N", "St,at,st+1~", "L(ba) :", "Z[s++G+f+1 - fi ($t, at, ..$+G, at+G,... ++G+f, at+G+)]?", "N", "t~Df=1", "3" ]

{ "arxiv_id": "2205.10629", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.09216", "0.824", "0.09216", "0.824", "0.1067", "0.1748", "0.1067" ], [ "0.1748", "0.108", "0.8228", "0.108", "0.8228", "0.1187", "0.1748", "0.1187" ], [ "0.1748", "0.1218", "0.4968", "0.1206", "0.4968", "0.132", "0.1748", "0.1332" ], [ "0.1748", "0.1401", "0.8228", "0.1401", "0.8228", "0.1547", "0.1748", "0.1547" ], [ "0.1748", "0.1547", "0.824", "0.1547", "0.824", "0.1692", "0.1748", "0.1692" ], [ "0.1765", "0.1718", "0.824", "0.1718", "0.824", "0.1818", "0.1765", "0.1818" ], [ "0.1748", "0.1843", "0.8228", "0.1843", "0.8228", "0.1976", "0.1748", "0.1976" ], [ "0.1748", "0.1976", "0.8228", "0.1976", "0.8228", "0.2108", "0.1748", "0.2108" ], [ "0.1765", "0.2134", "0.8228", "0.2134", "0.8228", "0.2235", "0.1765", "0.2235" ], [ "0.1748", "0.2247", "0.8228", "0.2247", "0.8228", "0.238", "0.1748", "0.238" ], [ "0.1732", "0.2374", "0.826", "0.238", "0.826", "0.2524", "0.1732", "0.252" ], [ "0.1748", "0.254", "0.415", "0.254", "0.415", "0.2644", "0.1748", "0.2644" ], [ "0.273", "0.286", "0.3447", "0.286", "0.3447", "0.3005", "0.273", "0.3005" ], [ "0.3447", "0.2874", "0.361", "0.2874", "0.361", "0.2954", "0.3447", "0.2954" ], [ "0.3774", "0.2854", "0.613", "0.2854", "0.613", "0.303", "0.3774", "0.303" ], [ "0.609", "0.2874", "0.726", "0.2874", "0.726", "0.3005", "0.609", "0.3005" ], [ "0.8022", "0.2874", "0.826", "0.2874", "0.826", "0.3018", "0.8022", "0.3018" ], [ "0.353", "0.308", "0.3743", "0.308", "0.3743", "0.3125", "0.353", "0.3125" ], [ "0.1732", "0.3225", "0.822", "0.3225", "0.822", "0.3372", "0.1732", "0.3372" ], [ "0.1732", "0.3384", "0.826", "0.339", "0.826", "0.3542", "0.1732", "0.3528" ], [ "0.1748", "0.3542", "0.8228", "0.3542", "0.8228", "0.3674", "0.1748", "0.3674" ], [ "0.1732", "0.3674", "0.824", "0.3674", "0.824", "0.3806", "0.1732", "0.3806" ], [ "0.1724", "0.3806", "0.2467", "0.379", "0.2467", "0.3933", "0.1724", "0.3958" ], [ "0.2173", "0.4185", "0.2296", "0.4185", "0.2296", "0.43", "0.2173", "0.43" ], [ "0.734", "0.4236", "0.7563", "0.4236", "0.7563", "0.483", "0.734", "0.483" ], [ "0.683", "0.4817", "0.714", "0.4817", "0.714", "0.4849", "0.683", "0.4849" ], [ "0.746", "0.4873", "0.751", "0.4905", "0.746", "0.4968", "0.74", "0.4937" ], [ "0.6724", "0.526", "0.705", "0.526", "0.705", "0.5327", "0.6724", "0.5327" ], [ "0.74", "0.5234", "0.796", "0.5234", "0.796", "0.535", "0.74", "0.535" ], [ "0.3057", "0.5737", "0.3179", "0.5835", "0.3057", "0.5923", "0.2932", "0.5835" ], [ "0.3594", "0.5757", "0.3857", "0.579", "0.3823", "0.5933", "0.357", "0.5903" ], [ "0.7793", "0.577", "0.794", "0.577", "0.794", "0.587", "0.7793", "0.587" ], [ "0.4265", "0.579", "0.4412", "0.579", "0.4412", "0.589", "0.4265", "0.589" ], [ "0.2009", "0.582", "0.21", "0.582", "0.21", "0.587", "0.2009", "0.587" ], [ "0.2427", "0.577", "0.2632", "0.577", "0.2632", "0.5903", "0.2427", "0.5903" ], [ "0.687", "0.582", "0.701", "0.582", "0.701", "0.5854", "0.687", "0.5854" ], [ "0.728", "0.5903", "0.7563", "0.5923", "0.755", "0.601", "0.726", "0.5986" ], [ "0.734", "0.5854", "0.7485", "0.5854", "0.7485", "0.589", "0.734", "0.589" ], [ "0.2852", "0.5913", "0.326", "0.5913", "0.326", "0.603", "0.2852", "0.603" ], [ "0.4902", "0.5913", "0.562", "0.595", "0.5605", "0.606", "0.4885", "0.6016" ], [ "0.681", "0.5913", "0.7075", "0.5913", "0.7075", "0.5996", "0.681", "0.5996" ], [ "0.6284", "0.5947", "0.6406", "0.5947", "0.6406", "0.5986", "0.6284", "0.5986" ], [ "0.1871", "0.5947", "0.2737", "0.5947", "0.2737", "0.605", "0.1871", "0.605" ], [ "0.6445", "0.5947", "0.6567", "0.5947", "0.6567", "0.5986", "0.6445", "0.5986" ], [ "0.774", "0.5947", "0.8037", "0.5947", "0.8037", "0.603", "0.774", "0.603" ], [ "0.4043", "0.595", "0.4592", "0.5933", "0.4592", "0.605", "0.4043", "0.607" ], [ "0.3381", "0.595", "0.4043", "0.595", "0.4043", "0.606", "0.3381", "0.606" ], [ "0.464", "0.595", "0.4763", "0.5933", "0.478", "0.5986", "0.4656", "0.601" ], [ "0.574", "0.595", "0.5864", "0.595", "0.5864", "0.5996", "0.574", "0.5996" ], [ "0.6807", "0.5967", "0.714", "0.5986", "0.7134", "0.61", "0.6787", "0.608" ], [ "0.722", "0.5996", "0.759", "0.5996", "0.759", "0.61", "0.722", "0.61" ], [ "0.576", "0.6016", "0.6094", "0.5996", "0.6113", "0.6113", "0.5767", "0.613" ], [ "0.2957", "0.6016", "0.3186", "0.6016", "0.3186", "0.61", "0.2957", "0.61" ], [ "0.4756", "0.604", "0.509", "0.604", "0.509", "0.6113", "0.4756", "0.6113" ], [ "0.6313", "0.6016", "0.661", "0.6016", "0.661", "0.61", "0.6313", "0.61" ], [ "0.1732", "0.627", "0.824", "0.6274", "0.824", "0.6426", "0.1732", "0.6416" ], [ "0.1748", "0.6426", "0.8228", "0.6426", "0.8228", "0.656", "0.1748", "0.656" ], [ "0.1765", "0.6577", "0.822", "0.6577", "0.822", "0.6685", "0.1765", "0.6685" ], [ "0.1765", "0.672", "0.822", "0.672", "0.822", "0.6816", "0.1765", "0.6816" ], [ "0.1748", "0.6846", "0.8228", "0.6846", "0.8228", "0.6978", "0.1748", "0.6978" ], [ "0.1748", "0.6963", "0.8228", "0.6963", "0.8228", "0.7104", "0.1748", "0.7104" ], [ "0.1732", "0.711", "0.826", "0.7104", "0.826", "0.7236", "0.1732", "0.7246" ], [ "0.1765", "0.7266", "0.824", "0.7266", "0.824", "0.7373", "0.1765", "0.7373" ], [ "0.1748", "0.7393", "0.824", "0.7393", "0.824", "0.7524", "0.1748", "0.7524" ], [ "0.1724", "0.752", "0.726", "0.7524", "0.726", "0.7666", "0.1724", "0.765" ], [ "0.1748", "0.78", "0.824", "0.78", "0.824", "0.7944", "0.1748", "0.7944" ], [ "0.1748", "0.7954", "0.8203", "0.7954", "0.8203", "0.806", "0.1748", "0.806" ], [ "0.1748", "0.8096", "0.824", "0.8096", "0.824", "0.8228", "0.1748", "0.8228" ], [ "0.1748", "0.8228", "0.694", "0.8228", "0.694", "0.836", "0.1748", "0.836" ], [ "0.1748", "0.8423", "0.824", "0.8423", "0.824", "0.856", "0.1748", "0.856" ], [ "0.1732", "0.857", "0.826", "0.856", "0.826", "0.869", "0.1732", "0.8706" ], [ "0.1748", "0.8706", "0.826", "0.8706", "0.826", "0.884", "0.1748", "0.884" ], [ "0.1732", "0.8833", "0.826", "0.884", "0.826", "0.8984", "0.1732", "0.898" ], [ "0.1732", "0.8984", "0.824", "0.8984", "0.824", "0.912", "0.1732", "0.912" ], [ "0.5", "0.9346", "0.5107", "0.9424", "0.5", "0.9507", "0.4885", "0.9424" ] ]

[ "where st+H+f are the model predictions that are fed back to be used as input again. For simplicity", "inthisnotationweassumethereward tobepartofthestate.Alsowedonotexplicitlyshowthe", "recurrence and carrying over of the hidden states.", "After having trained the two components β(s) and (fi. (s, a)), we can then move on to policy", "training. Similarly to MOOSE and WSBC, we optimize the policy e by sampling start states from D", "andperformingvirtualrolloutsthroughoutthedynamicsensembleusingthecurrentpolicycandidate", "In every step, the ensemble predicts the reward as the minimum among its members and the next state", "that goes with it. At the same time we collect the mean squared differences between the actions that", "TetookintherolloutandtheonethatBwouldhavetaken.Thelossisthencomputedasa weighted", "sum of the two components. Crucially, we sample the weighting factor randomly and pass it to the", "policy as an additional input - the policy thus needs to learn all behaviors ranging from pure behavior", "cloningtoentirelyfreeoptimization:", "L(0) = -", "", "[\\e(st, at) -(1 -X)p(at)]", "at=元e(st,入)", "(3)", "", "where we sample between o & 1, e(st, at) = minfr(fi(st, at))li e O,.., N - 1) denotes the", "outputof theensemblepredictionfor reward (we omitexplicitnotation of recurrenceforsimplicity)", "and p(at) = [β(st) - at]? denotes the penalty based on the mean squared distance between the", "original policy and the actions proposed by e. See Fig. 1 for a visualization of our proposed training", "procedure.", "三", "W", "", "", "L(0)=", "(1-A)P", "®", "", "", "A", "", "O", "", "Action", "", "Predicted", "al PolicyPolicyrt", "Policy", "Ra", "DatasetStartState", "au", "Reward", "Sampled入", "入Distribution", "Ori", "DV", "Action", "Distance", "Model", "State", "Modelβ", "Action", "Figurel:Schematic of LION policy training.Duringpolicytraining (Eq.3)onlyTe (in green)is", "adapted, while the original policy model β (orange) and the dynamics ensemble ( fu) (blue) are", "alreadytrained and remainunchanged.Fromlefttoright,wefirst samplea start state(black)from", "thedatasetand avaluefrom its distribution.Then,weletthe original policy(orange)as well asthe", "currently trained policy (green) predict actions - note that the newly trained policy is conditioned on", "X. Both actions are then compared to calculate the penalty for that timestep (red). The action from", "the currently trained policy is then also fed into the trained transition model (blue) together with the", "current state (black/blue),toget the reward forthattimestep (yellow)as well as thenext state (blue).", "This procedure is repeated until the horizon of the episode is reached. The rewards and penalties are", "then summedup and weighted by to be used as a loss function for policy training.", "We motivate our purely model-based approach (no value function involved) with the fact that we have", "fewermovingparts:Ourensemble canbekeptfixed once itistrained,whileavaluefunctionhasto", "be learned jointly with To, which is in our case more complex than usual. See experimental results in", "Fig. 10 a brief attempt at making our approach work in the model-free domain.", "In addition to Eq. 3, we need to penalize divergence not only from the learned model of the original", "policy during virtual roll-outs, but also from the actual actions in the dataset at = o. It seems that if", "this is not done, the trained policy sticks to the (also trained) original policy β during the rollouts,", "but during those rollouts, there are states that did not appear in the original dataset, enabling r to", "actually diverge from the true trajectory distribution. We thus penalize both rollout as well as data", "X" ]

{ "arxiv_id": "2205.10629", "language": "en", "page_idx": 3 }

[ [ "0.1748", "0.09216", "0.3105", "0.09216", "0.3105", "0.1061", "0.1748", "0.1061" ], [ "0.2255", "0.1237", "0.2769", "0.1237", "0.2769", "0.1383", "0.2255", "0.1383" ], [ "0.3464", "0.125", "0.5557", "0.1237", "0.5557", "0.1383", "0.3464", "0.1395" ], [ "0.6465", "0.125", "0.774", "0.1225", "0.774", "0.137", "0.6465", "0.1395" ], [ "0.8022", "0.1256", "0.824", "0.1256", "0.824", "0.1383", "0.8022", "0.1383" ], [ "0.6006", "0.132", "0.6113", "0.1288", "0.613", "0.1338", "0.601", "0.1364" ], [ "0.614", "0.1433", "0.653", "0.1401", "0.6543", "0.1515", "0.615", "0.1547" ], [ "0.556", "0.1276", "0.581", "0.1276", "0.581", "0.1351", "0.556", "0.1351" ], [ "0.2974", "0.1445", "0.327", "0.1445", "0.327", "0.1515", "0.2974", "0.1515" ], [ "0.1732", "0.1578", "0.824", "0.1578", "0.824", "0.1724", "0.1732", "0.1724" ], [ "0.1748", "0.1724", "0.824", "0.1724", "0.824", "0.1863", "0.1748", "0.1863" ], [ "0.1748", "0.1863", "0.824", "0.1863", "0.824", "0.1995", "0.1748", "0.1995" ], [ "0.1732", "0.2008", "0.824", "0.2008", "0.824", "0.214", "0.1732", "0.214" ], [ "0.1748", "0.2134", "0.824", "0.2134", "0.824", "0.2267", "0.1748", "0.2267" ], [ "0.1748", "0.2267", "0.5923", "0.2267", "0.5923", "0.2412", "0.1748", "0.2412" ], [ "0.1748", "0.2474", "0.824", "0.2474", "0.824", "0.2615", "0.1748", "0.2615" ], [ "0.1765", "0.2632", "0.8228", "0.2632", "0.8228", "0.2734", "0.1765", "0.2734" ], [ "0.1724", "0.2747", "0.8228", "0.2747", "0.8228", "0.2886", "0.1724", "0.2886" ], [ "0.1748", "0.289", "0.824", "0.289", "0.824", "0.303", "0.1748", "0.303" ], [ "0.1732", "0.303", "0.826", "0.303", "0.826", "0.3164", "0.1732", "0.3164" ], [ "0.1732", "0.3176", "0.6323", "0.3176", "0.6323", "0.3308", "0.1732", "0.3308" ], [ "0.1732", "0.3467", "0.2998", "0.3467", "0.2998", "0.361", "0.1732", "0.361" ], [ "0.1765", "0.3738", "0.4182", "0.3738", "0.4182", "0.3838", "0.1765", "0.3838" ], [ "0.1732", "0.3865", "0.4207", "0.3865", "0.4207", "0.3997", "0.1732", "0.3997" ], [ "0.433", "0.3894", "0.634", "0.3894", "0.634", "0.4028", "0.433", "0.4028" ], [ "0.1748", "0.4016", "0.4182", "0.4016", "0.4182", "0.4124", "0.1748", "0.4124" ], [ "0.4412", "0.4072", "0.826", "0.4072", "0.826", "0.4204", "0.4412", "0.4204" ], [ "0.1732", "0.4136", "0.4192", "0.4136", "0.4192", "0.4268", "0.1732", "0.4268" ], [ "0.4592", "0.4204", "0.826", "0.4204", "0.826", "0.4343", "0.4592", "0.4343" ], [ "0.1765", "0.4287", "0.4207", "0.4287", "0.4207", "0.4395", "0.1765", "0.4395" ], [ "0.46", "0.4343", "0.722", "0.4343", "0.722", "0.4475", "0.46", "0.4475" ], [ "0.1732", "0.4414", "0.4207", "0.4414", "0.4207", "0.455", "0.1732", "0.455" ], [ "0.4412", "0.4475", "0.826", "0.4475", "0.826", "0.461", "0.4412", "0.461" ], [ "0.1748", "0.457", "0.2418", "0.457", "0.2418", "0.4673", "0.1748", "0.4673" ], [ "0.4617", "0.4634", "0.7935", "0.4634", "0.7935", "0.4734", "0.4617", "0.4734" ], [ "0.1847", "0.476", "0.2834", "0.476", "0.2834", "0.4873", "0.1847", "0.4873" ], [ "0.2754", "0.473", "0.4192", "0.4734", "0.4192", "0.488", "0.2754", "0.4873" ], [ "0.4412", "0.476", "0.819", "0.476", "0.819", "0.4893", "0.4412", "0.4893" ], [ "0.1748", "0.4937", "0.4224", "0.4937", "0.4224", "0.507", "0.1748", "0.507" ], [ "0.4395", "0.4905", "0.7915", "0.4905", "0.7915", "0.505", "0.4395", "0.505" ], [ "0.1732", "0.507", "0.4224", "0.507", "0.4224", "0.521", "0.1732", "0.521" ], [ "0.4395", "0.506", "0.5884", "0.506", "0.5884", "0.5195", "0.4395", "0.5195" ], [ "0.1724", "0.5195", "0.4207", "0.521", "0.4207", "0.535", "0.1724", "0.534" ], [ "0.4412", "0.523", "0.455", "0.523", "0.455", "0.532", "0.4412", "0.532" ], [ "0.4836", "0.521", "0.664", "0.5176", "0.664", "0.532", "0.4836", "0.535" ], [ "0.1748", "0.535", "0.4207", "0.535", "0.4207", "0.549", "0.1748", "0.549" ], [ "0.4429", "0.535", "0.4558", "0.535", "0.4558", "0.5454", "0.4429", "0.5454" ], [ "0.4844", "0.534", "0.7363", "0.534", "0.7363", "0.5474", "0.4844", "0.5474" ], [ "0.1732", "0.549", "0.4224", "0.549", "0.4224", "0.5625", "0.1732", "0.5625" ], [ "0.4412", "0.55", "0.4558", "0.55", "0.4558", "0.56", "0.4412", "0.56" ], [ "0.486", "0.5474", "0.687", "0.547", "0.687", "0.56", "0.486", "0.561" ], [ "0.1765", "0.563", "0.4192", "0.563", "0.4192", "0.5737", "0.1765", "0.5737" ], [ "0.4412", "0.563", "0.455", "0.563", "0.455", "0.5737", "0.4412", "0.5737" ], [ "0.4844", "0.5625", "0.6094", "0.5625", "0.6094", "0.5728", "0.4844", "0.5728" ], [ "0.1732", "0.5757", "0.4224", "0.5757", "0.4224", "0.589", "0.1732", "0.589" ], [ "0.4307", "0.577", "0.4558", "0.577", "0.4558", "0.5884", "0.4307", "0.5884" ], [ "0.509", "0.5757", "0.7646", "0.5757", "0.7646", "0.589", "0.509", "0.589" ], [ "0.1732", "0.589", "0.423", "0.5884", "0.423", "0.603", "0.1732", "0.604" ], [ "0.4307", "0.5903", "0.4558", "0.5903", "0.4558", "0.6016", "0.4307", "0.6016" ], [ "0.5073", "0.587", "0.7754", "0.589", "0.7754", "0.604", "0.5073", "0.6016" ], [ "0.1748", "0.604", "0.4224", "0.604", "0.4224", "0.617", "0.1748", "0.617" ], [ "0.433", "0.604", "0.4558", "0.604", "0.4558", "0.616", "0.433", "0.616" ], [ "0.5063", "0.6016", "0.814", "0.601", "0.814", "0.6187", "0.5063", "0.619" ], [ "0.1732", "0.617", "0.3284", "0.617", "0.3284", "0.631", "0.1732", "0.631" ], [ "0.433", "0.619", "0.4575", "0.619", "0.4575", "0.632", "0.433", "0.632" ], [ "0.509", "0.619", "0.6562", "0.619", "0.6562", "0.634", "0.509", "0.634" ], [ "0.1748", "0.631", "0.4192", "0.631", "0.4192", "0.646", "0.1748", "0.646" ], [ "0.5254", "0.6353", "0.755", "0.6353", "0.755", "0.6494", "0.5254", "0.6494" ], [ "0.1772", "0.6465", "0.4192", "0.6465", "0.4192", "0.657", "0.1772", "0.657" ], [ "0.433", "0.655", "0.4558", "0.655", "0.4558", "0.6665", "0.433", "0.6665" ], [ "0.511", "0.654", "0.7427", "0.654", "0.7427", "0.6675", "0.511", "0.6675" ], [ "0.482", "0.6704", "0.778", "0.6694", "0.778", "0.6846", "0.482", "0.685" ], [ "0.1732", "0.659", "0.4224", "0.659", "0.4224", "0.6724", "0.1732", "0.6724" ], [ "0.1748", "0.6724", "0.4248", "0.6724", "0.4248", "0.686", "0.1748", "0.686" ], [ "0.423", "0.672", "0.4558", "0.672", "0.4558", "0.683", "0.423", "0.683" ], [ "0.1748", "0.686", "0.423", "0.686", "0.423", "0.6997", "0.1748", "0.6997" ], [ "0.429", "0.685", "0.536", "0.688", "0.536", "0.7026", "0.429", "0.6997" ], [ "0.1748", "0.7007", "0.4207", "0.7007", "0.4207", "0.714", "0.1748", "0.714" ], [ "0.1724", "0.712", "0.8228", "0.712", "0.8228", "0.7266", "0.1724", "0.7266" ], [ "0.1748", "0.729", "0.8228", "0.729", "0.8228", "0.7393", "0.1748", "0.7393" ], [ "0.1732", "0.7407", "0.8228", "0.7407", "0.8228", "0.754", "0.1732", "0.754" ], [ "0.1732", "0.755", "0.6978", "0.755", "0.6978", "0.7695", "0.1732", "0.7695" ], [ "0.1772", "0.7915", "0.1912", "0.7915", "0.1912", "0.8", "0.1772", "0.8" ], [ "0.2018", "0.7886", "0.3137", "0.7886", "0.3137", "0.803", "0.2018", "0.803" ], [ "0.1748", "0.8184", "0.824", "0.8184", "0.824", "0.832", "0.1748", "0.832" ], [ "0.1748", "0.8335", "0.824", "0.8335", "0.824", "0.8467", "0.1748", "0.8467" ], [ "0.1732", "0.8467", "0.8267", "0.8467", "0.8267", "0.86", "0.1732", "0.86" ], [ "0.1732", "0.8594", "0.5107", "0.86", "0.5107", "0.875", "0.1732", "0.8735" ], [ "0.2197", "0.884", "0.826", "0.884", "0.826", "0.8984", "0.2197", "0.8984" ], [ "0.2336", "0.8994", "0.7686", "0.8994", "0.7686", "0.91", "0.2336", "0.91" ], [ "0.4944", "0.936", "0.5063", "0.936", "0.5063", "0.9487", "0.4944", "0.9487" ] ]

[ "divergence at 入= O:", "L(0) =", "t[e(st, at) -(1 -)p(at)]", "[a -π(s, 入= 0)]2", "(4)", "", "8,a~D", "+", "Sa~D", "where n controls the penalty weight for not following dataset actions at A = O, see Appendix A for", "more details. Furthermore, we normalize states to have zero mean and unit variance during every", "forward pass through dynamics model or policy, using the mean and standard deviation observed", "in the dataset. We also normalize the rewards provided by the ensemble rt = e(st, at), so that they", "live in the same magnitude as the action penalties (we assume actions to be in -1, 1jD, so that the", "penalty can be in [0, 4] D where D is the action dimensionality).", "Intuitively, one might choose to sample uniformly between zero and one, however instead we", "chooseabetadistributionwithparameters(0.1,0.1),whichcouldbecalledbathtub-shaped.Similarly", "to (Seo et al., 2021), we find that it is important to put emphasis on the edge cases, so that the extreme", "behavior is properly learned, rather than putting equal probability mass on each value in the [O, 1]", "range. The interpolation between the edges seems to be easier and thus require less samples.Fig. 11", "shows policy results for different lambda distributions during training", "3.2Deployment", "Atinferencetime.thetrained policy", "can at any point be influenced by the", "Algorithm 1 LION (Training)", "userthatwould otherwisebe incon", "I: Require Dataset D = {T:], randomly initialized param-", "trol of the system, by choosing the ^", "eters , , , lambda distribution parameters Beta(a, b),", "thatis passed to thepolicytogether", "horizon H, number of policy updates U", "with the current system state to obtain", "2:// dynamics and original policy models can betrained", "an action:", "supervisedandindependently of othercomponents", "at=e(St,)", ") 入E User(st). (5)", "3: train original policy model β using D and Equation 1", "He or she may choose to be conserva-", "4: train dynamics models fi with D and Equation 2", "tive or adventurous, observe the feed-", "5: for j in 1..Udo", "back and always adjust the proximity", "6:", "sample start states So ~ D", "parameter of the policy accordingly.", "7:", "sample lambda values ^ ~ Beta(a, b)", "At this point, any disliked behavior", "8:", "initializepolicylossL(0)=0", "canimmediatelybecorrectedwithout", "9:", "fort in o..Hdo", "any time loss due to re-training and", "10:", "calculate policy actions at =Te(st,)", "deploying a new policy, even if the", "11:", "calculate behavioral actions bt=β(st)", "user's specific preferences were not", "12:", "calculate penalty term p(at) = [B(st) -at]?", "known at training time.", "13:", "rt, St+1 = fi (st, at)", "We propose to initially start with", "s.t. i = argmin,fr(fi (st,at)))", "入=0duringdeployment,inorder", "14:", "L(0)+ =-t[Art -(1-p(at)l", "update e using gradient VeL(0) and Adam", "to check whether the policy is actu-", "ally able to reproduce the original pol-", "15:", "icy and to gain the user's trust in the", "16: return 0;", "found solution. Then, depending on", "how critical failures are and how much time is at hand, A may be increased in small steps for as long", "astheuserisstillcomfortablewiththeobservedbehavior.Figure3showsanexampleofhowthe", "policy behavior changes over the course of . Once the performance stops to increase or the user is", "otherwise not satisfied, we can immediately return to the last satisfying A value.", "4", "Experiments", "At first, we intuitively showcase LION in a simple 2D-world in order to get an understanding of how", "the policy changes its behavior based on ). Afterwards, we move to a more serious test, evaluating", "our algorithm on the 16 industrial benchmark (IB) datasets (Hein et al., 2017a; Swazinna et al.,", "2021b).We aim to answer the following questions:", ": Do LION policies behave as expected, i.e. do they reproduce the original policy at ^ = 0", "anddeviatemoreandmorefromitwithincreasedfreedomtooptimizeforreturn?", "5" ]

{ "arxiv_id": "2205.10629", "language": "en", "page_idx": 4 }

[ [ "0.1748", "0.0341", "0.4795", "0.0341", "0.4795", "0.04736", "0.1748", "0.04736" ], [ "0.1765", "0.101", "0.7295", "0.1016", "0.7295", "0.1193", "0.1765", "0.1187" ], [ "0.1871", "0.1427", "0.71", "0.1427", "0.71", "0.1604", "0.1871", "0.1604" ], [ "0.183", "0.1604", "0.818", "0.159", "0.818", "0.1736", "0.183", "0.1749" ], [ "0.183", "0.1755", "0.2507", "0.1755", "0.2507", "0.1901", "0.183", "0.1901" ], [ "0.183", "0.1901", "0.4592", "0.1901", "0.4592", "0.2052", "0.183", "0.2052" ], [ "0.183", "0.2039", "0.4265", "0.2058", "0.4265", "0.2203", "0.183", "0.2184" ], [ "0.183", "0.2203", "0.5", "0.2203", "0.5", "0.2343", "0.183", "0.2343" ], [ "0.02777", "0.2715", "0.05392", "0.2715", "0.05392", "0.289", "0.02777", "0.289" ], [ "0.4534", "0.2708", "0.546", "0.2708", "0.546", "0.2854", "0.4534", "0.2854" ], [ "0.02287", "0.279", "0.06046", "0.279", "0.06046", "0.3738", "0.02287", "0.3738" ], [ "0.2336", "0.3018", "0.767", "0.3018", "0.767", "0.315", "0.2336", "0.315" ], [ "0.2336", "0.3164", "0.7656", "0.3164", "0.7656", "0.33", "0.2336", "0.33" ], [ "0.2336", "0.33", "0.7656", "0.33", "0.7656", "0.3435", "0.2336", "0.3435" ], [ "0.2336", "0.3435", "0.7656", "0.3435", "0.7656", "0.3567", "0.2336", "0.3567" ], [ "0.2336", "0.358", "0.767", "0.358", "0.767", "0.3718", "0.2336", "0.3718" ], [ "0.02696", "0.3687", "0.05554", "0.3687", "0.05554", "0.3977", "0.02696", "0.3977" ], [ "0.2345", "0.3718", "0.7656", "0.3718", "0.7656", "0.385", "0.2345", "0.385" ], [ "0.2345", "0.385", "0.767", "0.385", "0.767", "0.3984", "0.2345", "0.3984" ], [ "0.02124", "0.404", "0.06046", "0.404", "0.06046", "0.4873", "0.02124", "0.4873" ], [ "0.2336", "0.3997", "0.763", "0.3997", "0.763", "0.4136", "0.2336", "0.4136" ], [ "0.2361", "0.414", "0.7646", "0.414", "0.7646", "0.4248", "0.2361", "0.4248" ], [ "0.2336", "0.4268", "0.4478", "0.4268", "0.4478", "0.44", "0.2336", "0.44" ], [ "0.1814", "0.4685", "0.1871", "0.473", "0.1814", "0.4766", "0.1765", "0.473" ], [ "0.1993", "0.4673", "0.3367", "0.4673", "0.3367", "0.481", "0.1993", "0.481" ], [ "0.02124", "0.4912", "0.05963", "0.4912", "0.05963", "0.712", "0.02124", "0.712" ], [ "0.1748", "0.4976", "0.824", "0.4976", "0.824", "0.511", "0.1748", "0.511" ], [ "0.1748", "0.512", "0.824", "0.512", "0.824", "0.526", "0.1748", "0.526" ], [ "0.1748", "0.526", "0.824", "0.526", "0.824", "0.539", "0.1748", "0.539" ], [ "0.1748", "0.539", "0.824", "0.539", "0.824", "0.553", "0.1748", "0.553" ], [ "0.1732", "0.553", "0.824", "0.553", "0.824", "0.5664", "0.1732", "0.5664" ], [ "0.1748", "0.5674", "0.824", "0.5674", "0.824", "0.5806", "0.1748", "0.5806" ], [ "0.1748", "0.582", "0.826", "0.582", "0.826", "0.595", "0.1748", "0.595" ], [ "0.1748", "0.595", "0.7837", "0.595", "0.7837", "0.6094", "0.1748", "0.6094" ], [ "0.1748", "0.616", "0.824", "0.616", "0.824", "0.631", "0.1748", "0.631" ], [ "0.1748", "0.63", "0.826", "0.63", "0.826", "0.6445", "0.1748", "0.6445" ], [ "0.1748", "0.6445", "0.826", "0.6445", "0.826", "0.6577", "0.1748", "0.6577" ], [ "0.1732", "0.657", "0.824", "0.657", "0.824", "0.6704", "0.1732", "0.6704" ], [ "0.1748", "0.6704", "0.824", "0.6704", "0.824", "0.6846", "0.1748", "0.6846" ], [ "0.1748", "0.6875", "0.824", "0.6875", "0.824", "0.6978", "0.1748", "0.6978" ], [ "0.1748", "0.6997", "0.826", "0.6997", "0.826", "0.7134", "0.1748", "0.7134" ], [ "0.1732", "0.7134", "0.7485", "0.712", "0.7485", "0.7266", "0.1732", "0.728" ], [ "0.1748", "0.7344", "0.8228", "0.7344", "0.8228", "0.7485", "0.1748", "0.7485" ], [ "0.1748", "0.7485", "0.8228", "0.7485", "0.8228", "0.762", "0.1748", "0.762" ], [ "0.1765", "0.7646", "0.8228", "0.7646", "0.8228", "0.7744", "0.1765", "0.7744" ], [ "0.1732", "0.7764", "0.8228", "0.7764", "0.8228", "0.7905", "0.1732", "0.7905" ], [ "0.1748", "0.7905", "0.8228", "0.7905", "0.8228", "0.8037", "0.1748", "0.8037" ], [ "0.1748", "0.8047", "0.7754", "0.8047", "0.7754", "0.8184", "0.1748", "0.8184" ], [ "0.1765", "0.839", "0.1896", "0.839", "0.1896", "0.8496", "0.1765", "0.8496" ], [ "0.2035", "0.837", "0.327", "0.839", "0.327", "0.854", "0.2035", "0.8516" ], [ "0.1732", "0.869", "0.826", "0.869", "0.826", "0.8833", "0.1732", "0.8833" ], [ "0.1748", "0.884", "0.826", "0.884", "0.826", "0.898", "0.1748", "0.898" ], [ "0.1748", "0.898", "0.824", "0.898", "0.824", "0.911", "0.1748", "0.911" ], [ "0.1732", "0.911", "0.8228", "0.911", "0.8228", "0.9253", "0.1732", "0.9253" ], [ "0.4927", "0.9497", "0.505", "0.9497", "0.505", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2018", "PARAMETER SPACE NOISE FOR EXPLORATION", "Matthias Plappertt, Rein Houthooftt, Prafulla Dhariwalt, Szymon Sidort,", "Richard Y. Chent, Xi Chentt, Tamim Asfourt, Pieter Abbeeltt, and Marcin Andrychowiczt", "t OpenAI", "+ Karlsruhe Institute of Technology (KIT)", "tt Universityof California, Berkeley", "Correspondencetomatthias@openai.com", "8", "ABSTRACT", "310", "Deep reinforcement learning (RL) methods generally engage in exploratory be-", "havior through noise injection in the action space. An alternative is to add noise", "directly to the agent's parameters, which can lead to more consistent exploration", "and a richer set of behaviors. Methods such as evolutionary strategies use parameter", "perturbations, but discard all temporal structure in the process and require signif-", "3", "icantly more samples. Combining parameter noise with traditional RL methods", "allows to combine the best of both worlds. We demonstrate that both off- and", "[T'S]", "on-policy methods benefit from this approach through experimental comparison", "ofDQN,DDPG,andTRPOonhigh-dimensionaldiscreteactionenvironmentsas", "well as continuous control tasks.", "", "INTRODUCTION", "ZAS010901:AII0", "Exploration remains a key challenge in contemporary deep reinforcement learning (RL). Its main", "purpose is to ensure that the agent's behavior does not converge prematurely to a local optimum", "Enabling efficient and effective exploration is, however, not trivial since it is not directed by the", "reward function of the underlying Markov decision process (MDP). Although a plethora of methods", "have been proposed to tackle this challenge in high-dimensional and/or continuous-action MDPs,", "they often rely on complex additional structures such as counting tables (Tang et al., 2016), density", "modeling of the state space (Ostrovski et al., 2017), learned dynamics models (Houthooft et al., 2016;", "Achiam & Sastry, 2017; Stadie et al., 2015), or self-supervised curiosity (Pathak et al., 2017)", "An orthogonal way of increasing the exploratory nature of these algorithms is through the addition", "of temporally-correlated noise, for example as done in bootstrapped DQN (Osband et al., 2016a).", "Along the same lines, it was shown that the addition of parameter noise leads to better exploration by", "obtaining a policy that exhibits a larger variety of behaviors (Sun et al., 2009b; Salimans et al., 2017)", "We discuss these related approaches in greater detail in Section 5. Their main limitation, however,", "isthattheyareeitheronlyproposedandevaluatedfortheon-policysettingwithrelativelysmall", "and shallow function approximators (RickstieB et al., 2008) or disregard all temporal structure and", "gradient information (Salimans et al., 2017; Kober & Peters, 2008; Sehnke et al., 2010).", "This paper investigates how parameter space noise can be effectively combined with off-the-shelf deep", "RL algorithms such as DQN (Mnih et al., 2015), DDPG (Lillicrap et al., 2015), and TRPO (Schulman", "etal.,2ol5b)toimprovetheirexploratorybehavior.Experimentsshowthatthisformofexploration", "is applicable to both high-dimensional discrete environments and continuous control tasks, using", "on- and off-policy methods. Our results indicate that parameter noise outperforms traditional action", "space noise-based baselines, especially in tasks where the reward signal is extremely sparse.", "2", "BACKGROUND", "We consider the standard RL framework consisting of an agent interacting with an environment.", "To simplify the exposition we assume that the environment is fully observable. An environment is", "modeled as a Markov decision process (MDP) and is defined by a set of states S, a set of actions", "A, a distribution over initial states p(so), a reward function r : S x A→ R, transition probabilities", "1" ]

{ "arxiv_id": "1706.01905", "language": "en", "page_idx": 0 }

[ [ "0.1732", "0.03284", "0.4795", "0.0322", "0.4795", "0.04672", "0.1732", "0.04736" ], [ "0.1765", "0.1048", "0.822", "0.1048", "0.822", "0.1193", "0.1765", "0.1193" ], [ "0.1765", "0.1193", "0.8228", "0.1193", "0.8228", "0.1332", "0.1765", "0.1332" ], [ "0.1765", "0.1364", "0.8164", "0.1364", "0.8164", "0.1509", "0.1765", "0.1509" ], [ "0.1732", "0.1509", "0.826", "0.1509", "0.826", "0.1654", "0.1732", "0.1654" ], [ "0.1748", "0.1674", "0.706", "0.1674", "0.706", "0.1818", "0.1748", "0.1818" ], [ "0.1748", "0.1976", "0.3774", "0.1976", "0.3774", "0.2108", "0.1748", "0.2108" ], [ "0.1748", "0.2235", "0.8228", "0.2235", "0.8228", "0.2374", "0.1748", "0.2374" ], [ "0.1732", "0.2374", "0.824", "0.2361", "0.824", "0.2507", "0.1732", "0.252" ], [ "0.1748", "0.252", "0.6543", "0.252", "0.6543", "0.2651", "0.1748", "0.2651" ], [ "0.1732", "0.2798", "0.824", "0.281", "0.824", "0.2954", "0.1732", "0.2942" ], [ "0.1748", "0.2954", "0.714", "0.2954", "0.714", "0.3093", "0.1748", "0.3093" ], [ "0.357", "0.3145", "0.642", "0.3145", "0.642", "0.332", "0.357", "0.332" ], [ "0.1748", "0.3423", "0.8228", "0.3423", "0.8228", "0.3567", "0.1748", "0.3567" ], [ "0.1748", "0.3567", "0.8228", "0.356", "0.8228", "0.3706", "0.1748", "0.3718" ], [ "0.1765", "0.3706", "0.8228", "0.3706", "0.8228", "0.385", "0.1765", "0.385" ], [ "0.1748", "0.3838", "0.649", "0.3838", "0.649", "0.3977", "0.1748", "0.3977" ], [ "0.1724", "0.411", "0.824", "0.4124", "0.824", "0.43", "0.1724", "0.4287" ], [ "0.1748", "0.4287", "0.822", "0.4287", "0.822", "0.4426", "0.1748", "0.4426" ], [ "0.1732", "0.4414", "0.4429", "0.4426", "0.4429", "0.457", "0.1732", "0.4558" ], [ "0.353", "0.4634", "0.6465", "0.4634", "0.6465", "0.4778", "0.353", "0.4778" ], [ "0.1748", "0.483", "0.8228", "0.483", "0.8228", "0.4976", "0.1748", "0.4976" ], [ "0.1748", "0.4976", "0.822", "0.4976", "0.822", "0.512", "0.1748", "0.512" ], [ "0.1765", "0.512", "0.824", "0.512", "0.824", "0.5264", "0.1765", "0.5264" ], [ "0.1724", "0.526", "0.6406", "0.5244", "0.6406", "0.539", "0.1724", "0.5405" ], [ "0.1772", "0.5596", "0.2035", "0.5596", "0.2035", "0.5674", "0.1772", "0.5674" ], [ "0.2115", "0.556", "0.3718", "0.556", "0.3718", "0.5693", "0.2115", "0.5693" ], [ "0.1732", "0.582", "0.822", "0.582", "0.822", "0.595", "0.1732", "0.595" ], [ "0.1765", "0.5967", "0.822", "0.5967", "0.822", "0.61", "0.1765", "0.61" ], [ "0.1748", "0.61", "0.8228", "0.61", "0.8228", "0.6235", "0.1748", "0.6235" ], [ "0.1748", "0.625", "0.789", "0.625", "0.789", "0.638", "0.1748", "0.638" ], [ "0.1765", "0.653", "0.824", "0.653", "0.824", "0.6675", "0.1765", "0.6675" ], [ "0.1732", "0.6665", "0.824", "0.6665", "0.824", "0.6797", "0.1732", "0.6797" ], [ "0.1748", "0.681", "0.701", "0.681", "0.701", "0.696", "0.1748", "0.696" ], [ "0.5186", "0.6987", "0.5923", "0.7007", "0.591", "0.7197", "0.517", "0.7173" ], [ "0.3381", "0.7056", "0.5156", "0.711", "0.5146", "0.73", "0.3367", "0.7246" ], [ "0.581", "0.7104", "0.6284", "0.7104", "0.6284", "0.726", "0.581", "0.726" ], [ "0.5215", "0.7217", "0.576", "0.7217", "0.576", "0.733", "0.5215", "0.733" ], [ "0.3652", "0.735", "0.3896", "0.735", "0.3896", "0.7476", "0.3652", "0.7476" ], [ "0.423", "0.7344", "0.675", "0.7344", "0.675", "0.7485", "0.423", "0.7485" ], [ "0.1765", "0.754", "0.822", "0.754", "0.822", "0.768", "0.1765", "0.768" ], [ "0.1732", "0.7676", "0.822", "0.7676", "0.822", "0.781", "0.1732", "0.781" ], [ "0.1748", "0.7827", "0.7075", "0.7827", "0.7075", "0.797", "0.1748", "0.797" ], [ "0.1724", "0.8164", "0.5923", "0.8174", "0.5923", "0.832", "0.1724", "0.8306" ], [ "0.1732", "0.845", "0.824", "0.8477", "0.824", "0.8623", "0.1732", "0.86" ], [ "0.1748", "0.8623", "0.822", "0.8623", "0.822", "0.8755", "0.1748", "0.8755" ], [ "0.1978", "0.8853", "0.826", "0.8853", "0.826", "0.8994", "0.1978", "0.8994" ], [ "0.1748", "0.898", "0.2942", "0.898", "0.2942", "0.911", "0.1748", "0.911" ], [ "0.1936", "0.911", "0.6763", "0.91", "0.6763", "0.925", "0.1936", "0.9253" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.9634", "0.492", "0.9634" ] ]

[ "Published as a conference paper at ICLR 2018", "p(st+1st, at), a time horizon T, and a discount factor E [O, 1). We denote by g a policy", "parametrized by 0, which can be either deterministic, ☆ : S → A, or stochastic, π : S → P(A). The", "agent's goal is to maximize the expected discounted return n(πe) = E,[t=o tr(st, at)], where T :", "(so, ao,..., sT) denotes a trajectory with so ~ p(so), at ~ Te(atst), and st+1 ~ p(st+i/st, at).", "Experimental evaluation is based on the undiscounted return E- [t=o r(st, at)].", "2.1 OFF-POLICYMETHODS", "Off-policy RL methods allow learning based on data captured by arbitrary policies. This paper", "considers two popular off-policy algorithms, namely Deep Q-Networks (DQN, Mnih et al. (2015))", "and Deep Deterministic Policy Gradients (DDPG, Lillicrap et al. (2015)).", "Deep Q-Networks (DQN)DQN uses a deep neural network as a function approximator to estimate", "the optimal Q-value function, which conforms to the Bellman optimality equation:", "Q(st, at)= r(st, at) +maxQ(st+1, a').", "The policy is implicitly defined by Q as (st) = argmaxa'eAQ(st,a'). Typically, a stochastic e", "greedy or Boltzmann policy (Sutton & Barto, 1998)is derived from the Q-value function to encourage", "exploration, which relies on sampling noise in the action space. The Q-network predicts a Q-value", "for each action and is updated using off-policy data from a replay buffer.", "Deep Deterministic Policy Gradients (DDPG)DDPG is an actor-critic algorithm, applicable to", "continuous action spaces. Similar to DQN, the critic estimates the Q-value function using off-policy", "data and therecursive Bellman equation:", "Q(st,at)=r(st,at) +Q(st+1,元e(st+1))", "where e is the actor or policy. The actor is trained to maximize the critic's estimated Q-values by", "back-propagating through both networks. For exploration, DDPG uses a stochastic policy of the", "form 元(st), = πe(st) + w, where w is either w ~ N(O,α21) (uncorrelated) or w ~ OU(0, α2)", "(correlated).2 Again, exploration is realized through action space noise.", "2.2", "ON-POLICY METHODS", "In contrast to off-policy algorithms, on-policy methods require updating function approximators", "according to the currently followed policy. In particular, we will consider Trust Region Policy", "Optimization (TRPO, Schulman et al. (2015a), an extension of traditional policy gradient methods", "(Williams, 1992b) using the natural gradient direction (Peters & Schaal,2008; Kakade,2001).", "Trust Region Policy Optimization (TRPO) TRPO improves upon REINFORCE (Williams,", "1992b) by computing an ascent direction that ensures a small change in the policy distribution.", "More specifically, TRPO solves the following constrained optimization problem:", "[e(a]s) A", "maximizegEs~Pa,a~To'", "A(s,a)", "Te(as", "s.t.", "Es~p[Dk((s)]/(|s))]≤ 0kL", "where pe = P。 is the discounted state-visitation frequencies induced by g, A(s,a) denotes the", "advantage function estimated by the empirical return minus the baseline, and kL is a step size", "parameter which controls how much the policy is allowed to change per iteration.", "3 PARAMETERSPACENOISEFOREXPLORATION", "This work considers policies that are realized as parameterized functions, which we denote as e.", "with being the parameter vector. We represent policies as neural networks but our technique can", "'If t = T, we write r(sT, ar) to denote the terminal reward, even though it has has no dependence on aT,", "to simplify notation.", "20U(,) denotes the Ornstein-Uhlenbeck process (Uhlenbeck & Ornstein, 1930)", "2" ]

{ "arxiv_id": "1706.01905", "language": "en", "page_idx": 1 }

[ [ "0.1732", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1732", "0.04736" ], [ "0.1748", "0.1048", "0.8228", "0.1048", "0.8228", "0.1187", "0.1748", "0.1187" ], [ "0.1748", "0.1193", "0.8267", "0.1193", "0.8267", "0.1332", "0.1748", "0.1332" ], [ "0.1748", "0.1351", "0.8228", "0.1351", "0.8228", "0.1483", "0.1748", "0.1483" ], [ "0.1732", "0.1477", "0.8228", "0.1497", "0.8228", "0.1635", "0.1732", "0.161" ], [ "0.1724", "0.1622", "0.4592", "0.1622", "0.4592", "0.1755", "0.1724", "0.1755" ], [ "0.1748", "0.1925", "0.3816", "0.1925", "0.3816", "0.2058", "0.1748", "0.2058" ], [ "0.3733", "0.1913", "0.8228", "0.1913", "0.8228", "0.2052", "0.3733", "0.2052" ], [ "0.1732", "0.207", "0.8228", "0.207", "0.8228", "0.2203", "0.1732", "0.2203" ], [ "0.1765", "0.2229", "0.8228", "0.2229", "0.8228", "0.2329", "0.1765", "0.2329" ], [ "0.1748", "0.2349", "0.8228", "0.2349", "0.8228", "0.2488", "0.1748", "0.2488" ], [ "0.1748", "0.2488", "0.8228", "0.2488", "0.8228", "0.262", "0.1748", "0.262" ], [ "0.1765", "0.2644", "0.822", "0.2644", "0.822", "0.2747", "0.1765", "0.2747" ], [ "0.1732", "0.276", "0.822", "0.276", "0.822", "0.289", "0.1732", "0.289" ], [ "0.1748", "0.2903", "0.824", "0.2903", "0.824", "0.3037", "0.1748", "0.3037" ], [ "0.1748", "0.3037", "0.822", "0.305", "0.822", "0.3164", "0.1748", "0.315" ], [ "0.1732", "0.3176", "0.509", "0.3176", "0.509", "0.3308", "0.1732", "0.3308" ], [ "0.1732", "0.3467", "0.824", "0.3467", "0.824", "0.361", "0.1732", "0.361" ], [ "0.1732", "0.3623", "0.824", "0.361", "0.824", "0.375", "0.1732", "0.3757" ], [ "0.1732", "0.375", "0.822", "0.375", "0.822", "0.3882", "0.1732", "0.3882" ], [ "0.1748", "0.3894", "0.8228", "0.3894", "0.8228", "0.4028", "0.1748", "0.4028" ], [ "0.1765", "0.4048", "0.822", "0.4048", "0.822", "0.4153", "0.1765", "0.4153" ], [ "0.1748", "0.4172", "0.8228", "0.4172", "0.8228", "0.4312", "0.1748", "0.4312" ], [ "0.1732", "0.4312", "0.556", "0.4312", "0.556", "0.4443", "0.1732", "0.4443" ], [ "0.1748", "0.461", "0.8228", "0.461", "0.8228", "0.4749", "0.1748", "0.4749" ], [ "0.1748", "0.4749", "0.8228", "0.4749", "0.8228", "0.488", "0.1748", "0.488" ], [ "0.1732", "0.488", "0.826", "0.4873", "0.826", "0.502", "0.1732", "0.5024" ], [ "0.1748", "0.504", "0.8228", "0.504", "0.8228", "0.5146", "0.1748", "0.5146" ], [ "0.1732", "0.5166", "0.8228", "0.5166", "0.8228", "0.53", "0.1732", "0.53" ], [ "0.1732", "0.529", "0.822", "0.529", "0.822", "0.5425", "0.1732", "0.5425" ], [ "0.1748", "0.5454", "0.822", "0.5454", "0.822", "0.556", "0.1748", "0.556" ], [ "0.1748", "0.558", "0.8228", "0.558", "0.8228", "0.5713", "0.1748", "0.5713" ], [ "0.1748", "0.5728", "0.824", "0.5728", "0.824", "0.586", "0.1748", "0.586" ], [ "0.1748", "0.586", "0.2418", "0.586", "0.2418", "0.5977", "0.1748", "0.5977" ], [ "0.4558", "0.5986", "0.6055", "0.5986", "0.6055", "0.613", "0.4558", "0.613" ], [ "0.3938", "0.61", "0.423", "0.61", "0.423", "0.622", "0.3938", "0.622" ], [ "0.8013", "0.606", "0.8286", "0.606", "0.8286", "0.6226", "0.8013", "0.6226" ], [ "0.4927", "0.613", "0.576", "0.6157", "0.576", "0.63", "0.4927", "0.6274" ], [ "0.464", "0.6206", "0.5", "0.6206", "0.5", "0.6274", "0.464", "0.6274" ], [ "0.1748", "0.636", "0.824", "0.636", "0.824", "0.6494", "0.1748", "0.6494" ], [ "0.1748", "0.651", "0.826", "0.651", "0.826", "0.6655", "0.1748", "0.6655" ], [ "0.1732", "0.664", "0.3513", "0.664", "0.3513", "0.6787", "0.1732", "0.6787" ], [ "0.1748", "0.6934", "0.8228", "0.6934", "0.8228", "0.707", "0.1748", "0.707" ], [ "0.1732", "0.7085", "0.8228", "0.707", "0.8228", "0.7217", "0.1732", "0.7227" ], [ "0.1748", "0.7227", "0.826", "0.7227", "0.826", "0.7363", "0.1748", "0.7363" ], [ "0.1732", "0.735", "0.3096", "0.735", "0.3096", "0.7485", "0.1732", "0.7485" ], [ "0.1732", "0.765", "0.8267", "0.7646", "0.8267", "0.779", "0.1732", "0.78" ], [ "0.1732", "0.78", "0.824", "0.778", "0.824", "0.7925", "0.1732", "0.7944" ], [ "0.1732", "0.7935", "0.826", "0.7925", "0.826", "0.8066", "0.1732", "0.808" ], [ "0.1748", "0.808", "0.824", "0.808", "0.824", "0.8213", "0.1748", "0.8213" ], [ "0.1732", "0.8213", "0.3096", "0.8213", "0.3096", "0.8354", "0.1732", "0.8354" ], [ "0.4805", "0.856", "0.705", "0.854", "0.705", "0.872", "0.4805", "0.8735" ], [ "0.8013", "0.856", "0.8267", "0.856", "0.8267", "0.8735", "0.8013", "0.8735" ], [ "0.3782", "0.8623", "0.384", "0.8623", "0.384", "0.867", "0.3782", "0.867" ], [ "0.398", "0.869", "0.407", "0.869", "0.407", "0.8735", "0.398", "0.8735" ], [ "0.2754", "0.857", "0.3823", "0.857", "0.3823", "0.872", "0.2754", "0.872" ], [ "0.4478", "0.8755", "0.478", "0.8755", "0.478", "0.8833", "0.4478", "0.8833" ], [ "0.1929", "0.8945", "0.826", "0.8965", "0.826", "0.914", "0.1929", "0.912" ], [ "0.1732", "0.911", "0.2712", "0.912", "0.2712", "0.924", "0.1732", "0.9224" ], [ "0.4927", "0.9487", "0.5063", "0.9487", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2018", "be applied to arbitrary parametric models. To achieve structured exploration, we sample from a", "set of policies by applying additive Gaussian noise to the parameter vector of the current policy:", " = + N(0, 2I). Importantly, the perturbed policy is sampled at the beginning of each episode", "and keptfixedfor the entirerollout.For convenience and readability,we denote this perturbed policy", "as元：=andanalogouslydefineπ=Te", "State-dependent exploration", "As pointed out by RickstieB et al. (2008), there is a crucial difference", "between action space noise and parameter space noise. Consider the continuous action space", "case.WhenusingGaussianactionnoise,actions aresampledaccordingtosomestochasticpolicy", "generating at = (st) + N(O, α?I). Therefore, even for a fixed state s, we will almost certainly", "obtain a different action whenever that state is sampled again in the rollout, since action space noise", "iscompletelyindependentofthecurrentstateSt(noticethatthisisequallytrueforcorrelatedaction", "space noise). In contrast, if the parameters of the policy are perturbed at the beginning of each", "episode, we get at = π(st). In this case, the same action will be taken every time the same state st is", "sampledintherollout.Thisensuresconsistencyinactions,and directlyintroducesadependence", "between the state and the exploratory action taken.", "Perturbing deep neural networks It is not immediately obvious that deep neural networks, with", "potentially millions of parameters and complicated nonlinear interactions, can be perturbed in", "meaningful ways by applying spherical Gaussian noise. However, as recently shown by Salimans", "et al. (2017), a simple reparameterization of the network achieves exactly this. More concretely, we", "uselayernormalization(Baetal.,20l6)betweenperturbedlayers.3Duetothisnormalizingacross", "activations within a layer, the same perturbation scale can be used across all layers, even though", "different layers may exhibit different sensitivities to noise.", "Adaptive noise scalingParameter space noise requires us to pick a suitable scale o. This can be", "problematic since the scale will strongly depend on the specific network architecture, and is likely to", "vary over time as parameters become more sensitive to noise as learning progresses. Additionally,", "whileitis easyto intuitivelygraspthe scale of action spacenoise,it isfar harderto understand the", "scale in parameter space. We propose a simple solution that resolves all aforementioned limitations", "in an easy and straightforward way. This is achieved by adapting the scale of the parameter space", "noise overtimeand relating ittothe variancein action spacethatit induces.More concretely,wecan", "define a distance measure between perturbed and non-perturbed policy in action space and adaptively", "increase or decrease the parameter space noise depending on whether it is below or above a certain", "threshold:", "[aokifd(元,元)≤,", "Ok+", "(1)", "otherwise,", "-ok", "where α R>o is a scaling factor and e R>o a threshold value. The concrete realization of d(, -)", "depends on the algorithm at hand and we describe appropriate distance measures for DQN, DDPG,", "and TRPO in Appendix C.", "Parameter space noisefor off-policy methodsIn the off-policy case,parameter space noise can", "be applied straightforwardly since, by definition, data that was collected off-policy can be used. More", "concretely, we only perturb the policy for exploration and train the non-perturbed network on this", "data by replaying it.", "Parameter space noise for on-policy methods Parameter noise can be incorporated in an on-", "policy setting, using an adapted policy gradient, as set forth by RuckstieB et al. (2008). Policy", "gradient methods optimize E~(元,p)[R()]. Given a stochastic policy e(a|s) with ~N(,),", "the expected return can be expanded using likelihood ratios and the re-parametrization trick (Kingma", "& Welling, 2013) as", "Vo, log (atlst;Φ+e)Rt(T", "(2)", "5", "", "VS,E.[R(T)] :", "Ft=0", "3This is in contrast to Salimans etal. (2017), who use virtual batch normalization, which we found to perform", "less consistently", "3" ]

{ "arxiv_id": "1706.01905", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1748", "0.04736" ], [ "0.1708", "0.101", "0.826", "0.1016", "0.826", "0.1237", "0.1708", "0.1225" ], [ "0.1732", "0.125", "0.8228", "0.125", "0.8228", "0.1395", "0.1732", "0.1395" ], [ "0.1748", "0.1395", "0.4453", "0.1395", "0.4453", "0.1528", "0.1748", "0.1528" ], [ "0.1765", "0.178", "0.1912", "0.178", "0.1912", "0.1875", "0.1765", "0.1875" ], [ "0.2051", "0.1755", "0.3284", "0.1755", "0.3284", "0.1895", "0.2051", "0.1895" ], [ "0.1748", "0.207", "0.4714", "0.207", "0.4714", "0.2203", "0.1748", "0.2203" ], [ "0.2115", "0.2343", "0.824", "0.2343", "0.824", "0.2474", "0.2115", "0.2474" ], [ "0.2059", "0.254", "0.824", "0.254", "0.824", "0.2683", "0.2059", "0.2683" ], [ "0.2009", "0.2727", "0.822", "0.2734", "0.822", "0.2886", "0.2009", "0.2874" ], [ "0.2336", "0.2886", "0.664", "0.2886", "0.664", "0.3018", "0.2336", "0.3018" ], [ "0.1732", "0.315", "0.824", "0.315", "0.824", "0.3289", "0.1732", "0.3289" ], [ "0.1748", "0.33", "0.2263", "0.33", "0.2263", "0.3416", "0.1748", "0.3416" ], [ "0.1748", "0.361", "0.671", "0.361", "0.671", "0.375", "0.1748", "0.375" ], [ "0.1748", "0.387", "0.826", "0.387", "0.826", "0.4016", "0.1748", "0.4016" ], [ "0.1748", "0.4016", "0.826", "0.4016", "0.826", "0.4153", "0.1748", "0.4153" ], [ "0.1724", "0.4153", "0.8267", "0.414", "0.8267", "0.4287", "0.1724", "0.43" ], [ "0.1732", "0.4458", "0.8228", "0.4463", "0.8228", "0.461", "0.1732", "0.4602" ], [ "0.1765", "0.4622", "0.824", "0.4622", "0.824", "0.473", "0.1765", "0.473" ], [ "0.1748", "0.4749", "0.824", "0.4749", "0.824", "0.488", "0.1748", "0.488" ], [ "0.1732", "0.488", "0.824", "0.4873", "0.824", "0.502", "0.1732", "0.5024" ], [ "0.1732", "0.502", "0.8228", "0.502", "0.8228", "0.515", "0.1732", "0.515" ], [ "0.1748", "0.5166", "0.8228", "0.5166", "0.8228", "0.53", "0.1748", "0.53" ], [ "0.1748", "0.53", "0.824", "0.53", "0.824", "0.5435", "0.1748", "0.5435" ], [ "0.1748", "0.5435", "0.824", "0.5435", "0.824", "0.5566", "0.1748", "0.5566" ], [ "0.1732", "0.558", "0.699", "0.558", "0.699", "0.5713", "0.1732", "0.5713" ], [ "0.1748", "0.5776", "0.824", "0.5776", "0.824", "0.5923", "0.1748", "0.5923" ], [ "0.1748", "0.5933", "0.826", "0.5933", "0.826", "0.607", "0.1748", "0.607" ], [ "0.1732", "0.607", "0.824", "0.607", "0.824", "0.6206", "0.1732", "0.6206" ], [ "0.1765", "0.6226", "0.822", "0.6226", "0.822", "0.6333", "0.1765", "0.6333" ], [ "0.1765", "0.634", "0.8228", "0.634", "0.8228", "0.6484", "0.1765", "0.6484" ], [ "0.1748", "0.6484", "0.8228", "0.6484", "0.8228", "0.662", "0.1748", "0.662" ], [ "0.1748", "0.662", "0.8228", "0.662", "0.8228", "0.6753", "0.1748", "0.6753" ], [ "0.1732", "0.675", "0.822", "0.675", "0.822", "0.688", "0.1732", "0.688" ], [ "0.1748", "0.6895", "0.824", "0.6895", "0.824", "0.7026", "0.1748", "0.7026" ], [ "0.1765", "0.705", "0.822", "0.705", "0.822", "0.7153", "0.1765", "0.7153" ], [ "0.1748", "0.7173", "0.824", "0.7173", "0.824", "0.731", "0.1748", "0.731" ], [ "0.1724", "0.732", "0.2532", "0.73", "0.2532", "0.7437", "0.1724", "0.7456" ], [ "0.1732", "0.752", "0.8228", "0.752", "0.8228", "0.765", "0.1732", "0.765" ], [ "0.1748", "0.7666", "0.824", "0.7666", "0.824", "0.78", "0.1748", "0.78" ], [ "0.1748", "0.78", "0.824", "0.78", "0.824", "0.7944", "0.1748", "0.7944" ], [ "0.1732", "0.7944", "0.8203", "0.7944", "0.8203", "0.808", "0.1732", "0.808" ], [ "0.1748", "0.815", "0.8228", "0.815", "0.8228", "0.829", "0.1748", "0.829" ], [ "0.1748", "0.829", "0.824", "0.829", "0.824", "0.8423", "0.1748", "0.8423" ], [ "0.1748", "0.8433", "0.8228", "0.8433", "0.8228", "0.857", "0.1748", "0.857" ], [ "0.1748", "0.857", "0.824", "0.857", "0.824", "0.8706", "0.1748", "0.8706" ], [ "0.1748", "0.872", "0.8228", "0.872", "0.8228", "0.8853", "0.1748", "0.8853" ], [ "0.1748", "0.8853", "0.4944", "0.8853", "0.4944", "0.8984", "0.1748", "0.8984" ], [ "0.1953", "0.911", "0.5107", "0.911", "0.5107", "0.9253", "0.1953", "0.9253" ], [ "0.4927", "0.9497", "0.5063", "0.9497", "0.5063", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2018", "for N samples e' ~ N(O, I) and + ~ ($+e',P) (see Appendix B for a full derivation). Rather", "than updating according to the previously derived policy gradient, we fix its value to ?I and scale", "it adaptively as described in Appendix C.", "4", "EXPERIMENTS", "This section answers the following questions:", "(i) Do existing state-of-the-art RL algorithms benefit from incorporating parameter space noise?", "(ii) Does parameter space noise aid in exploring sparse reward environments more effectively?", "(ii)How does parameter spacenoise exploration compare against evolution strategies fordeep", "policies (Salimans et al., 2017) with respect to sample efficiency?", "Reference implementations of DQN and DDPG with adaptive parameter space noise are available", "online.4", "4.1COMPARINGPARAMETER SPACENOISETOACTIONSPACENOISE", "The added value of parameter space noise over action space noise is measured on both high-", "dimensional discrete-action environments and continuous control tasks. For the discrete environments,", "comparisons are made using DQN, while DDPG and TRPO are used on the continuous control tasks.", "Discrete-action environments For discrete-action environments, we use the Arcade Learning", "Environment(ALE,Bellemareetal.(2013))benchmarkalongwithastandardDQNimplementation", "We compare a baseline DQN agent with E-greedy action noise against a version of DQN with", "parameter noise. We linearly anneal e from 1.0 to 0.i over the first 1 million timesteps. For parameter", "noise, we adapt the scale using a simple heuristic that increases the scale if the KL divergence between", "perturbed and non-perturbed policy is less than the KL divergence between greedy and e-greedy", "policy and decreases it otherwise (see Section C.1 for details). By using this approach, we achieve a", "fair comparison between action space noise and parameter space noise since the magnitude of the", "noise is similar and also avoid the introduction of an additional hyperparameter.", "For parameter perturbation, we found it useful to reparametrize the network in terms of an explicit", "policy that represents the greedy policy π implied by the Q-values, rather than perturbing the Q-", "function directly. To represent the policy (as), we add a single fully connected layer after the", "convolutionalpartof thenetwork,followedby a softmax outputlayer.Thus,predictsa discrete", "probability distribution over actions, given a state. We find that perturbing instead of Q results", "in more meaningful changes since we now define an explicit behavioral policy. In this setting, the", "Q-network is trained according to standard DQN practices. The policy π is trained by maximizing", "the probability of outputting the greedy action accordingly to the current Q-network. Essentially, the", "policy is trained to exhibit the same behavior as running greedy DQN. To rule out this double-headed", "versionofDQNaloneexhibitssignificantlydifferentbehavior,wealwayscompareourparametei", "space noise approach against two baselines, regular DQN and two-headed DQN, both with e-greedy", "exploration", "We furthermore randomly sample actions for the first 50 thousand timesteps in all cases to fill the", "replay buffer before starting training. Moreover, we found that parameter space noise performs better", "if it is combined with a bit of action space noise (we use a e-greedy behavioral policy with e = O.01", "for the parameter space noise experiments). Full experimental details are described in Section A.1.", "We chose 21 games of varying complexity, according to the taxonomy presented by (Bellemare et al.", "2016). The learning curves are shown in Figure 1 for a selection of games (see Appendix D for full", "results). Each agent is trained for 40 M frames. The overall performance is estimated by running", "each configuration with three different random seeds, and we plot the median return (line) as well as", "the interquartile range (shaded area). Note that performance is evaluated on the exploratory policy", "since we are interested in its behavior especially.", "thttps://github.com/openai/baselines", "4" ]

{ "arxiv_id": "1706.01905", "language": "en", "page_idx": 3 }

[ [ "0.1748", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1748", "0.04736" ], [ "0.2133", "0.3542", "0.4631", "0.3542", "0.4631", "0.3643", "0.2133", "0.3643" ], [ "0.527", "0.3567", "0.5933", "0.3567", "0.5933", "0.3643", "0.527", "0.3643" ], [ "0.607", "0.356", "0.785", "0.356", "0.785", "0.3662", "0.607", "0.3662" ], [ "0.603", "0.3591", "0.607", "0.3567", "0.6094", "0.3623", "0.6055", "0.3643" ], [ "0.4631", "0.358", "0.527", "0.358", "0.527", "0.3623", "0.4631", "0.3623" ], [ "0.1732", "0.3958", "0.8228", "0.3965", "0.8228", "0.411", "0.1732", "0.4104" ], [ "0.1765", "0.411", "0.8228", "0.411", "0.8228", "0.4248", "0.1765", "0.4248" ], [ "0.1748", "0.4236", "0.824", "0.4255", "0.824", "0.4395", "0.1748", "0.4375" ], [ "0.1748", "0.4395", "0.826", "0.4395", "0.826", "0.4526", "0.1748", "0.4526" ], [ "0.1748", "0.4526", "0.824", "0.4526", "0.824", "0.4666", "0.1748", "0.4666" ], [ "0.1748", "0.4666", "0.5", "0.4666", "0.5", "0.4797", "0.1748", "0.4797" ], [ "0.1748", "0.4873", "0.8228", "0.4873", "0.8228", "0.5005", "0.1748", "0.5005" ], [ "0.1748", "0.5005", "0.826", "0.5005", "0.826", "0.515", "0.1748", "0.515" ], [ "0.1748", "0.515", "0.8228", "0.515", "0.8228", "0.529", "0.1748", "0.529" ], [ "0.1772", "0.53", "0.822", "0.53", "0.822", "0.5405", "0.1772", "0.5405" ], [ "0.1748", "0.5435", "0.818", "0.5435", "0.818", "0.5566", "0.1748", "0.5566" ], [ "0.1748", "0.563", "0.826", "0.563", "0.826", "0.5776", "0.1748", "0.5776" ], [ "0.1748", "0.5776", "0.8228", "0.5776", "0.8228", "0.5913", "0.1748", "0.5913" ], [ "0.1748", "0.5923", "0.824", "0.5923", "0.824", "0.606", "0.1748", "0.606" ], [ "0.1748", "0.606", "0.433", "0.606", "0.433", "0.619", "0.1748", "0.619" ], [ "0.1748", "0.638", "0.8228", "0.638", "0.8228", "0.6514", "0.1748", "0.6514" ], [ "0.1765", "0.654", "0.8228", "0.654", "0.8228", "0.664", "0.1765", "0.664" ], [ "0.1732", "0.6655", "0.822", "0.6675", "0.822", "0.681", "0.1732", "0.6787" ], [ "0.1732", "0.681", "0.824", "0.681", "0.824", "0.6943", "0.1732", "0.6943" ], [ "0.1748", "0.6943", "0.8228", "0.6943", "0.8228", "0.7085", "0.1748", "0.7085" ], [ "0.1748", "0.709", "0.415", "0.709", "0.415", "0.7227", "0.1748", "0.7227" ], [ "0.1748", "0.729", "0.824", "0.729", "0.824", "0.742", "0.1748", "0.742" ], [ "0.1748", "0.7437", "0.455", "0.7437", "0.455", "0.757", "0.1748", "0.757" ], [ "0.4617", "0.7437", "0.8228", "0.7437", "0.8228", "0.757", "0.4617", "0.757" ], [ "0.1765", "0.759", "0.822", "0.759", "0.822", "0.7695", "0.1765", "0.7695" ], [ "0.1748", "0.7715", "0.8228", "0.7715", "0.8228", "0.785", "0.1748", "0.785" ], [ "0.1748", "0.785", "0.826", "0.785", "0.826", "0.799", "0.1748", "0.799" ], [ "0.1748", "0.799", "0.3865", "0.799", "0.3865", "0.8125", "0.1748", "0.8125" ], [ "0.1748", "0.8193", "0.8228", "0.8193", "0.8228", "0.8335", "0.1748", "0.8335" ], [ "0.1748", "0.834", "0.824", "0.834", "0.824", "0.8477", "0.1748", "0.8477" ], [ "0.1748", "0.8477", "0.824", "0.8477", "0.824", "0.8613", "0.1748", "0.8613" ], [ "0.1748", "0.8623", "0.726", "0.8623", "0.726", "0.8755", "0.1748", "0.8755" ], [ "0.1732", "0.8833", "0.822", "0.8833", "0.822", "0.8965", "0.1732", "0.8965" ], [ "0.1732", "0.896", "0.824", "0.8965", "0.824", "0.911", "0.1732", "0.91" ], [ "0.1748", "0.911", "0.826", "0.911", "0.826", "0.925", "0.1748", "0.925" ], [ "0.4944", "0.9487", "0.5063", "0.9487", "0.5063", "0.9604", "0.4944", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2018", "Figurel:MedianDONreturnsforsev", "environme", "tplottedovertrainingsteps", "7", "eraAH", "Overall, our results show that parameter space noise often outperforms action space noise, especially", "on games that require consistency (e.g. Enduro, Freeway) and performs comparably on the remaining", "ones. Additionally,learning progress usually starts much sooner when using parameter space noise", "Finally, we also compare against a double-headed version of DQN with e-greedy exploration to", "ensure that this change in architecture is not responsible for improved exploration, which our results", "confirm. Full results are available in Appendix D.", "That being said, parameter space noise is unable to sufficiently explore in extremely challenging", "games like Montezuma's Revenge. More sophisticated exploration methods like Bellemare et al.", "(2016) are likely necessary to successfully learn these games: However, such methods often rely on", "someformof\"innerexploration method,whichis usuallytraditional action spacenoise.Itwouldbe", "interesting to evaluate the effect of parameter space noise when combined with exploration methods.", "On a final note, proposed improvements to DQN like double DQN (Hasselt, 2010), prioritized", "experience replay (Schaul et al., 2015), and dueling networks (Wang et al., 2015) are orthogonal to", "our improvements and would therefore likely improve results further. We leave the experimental", "validation of this theory to future work.", "Continuous control environmentsWe now compare parameter noise with action noise on the", "continuouscontrolenvironmentsimplementedinOpenAIGym(Brockmanetal.,2o16).Weuse", "DDPG(Lillicrap etal.,2015)as theRL algorithmforall environments with similarhyperparameters", "as outlined in the original paper except for the fact that layer normalization (Ba et al., 2016) is applied", "after each layer before the nonlinearity, which we found to be useful in either case and especially", "important for parameter space noise.", "We compare the performance of the following configurations: (a) no noise at all, (b)uncorrelated", "additive Gaussian action space noise (g", "=0.2),(c)correlated additiveGaussian action space", "noise(Ornstein-Uhlenbeckprocess(Uhlenbeck&Ornstein,1930)with=0.2),and(d)adaptive", "parameter space noise. In the case of parameter space noise, we adapt the scale so that the resulting", "change in action space is comparable toourbaselines with uncorrelated Gaussian action spacenoise", "(see Section C.2 for full details)", "We evaluate the performance on several continuous control tasks. Figure 2 depicts the results for", "three exemplary environments. Each agent is trained for 1 M timesteps, where 1 epoch consists of", "10 thousand timesteps. In order to make results comparable between configurations, we evaluate the", "performance of the agent every 10 thousand steps by using no noise for 20 episodes.", "On HalfCheetah, parameter space noise achieves significantly higher returns than all other configura", "tions. We find that, in this environment, all other exploration schemes quickly converge to a local", "optimum (in which the agent learns to flip on its back and then “wiggles\" its way forward). Parameter", "5" ]

{ "arxiv_id": "1706.01905", "language": "en", "page_idx": 4 }

[ [ "0.2284", "0.08075", "0.7715", "0.082", "0.7715", "0.0998", "0.2284", "0.0986" ], [ "0.351", "0.12354", "0.665", "0.12054", "0.665", "0.1383", "0.351", "0.1414" ], [ "0.1663", "0.1383", "0.8403", "0.1414", "0.8403", "0.1592", "0.1663", "0.1561" ], [ "0.1814", "0.155", "0.825", "0.158", "0.825", "0.1758", "0.1814", "0.1729" ], [ "0.1948", "0.1758", "0.812", "0.1758", "0.812", "0.1907", "0.1948", "0.1907" ], [ "0.1587", "0.1882", "0.8447", "0.1907", "0.8447", "0.2079", "0.1587", "0.206" ], [ "0.2435", "0.209", "0.764", "0.209", "0.764", "0.2227", "0.2435", "0.2227" ], [ "0.2644", "0.2632", "0.3435", "0.2654", "0.3435", "0.2808", "0.2644", "0.279" ], [ "0.525", "0.2715", "0.817", "0.2715", "0.817", "0.282", "0.525", "0.282" ], [ "0.523", "0.284", "0.686", "0.284", "0.686", "0.2976", "0.523", "0.2976" ], [ "0.147", "0.2903", "0.461", "0.2903", "0.461", "0.3047", "0.147", "0.3047" ], [ "0.5225", "0.2964", "0.686", "0.2964", "0.686", "0.31", "0.5225", "0.31" ], [ "0.02016", "0.3027", "0.06635", "0.3027", "0.06635", "0.4407", "0.02016", "0.4407" ], [ "0.1453", "0.3047", "0.4626", "0.3047", "0.4626", "0.3184", "0.1453", "0.3184" ], [ "0.5225", "0.3113", "0.889", "0.3113", "0.889", "0.3247", "0.5225", "0.3247" ], [ "0.147", "0.3196", "0.4626", "0.3196", "0.4626", "0.3333", "0.147", "0.3333" ], [ "0.525", "0.3267", "0.7583", "0.3267", "0.7583", "0.3374", "0.525", "0.3374" ], [ "0.147", "0.3342", "0.461", "0.3342", "0.461", "0.3474", "0.147", "0.3474" ], [ "0.5225", "0.3384", "0.8623", "0.3384", "0.8623", "0.3516", "0.5225", "0.3516" ], [ "0.1486", "0.3499", "0.4592", "0.3499", "0.4592", "0.3599", "0.1486", "0.3599" ], [ "0.525", "0.354", "0.671", "0.354", "0.671", "0.364", "0.525", "0.364" ], [ "0.1486", "0.3623", "0.4626", "0.3623", "0.4626", "0.376", "0.1486", "0.376" ], [ "0.523", "0.3665", "0.8213", "0.3665", "0.8213", "0.3801", "0.523", "0.3801" ], [ "0.147", "0.376", "0.4592", "0.376", "0.4592", "0.3896", "0.147", "0.3896" ], [ "0.523", "0.3801", "0.8037", "0.3801", "0.8037", "0.3938", "0.523", "0.3938" ], [ "0.147", "0.3901", "0.461", "0.3901", "0.461", "0.4038", "0.147", "0.4038" ], [ "0.5225", "0.3938", "0.791", "0.3938", "0.791", "0.4067", "0.5225", "0.4067" ], [ "0.147", "0.405", "0.4592", "0.405", "0.4592", "0.4187", "0.147", "0.4187" ], [ "0.147", "0.4187", "0.4626", "0.4187", "0.4626", "0.4316", "0.147", "0.4316" ], [ "0.513", "0.427", "0.8833", "0.4258", "0.8833", "0.4407", "0.513", "0.4412" ], [ "0.02184", "0.437", "0.06464", "0.437", "0.06464", "0.5186", "0.02184", "0.5186" ], [ "0.147", "0.4329", "0.4592", "0.4329", "0.4592", "0.4465", "0.147", "0.4465" ], [ "0.03864", "0.4465", "0.0487", "0.4465", "0.0487", "0.4507", "0.03864", "0.4507" ], [ "0.511", "0.4412", "0.852", "0.4412", "0.852", "0.4548", "0.511", "0.4548" ], [ "0.147", "0.4478", "0.461", "0.4478", "0.461", "0.4614", "0.147", "0.4614" ], [ "0.1453", "0.4602", "0.4626", "0.4614", "0.4626", "0.4756", "0.1453", "0.4744" ], [ "0.147", "0.4756", "0.461", "0.4756", "0.461", "0.4893", "0.147", "0.4893" ], [ "0.147", "0.4905", "0.461", "0.4905", "0.461", "0.5044", "0.147", "0.5044" ], [ "0.5146", "0.488", "0.8833", "0.488", "0.8833", "0.503", "0.5146", "0.503" ], [ "0.147", "0.5044", "0.4626", "0.5044", "0.4626", "0.517", "0.147", "0.517" ], [ "0.513", "0.506", "0.885", "0.506", "0.885", "0.5195", "0.513", "0.5195" ], [ "0.02184", "0.521", "0.06213", "0.521", "0.06213", "0.73", "0.02184", "0.73" ], [ "0.147", "0.5186", "0.461", "0.5186", "0.461", "0.532", "0.147", "0.532" ], [ "0.513", "0.5215", "0.8867", "0.5215", "0.8867", "0.535", "0.513", "0.535" ], [ "0.147", "0.533", "0.461", "0.533", "0.461", "0.5464", "0.147", "0.5464" ], [ "0.513", "0.536", "0.8867", "0.5376", "0.8867", "0.552", "0.513", "0.5503" ], [ "0.147", "0.5464", "0.4626", "0.5464", "0.4626", "0.56", "0.147", "0.56" ], [ "0.513", "0.554", "0.8867", "0.554", "0.8867", "0.568", "0.513", "0.568" ], [ "0.147", "0.5625", "0.4592", "0.5625", "0.4592", "0.5723", "0.147", "0.5723" ], [ "0.147", "0.5747", "0.4626", "0.5747", "0.4626", "0.5884", "0.147", "0.5884" ], [ "0.5146", "0.5693", "0.885", "0.5693", "0.885", "0.583", "0.5146", "0.583" ], [ "0.1453", "0.589", "0.461", "0.589", "0.461", "0.6025", "0.1453", "0.6025" ], [ "0.5146", "0.5874", "0.8833", "0.5874", "0.8833", "0.597", "0.5146", "0.597" ], [ "0.147", "0.605", "0.294", "0.605", "0.294", "0.615", "0.147", "0.615" ], [ "0.5146", "0.6016", "0.8833", "0.6016", "0.8833", "0.615", "0.5146", "0.615" ], [ "0.5146", "0.618", "0.882", "0.618", "0.882", "0.632", "0.5146", "0.632" ], [ "0.12427", "0.632", "0.1301", "0.632", "0.1301", "0.637", "0.12427", "0.637" ], [ "0.126", "0.6274", "0.2603", "0.6274", "0.2603", "0.641", "0.126", "0.641" ], [ "0.5146", "0.6343", "0.8833", "0.6343", "0.8833", "0.648", "0.5146", "0.648" ], [ "0.12006", "0.654", "0.4937", "0.654", "0.4937", "0.6675", "0.12006", "0.6675" ], [ "0.511", "0.6494", "0.885", "0.6484", "0.885", "0.663", "0.511", "0.6646" ], [ "0.11755", "0.6704", "0.4902", "0.6694", "0.4902", "0.684", "0.11755", "0.685" ], [ "0.5146", "0.6665", "0.882", "0.6665", "0.882", "0.6797", "0.5146", "0.6797" ], [ "0.11755", "0.686", "0.4902", "0.687", "0.4902", "0.7007", "0.11755", "0.6997" ], [ "0.511", "0.6816", "0.8833", "0.681", "0.8833", "0.6953", "0.511", "0.697" ], [ "0.1184", "0.703", "0.4902", "0.703", "0.4902", "0.7163", "0.1184", "0.7163" ], [ "0.511", "0.6978", "0.8833", "0.6987", "0.8833", "0.7134", "0.511", "0.712" ], [ "0.12006", "0.7188", "0.4895", "0.7188", "0.4895", "0.7324", "0.12006", "0.7324" ], [ "0.513", "0.7153", "0.8833", "0.7153", "0.8833", "0.7285", "0.513", "0.7285" ], [ "0.1184", "0.733", "0.4895", "0.734", "0.4895", "0.749", "0.1184", "0.7476" ], [ "0.513", "0.731", "0.885", "0.731", "0.885", "0.7446", "0.513", "0.7446" ], [ "0.12006", "0.7505", "0.4878", "0.7505", "0.4878", "0.764", "0.12006", "0.764" ], [ "0.513", "0.749", "0.581", "0.749", "0.581", "0.7603", "0.513", "0.7603" ], [ "0.11755", "0.7666", "0.4895", "0.7656", "0.4895", "0.78", "0.11755", "0.781" ], [ "0.532", "0.7666", "0.885", "0.7666", "0.885", "0.78", "0.532", "0.78" ], [ "0.12006", "0.783", "0.4902", "0.783", "0.4902", "0.7964", "0.12006", "0.7964" ], [ "0.513", "0.783", "0.885", "0.783", "0.885", "0.7964", "0.513", "0.7964" ], [ "0.1184", "0.799", "0.4902", "0.7974", "0.4902", "0.8125", "0.1184", "0.8135" ], [ "0.513", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.8154", "0.4937", "0.8154", "0.4937", "0.829", "0.1184", "0.829" ], [ "0.513", "0.814", "0.8833", "0.814", "0.8833", "0.8276", "0.513", "0.8276" ], [ "0.1427", "0.835", "0.4902", "0.836", "0.4902", "0.8496", "0.1427", "0.8486" ], [ "0.511", "0.83", "0.8833", "0.8306", "0.8833", "0.8457", "0.511", "0.844" ], [ "0.11755", "0.8486", "0.4902", "0.847", "0.4902", "0.861", "0.11755", "0.8623" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.11755", "0.86", "0.4902", "0.86", "0.4902", "0.8735", "0.11755", "0.8735" ], [ "0.513", "0.8633", "0.8867", "0.8633", "0.8867", "0.877", "0.513", "0.877" ], [ "0.1184", "0.8716", "0.4895", "0.8716", "0.4895", "0.8853", "0.1184", "0.8853" ], [ "0.11755", "0.883", "0.4895", "0.8843", "0.4895", "0.8984", "0.11755", "0.898" ], [ "0.513", "0.88", "0.885", "0.88", "0.885", "0.8936", "0.513", "0.8936" ], [ "0.11755", "0.8965", "0.451", "0.8965", "0.451", "0.91", "0.11755", "0.91" ], [ "0.513", "0.8955", "0.885", "0.8955", "0.885", "0.909", "0.513", "0.909" ] ]

[ "Structural Pre-training for Dialogue Comprehension", "Zhuosheng Zhangl,2,3, Hai Zhaol,2,3,*", "'Department of Computer Science and Engineering, Shanghai Jiao Tong University", "2Key Laboratory of Shanghai Education Commission for Intelligent Interaction", "and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, China", "3MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University", "zhangzs@sjtu.edu.cn, zhaohai@cs.sjtu.edu.cn", "Abstract", "U,:how.can iconnect to URLstreaming server?", "Uz: i think mplayer does it", "Pre-trained language models (PrLMs) have", "Us: fusekino,could be?", " ", "demonstrated superior performance due to", "U:kino is editing software.idon't think it supports media", "their strong ability to learn universal lan-", "Us:sorryEMOJI,imeanmovieplayer", "guage representations from self-supervised", "U:not sure about it.if it can't try vlc and mplayer too", "pre-training.However,even withthehelp of", "U,: is mplayer aptable?", "the powerful PrLMs, it is still challenging to ef-", "Ugyes but i'm not sure if it is in themainrepo.", "fectively capture task-related knowledge from", "U,: have you ever updated your sources.list?", "dialogue texts which are enriched by corre-", "Uo: i have noidea.j use adept on kubuntu..", "lations among speaker-aware utterances.In", "this work, we present SPIDER, Structural Pre-", "Figure l: A multi-turn dialogue example. Different col", "[T']", "tralned DialoguE Reader, to capture dialogue", "", "ors indicate the utterances from different speakers", "exclusive features. To simulate the dialogue", "like features, we propose two training objec-", "tives in addition to the original LM objectives:", "1) utterance order restoration, which predicts", "lenging to effectively adapt task-related knowledge", "the order of the permuted utterances in dia-", "during the detailed task-specific training which is", "1A95601'S1Z:AIX", "logue context; 2) sentence backbone regular-", "usually in a way of fine-tuning (Gururangan et al.,", "ization, which regularizes the model to im-", "2020). Generally, those PrLMs handle the whole", "prove the factual correctness of summarized", "input text as a linear sequence of successive tokens", "subject-verb-objecttriplets.Experimental re-", "sults on widely used dialogue benchmarks ver-", "and implicitly capture the contextualized represen-", "ify the effectiveness of the newly introduced", "tations ofthosetokensthrough self-attention. Such", "self-supervised tasks.", "fine-tuning paradigm of exploiting PrLMs would", "be suboptimal to model dialogue task which holds", "", "Introduction", "exclusive text features that plain text for PrLM", "Recent advances in large-scale pre-training lan-", "training may hardly embody. Therefore, we ex-", "guage models (PrLMs) have achieved remarkable", "plore a fundamental way to alleviate this difficulty", "successes in a variety of natural language process-", "by improving the training of PrLM. This work de-", "ing (NLP) tasks (Peters et al., 2018; Radford et al.,", "votes itself to designing the natural way of adapting", "2018; Devlin et al., 2019a; Yang et al., 2019; Clark", "the language modeling to the dialogue scenario mo-", "et al., 2020). Providing fine-grained contextualized", "tivated by the natural characteristics of dialogue", "embedding, these pre-trained models are widely", "contexts.", "employed as encoders for various downstream NLP", "As an active research topic in the NLP field,", "tasks. Although the PrLMs demonstrate superior", "multi-turn dialogue modeling has attracted great in-", "performance due to their strong representation abil-", "terest. The typical task is response selection (Lowe", "ity from self-supervised pre-training, it is still chal-", "et al., 2015; Wu et al., 2017; Zhang et al., 2018) that", "*Corresponding author.This paper was partially sup-", "aims to select the appropriate response according", "ported by National Key Research and Development Program", "to a given dialogue context containing a number of", "of China (No. 2017YFB0304100), Key Projects of Na-", "utterances, which is the focus in this work. How-", "tional Natural Science Foundation of China (Ui836222 and", "61733011), Huawei-SJTU long term AI project, Cutting-edge", "ever, selecting a coherent and informative response", "Machine reading comprehension and language model.", "for a given dialogue context remains a challenge." ]

{ "arxiv_id": "2105.10956", "language": "en", "page_idx": 0 }

[ [ "0.1184", "0.076", "0.4878", "0.076", "0.4878", "0.08905", "0.1184", "0.08905" ], [ "0.513", "0.0766", "0.8833", "0.0766", "0.8833", "0.0903", "0.513", "0.0903" ], [ "0.1184", "0.09143", "0.4902", "0.0932", "0.4902", "0.1069", "0.1184", "0.1051" ], [ "0.513", "0.0932", "0.8833", "0.0932", "0.8833", "0.1069", "0.513", "0.1069" ], [ "0.11755", "0.1081", "0.4937", "0.1069", "0.4937", "0.1217", "0.11755", "0.1223" ], [ "0.513", "0.1081", "0.882", "0.1081", "0.882", "0.1217", "0.513", "0.1217" ], [ "0.12006", "0.12354", "0.492", "0.12354", "0.492", "0.1383", "0.12006", "0.1383" ], [ "0.513", "0.1247", "0.8867", "0.1247", "0.8867", "0.1383", "0.513", "0.1383" ], [ "0.12177", "0.1426", "0.492", "0.1426", "0.492", "0.1526", "0.12177", "0.1526" ], [ "0.513", "0.1414", "0.885", "0.1414", "0.885", "0.155", "0.513", "0.155" ], [ "0.1184", "0.1567", "0.4937", "0.1567", "0.4937", "0.1704", "0.1184", "0.1704" ], [ "0.513", "0.1567", "0.8833", "0.1567", "0.8833", "0.1704", "0.513", "0.1704" ], [ "0.1184", "0.1729", "0.492", "0.1729", "0.492", "0.1859", "0.1184", "0.1859" ], [ "0.513", "0.1716", "0.8833", "0.1729", "0.8833", "0.187", "0.513", "0.1859" ], [ "0.12006", "0.1907", "0.4895", "0.1907", "0.4895", "0.2007", "0.12006", "0.2007" ], [ "0.513", "0.1895", "0.885", "0.1895", "0.885", "0.2031", "0.513", "0.2031" ], [ "0.1184", "0.2048", "0.4895", "0.2048", "0.4895", "0.2185", "0.1184", "0.2185" ], [ "0.513", "0.2048", "0.8867", "0.2048", "0.8867", "0.2185", "0.513", "0.2185" ], [ "0.1184", "0.2214", "0.4937", "0.2214", "0.4937", "0.2351", "0.1184", "0.2351" ], [ "0.511", "0.2214", "0.8833", "0.2214", "0.8833", "0.2351", "0.511", "0.2351" ], [ "0.1184", "0.2382", "0.492", "0.2382", "0.492", "0.2517", "0.1184", "0.2517" ], [ "0.513", "0.2382", "0.8867", "0.2382", "0.8867", "0.2517", "0.513", "0.2517" ], [ "0.1184", "0.2542", "0.4902", "0.2542", "0.4902", "0.2673", "0.1184", "0.2673" ], [ "0.5146", "0.2546", "0.602", "0.2546", "0.602", "0.2666", "0.5146", "0.2666" ], [ "0.12006", "0.2683", "0.4895", "0.2683", "0.4895", "0.282", "0.12006", "0.282" ], [ "0.1184", "0.2852", "0.4895", "0.284", "0.4895", "0.2986", "0.1184", "0.2998" ], [ "0.5156", "0.2861", "0.534", "0.2861", "0.534", "0.2944", "0.5156", "0.2944" ], [ "0.534", "0.284", "0.6323", "0.284", "0.6323", "0.2986", "0.534", "0.2986" ], [ "0.11755", "0.3005", "0.4878", "0.3005", "0.4878", "0.3142", "0.11755", "0.3142" ], [ "0.5146", "0.3113", "0.8833", "0.3113", "0.8833", "0.326", "0.5146", "0.326" ], [ "0.11755", "0.3176", "0.4895", "0.3164", "0.4895", "0.3308", "0.11755", "0.332" ], [ "0.513", "0.3267", "0.8867", "0.3267", "0.8867", "0.3403", "0.513", "0.3403" ], [ "0.12006", "0.335", "0.4878", "0.335", "0.4878", "0.3457", "0.12006", "0.3457" ], [ "0.5146", "0.3433", "0.8833", "0.3433", "0.8833", "0.357", "0.5146", "0.357" ], [ "0.12006", "0.3499", "0.4895", "0.3499", "0.4895", "0.3635", "0.12006", "0.3635" ], [ "0.5146", "0.3594", "0.8833", "0.3594", "0.8833", "0.3723", "0.5146", "0.3723" ], [ "0.1184", "0.3652", "0.492", "0.3652", "0.492", "0.379", "0.1184", "0.379" ], [ "0.5146", "0.376", "0.885", "0.376", "0.885", "0.3896", "0.5146", "0.3896" ], [ "0.1184", "0.3818", "0.492", "0.3818", "0.492", "0.3955", "0.1184", "0.3955" ], [ "0.5146", "0.3914", "0.8833", "0.3914", "0.8833", "0.405", "0.5146", "0.405" ], [ "0.1184", "0.3984", "0.4902", "0.3967", "0.4902", "0.411", "0.1184", "0.4133" ], [ "0.513", "0.4067", "0.8833", "0.4062", "0.8833", "0.4204", "0.513", "0.4216" ], [ "0.12006", "0.415", "0.4902", "0.415", "0.4902", "0.4287", "0.12006", "0.4287" ], [ "0.511", "0.4229", "0.8867", "0.4233", "0.8867", "0.4382", "0.511", "0.437" ], [ "0.1184", "0.4312", "0.4902", "0.4312", "0.4902", "0.4448", "0.1184", "0.4448" ], [ "0.511", "0.4395", "0.882", "0.4395", "0.882", "0.4536", "0.511", "0.4536" ], [ "0.11755", "0.4453", "0.444", "0.4465", "0.444", "0.4614", "0.11755", "0.4602" ], [ "0.513", "0.456", "0.8867", "0.456", "0.8867", "0.4697", "0.513", "0.4697" ], [ "0.1351", "0.4631", "0.4902", "0.4644", "0.4902", "0.4785", "0.1351", "0.478" ], [ "0.513", "0.4714", "0.8867", "0.4714", "0.8867", "0.4863", "0.513", "0.4863" ], [ "0.12006", "0.4827", "0.4895", "0.4827", "0.4895", "0.4934", "0.12006", "0.4934" ], [ "0.511", "0.488", "0.885", "0.488", "0.885", "0.502", "0.511", "0.502" ], [ "0.1184", "0.4954", "0.4902", "0.4963", "0.4902", "0.511", "0.1184", "0.5103" ], [ "0.513", "0.5044", "0.8867", "0.5044", "0.8867", "0.5186", "0.513", "0.5186" ], [ "0.12006", "0.513", "0.4902", "0.513", "0.4902", "0.527", "0.12006", "0.527" ], [ "0.513", "0.521", "0.711", "0.521", "0.711", "0.5337", "0.513", "0.5337" ], [ "0.1184", "0.5293", "0.4895", "0.5293", "0.4895", "0.542", "0.1184", "0.542" ], [ "0.1184", "0.5444", "0.4895", "0.5435", "0.4895", "0.558", "0.1184", "0.5586" ], [ "0.513", "0.549", "0.7305", "0.549", "0.7305", "0.5625", "0.513", "0.5625" ], [ "0.12006", "0.56", "0.4878", "0.56", "0.4878", "0.5737", "0.12006", "0.5737" ], [ "0.5146", "0.571", "0.8833", "0.571", "0.8833", "0.585", "0.5146", "0.585" ], [ "0.1184", "0.5767", "0.4895", "0.5767", "0.4895", "0.5903", "0.1184", "0.5903" ], [ "0.511", "0.586", "0.8833", "0.586", "0.8833", "0.5996", "0.511", "0.5996" ], [ "0.1184", "0.593", "0.4878", "0.593", "0.4878", "0.6055", "0.1184", "0.6055" ], [ "0.513", "0.604", "0.885", "0.604", "0.885", "0.6177", "0.513", "0.6177" ], [ "0.1184", "0.61", "0.4902", "0.61", "0.4902", "0.6235", "0.1184", "0.6235" ], [ "0.11755", "0.626", "0.1704", "0.626", "0.1704", "0.64", "0.11755", "0.64" ], [ "0.511", "0.6206", "0.8833", "0.619", "0.8833", "0.633", "0.511", "0.6343" ], [ "0.511", "0.6353", "0.882", "0.6353", "0.882", "0.6484", "0.511", "0.6484" ], [ "0.513", "0.652", "0.8833", "0.652", "0.8833", "0.665", "0.513", "0.665" ], [ "0.12006", "0.66", "0.451", "0.66", "0.451", "0.6733", "0.12006", "0.6733" ], [ "0.513", "0.6685", "0.8833", "0.6685", "0.8833", "0.6816", "0.513", "0.6816" ], [ "0.1184", "0.684", "0.4902", "0.685", "0.4902", "0.6997", "0.1184", "0.6987" ], [ "0.511", "0.683", "0.876", "0.683", "0.876", "0.697", "0.511", "0.697" ], [ "0.12006", "0.7017", "0.4895", "0.7017", "0.4895", "0.7153", "0.12006", "0.7153" ], [ "0.5146", "0.7007", "0.895", "0.7007", "0.895", "0.7144", "0.5146", "0.7144" ], [ "0.12006", "0.7173", "0.4902", "0.7173", "0.4902", "0.731", "0.12006", "0.731" ], [ "0.51", "0.7153", "0.8833", "0.7153", "0.8833", "0.7285", "0.51", "0.7285" ], [ "0.12006", "0.734", "0.4895", "0.734", "0.4895", "0.7476", "0.12006", "0.7476" ], [ "0.5156", "0.731", "0.882", "0.731", "0.882", "0.7446", "0.5156", "0.7446" ], [ "0.12006", "0.75", "0.4895", "0.75", "0.4895", "0.763", "0.12006", "0.763" ], [ "0.513", "0.749", "0.8833", "0.749", "0.8833", "0.7627", "0.513", "0.7627" ], [ "0.12006", "0.7666", "0.4902", "0.7666", "0.4902", "0.78", "0.12006", "0.78" ], [ "0.5146", "0.764", "0.8833", "0.764", "0.8833", "0.7793", "0.5146", "0.7793" ], [ "0.1184", "0.781", "0.4902", "0.781", "0.4902", "0.7944", "0.1184", "0.7944" ], [ "0.51", "0.78", "0.8833", "0.78", "0.8833", "0.7935", "0.51", "0.7935" ], [ "0.12006", "0.7974", "0.4902", "0.7974", "0.4902", "0.811", "0.12006", "0.811" ], [ "0.511", "0.7974", "0.6465", "0.796", "0.6465", "0.81", "0.511", "0.8125" ], [ "0.12006", "0.814", "0.4902", "0.814", "0.4902", "0.8276", "0.12006", "0.8276" ], [ "0.529", "0.8125", "0.888", "0.8125", "0.888", "0.83", "0.529", "0.83" ], [ "0.1184", "0.8306", "0.4902", "0.8306", "0.4902", "0.844", "0.1184", "0.844" ], [ "0.513", "0.83", "0.888", "0.83", "0.888", "0.8433", "0.513", "0.8433" ], [ "0.1184", "0.8457", "0.492", "0.8467", "0.492", "0.861", "0.1184", "0.86" ], [ "0.513", "0.8457", "0.8867", "0.8457", "0.8867", "0.86", "0.513", "0.86" ], [ "0.1184", "0.8633", "0.4937", "0.8633", "0.4937", "0.877", "0.1184", "0.877" ], [ "0.5146", "0.8623", "0.885", "0.8623", "0.885", "0.876", "0.5146", "0.876" ], [ "0.12006", "0.879", "0.4937", "0.879", "0.4937", "0.8936", "0.12006", "0.8936" ], [ "0.5146", "0.879", "0.8486", "0.879", "0.8486", "0.8926", "0.5146", "0.8926" ], [ "0.12006", "0.8955", "0.492", "0.8955", "0.492", "0.909", "0.12006", "0.909" ], [ "0.532", "0.8945", "0.885", "0.8945", "0.885", "0.909", "0.532", "0.909" ] ]

[ "guide the language modeling Transformers with", "quence of successive tokens and implicitly capture", "explicit knowledge and complementing objectives", "the contextualized representations of those tokens", "(Zhang et al., 2019; Sun et al., 2019b; Xu et al.,", "through self-attention (Qu et al., 2019; Liu et al.", "2020). A most related work is Xu et al. (2020),", "2020). Such a way of modeling would be subopti-", "whichintroducesfour self-supervisedtasksinclud-", "mal to capture the high-level relationships between", "ing next session prediction, utterance restoration,", "utterances in the dialogue history. In this work, we", "incoherence detection and consistency discrimina-", "are motivated to model the structural relationships", "tion. Our work differs from Xu et al.(2020)by", "between utterances from utterance order restoration", "three sides. 1) Motivation: our method is designed", "and the factual correctness inside each utterance in", "for a general-purpose in broad dialogue compre-", "the perspective of language modeling pre-training", "hension tasks whose goals may be either utterance-", "instead of heuristically stacking deeper model ar-", "level discourse coherence or inner-utterance factual", "chitectures.", "correctness, instead of only motivated for down-", "stream context-response matching, whose goal is", "3", "Approach", "to measure if two sequences are related or not. 2)", "This section presents our proposed method SPI-", "Technique: we propose both sides of intra- and", "DER (Structural Pre-tralned DialoguE Reader).", "inter- utterance objectives.In contrast, the four", "First, we will present the standard dialogue com-", "objectives proposed in Xu et al. (2020) are natural", "prehension model as the backbone. Then, we will", "variants of NSP in BERT, which are all utterance-", "introduce our designed language modeling objec-", "level. 3) Training: we empirically evaluate domain-", "tives for dialogue scenarios, including utterance", "adaptive training and multi-task learning, instead of", "order restoration (UOR) and sentence backbone", "only employing multi-task learning, which requires", "regularization (SBR). In terms of model training,", "many efforts of optimizing coefficients in the loss", "we employ two strategies, i.e., 1) domain adap.", "functions, which would be time-consuming.", "tive post-training that first trains a language model", "In terms of factual backbone modeling, com-", "based on newly proposed objectives and then fine-", "pared withthe existing studies that enhance the", "tunes the response selection task; 2) multi-task fine-", "PrLMs by annotating named entities or incorporat-", "tuning that trains the model for downstream tasks,", "ing external knowledge graphs (Eric et al., 2017;", "along with LM objectives.", "Liu et al., 2018), the SVO triplets extracted in", "our sentencebackbone predication objective(SBP)", "3.1Transformer Encoder", "method, appear more widely in the text itself. Such", "We first employ a pre-trained language model", "triplets ensure the correctness of SvO and enable", " such as BERT (Devlin et al., 2019a) to obtain the", "our model to discover the salient facts from the", "initial word representations. The utterances and", "lengthy texts, sensing the intuition of \"who did", "what\".", "response are concatenated and then fed into the", "encoder. Given the context C and response R, we", "concatenate all utterances in the context and the", "2.2Multi-turn Dialogue Comprehension", "response candidate as a single consecutive token se-", "Multi-turn dialogue comprehension aims to teach", "quence with special tokens separating them: X ", "machines to read dialogue contexts and solve tasks", "[[CLS]R[SEP]Ui[EOU]... [EOU]Un[SEP]]", "such as response selection (Lowe et al., 2015; Wu", "where [CLs] and [SEP] are special tokens.", "et al., 2017; Zhang et al., 2018) and answering ques-", "[EoU] is the \"End Of Utterance\" tag designed", "tions (Sun et al., 2019a; Cui et al., 2020), whose", "for multiturn context. X is then fed into the", "common application is building intelligent human-", "BERT encoder, which is a deep multi-layer bidi-", "computer interactive systems (Chen et al., 2017a;", " rectional Transformer, to obtain a contextualized", "Shum et al., 2018; Li et al., 2017; Zhu et al., 2018b).", "representation H.", "Early studies mainly focus on the matching be-", "In detail, let X = {αi, ..., &n] be the embed-", "tween the dialogue context and question (Huang", "ding of the sequence, which are features of encod-", "et al., 2018; Zhu et al., 2018a). Recently, inspired", "ing sentence words of length n. The input em-", "by the impressive performance of PrLMs, the main-", "beddings are then fed into the multi-head attention", "stream is employing PrLMs to handle the whole", "layer to obtain the contextual representations.", "input texts of context and question, as a linear se-", "The embedding sequence X is processed to a" ]

{ "arxiv_id": "2105.10956", "language": "en", "page_idx": 2 }

[ [ "0.1335", "0.08435", "0.3838", "0.08435", "0.3838", "0.0974", "0.1335", "0.0974" ], [ "0.4534", "0.08435", "0.7", "0.08435", "0.7", "0.0944", "0.4534", "0.0944" ], [ "0.754", "0.0861", "0.827", "0.0861", "0.827", "0.0968", "0.754", "0.0968" ], [ "0.843", "0.0903", "0.8623", "0.0903", "0.8623", "0.0932", "0.843", "0.0932" ], [ "0.4517", "0.1069", "0.596", "0.1069", "0.596", "0.12054", "0.4517", "0.12054" ], [ "0.7524", "0.10156", "0.806", "0.10156", "0.806", "0.1134", "0.7524", "0.1134" ], [ "0.1351", "0.10925", "0.2795", "0.10925", "0.2795", "0.1223", "0.1351", "0.1223" ], [ "0.7524", "0.11816", "0.8193", "0.11816", "0.8193", "0.13", "0.7524", "0.13" ], [ "0.1335", "0.1318", "0.4014", "0.1318", "0.4014", "0.1455", "0.1335", "0.1455" ], [ "0.444", "0.13", "0.714", "0.13", "0.714", "0.1437", "0.444", "0.1437" ], [ "0.754", "0.136", "0.8345", "0.136", "0.8345", "0.1478", "0.754", "0.1478" ], [ "0.1318", "0.155", "0.398", "0.1567", "0.398", "0.1704", "0.1318", "0.1687" ], [ "0.4534", "0.155", "0.6514", "0.155", "0.6514", "0.1687", "0.4534", "0.1687" ], [ "0.756", "0.1509", "0.839", "0.1509", "0.839", "0.1621", "0.756", "0.1621" ], [ "0.7573", "0.1692", "0.837", "0.1692", "0.837", "0.1799", "0.7573", "0.1799" ], [ "0.1335", "0.1799", "0.3342", "0.1799", "0.3342", "0.1936", "0.1335", "0.1936" ], [ "0.455", "0.1787", "0.623", "0.1787", "0.623", "0.1895", "0.455", "0.1895" ], [ "0.621", "0.1812", "0.711", "0.1812", "0.711", "0.1882", "0.621", "0.1882" ], [ "0.754", "0.1859", "0.848", "0.1859", "0.848", "0.1965", "0.754", "0.1965" ], [ "0.1318", "0.2019", "0.392", "0.2031", "0.392", "0.2162", "0.1318", "0.2156" ], [ "0.4492", "0.2031", "0.7197", "0.2031", "0.7197", "0.2131", "0.4492", "0.2131" ], [ "0.754", "0.2019", "0.831", "0.2019", "0.831", "0.2131", "0.754", "0.2131" ], [ "0.215", "0.2227", "0.3123", "0.2227", "0.3123", "0.234", "0.215", "0.234" ], [ "0.55", "0.2245", "0.636", "0.2245", "0.636", "0.2322", "0.55", "0.2322" ], [ "0.781", "0.2227", "0.854", "0.2227", "0.854", "0.2328", "0.781", "0.2328" ], [ "0.3167", "0.2505", "0.686", "0.2505", "0.686", "0.2642", "0.3167", "0.2642" ], [ "0.1184", "0.2915", "0.4895", "0.2915", "0.4895", "0.3047", "0.1184", "0.3047" ], [ "0.513", "0.2915", "0.8833", "0.2915", "0.8833", "0.3047", "0.513", "0.3047" ], [ "0.12177", "0.3088", "0.492", "0.3088", "0.492", "0.3196", "0.12177", "0.3196" ], [ "0.511", "0.3071", "0.885", "0.308", "0.885", "0.3218", "0.511", "0.3206" ], [ "0.513", "0.3237", "0.8833", "0.3237", "0.8833", "0.3374", "0.513", "0.3374" ], [ "0.178", "0.3374", "0.434", "0.3374", "0.434", "0.3516", "0.178", "0.3516" ], [ "0.4626", "0.3384", "0.4895", "0.3384", "0.4895", "0.3508", "0.4626", "0.3508" ], [ "0.513", "0.3403", "0.885", "0.3403", "0.885", "0.354", "0.513", "0.354" ], [ "0.5146", "0.357", "0.885", "0.357", "0.885", "0.3706", "0.5146", "0.3706" ], [ "0.12006", "0.3682", "0.4878", "0.3682", "0.4878", "0.379", "0.12006", "0.379" ], [ "0.513", "0.3723", "0.885", "0.3723", "0.885", "0.386", "0.513", "0.386" ], [ "0.1184", "0.3818", "0.4895", "0.383", "0.4895", "0.398", "0.1184", "0.3967" ], [ "0.511", "0.3872", "0.8867", "0.3884", "0.8867", "0.4026", "0.511", "0.402" ], [ "0.1184", "0.398", "0.3862", "0.3997", "0.3862", "0.4133", "0.1184", "0.411" ], [ "0.513", "0.4062", "0.5996", "0.4062", "0.5996", "0.4175", "0.513", "0.4175" ], [ "0.1335", "0.4133", "0.4534", "0.412", "0.4534", "0.43", "0.1335", "0.4312" ], [ "0.532", "0.4204", "0.8833", "0.4204", "0.8833", "0.4353", "0.532", "0.4353" ], [ "0.12006", "0.4316", "0.4895", "0.4316", "0.4895", "0.4465", "0.12006", "0.4465" ], [ "0.511", "0.4358", "0.8867", "0.4353", "0.8867", "0.4495", "0.511", "0.4507" ], [ "0.1184", "0.449", "0.2822", "0.449", "0.2822", "0.462", "0.1184", "0.462" ], [ "0.513", "0.4531", "0.8833", "0.4531", "0.8833", "0.466", "0.513", "0.466" ], [ "0.12006", "0.4739", "0.3896", "0.4739", "0.3896", "0.488", "0.12006", "0.488" ], [ "0.5146", "0.4685", "0.882", "0.4685", "0.882", "0.4822", "0.5146", "0.4822" ], [ "0.513", "0.485", "0.885", "0.485", "0.885", "0.4988", "0.513", "0.4988" ], [ "0.11584", "0.4946", "0.4895", "0.4954", "0.4895", "0.5103", "0.11584", "0.509" ], [ "0.513", "0.5005", "0.8833", "0.5005", "0.8833", "0.5156", "0.513", "0.5156" ], [ "0.1184", "0.5127", "0.4902", "0.5127", "0.4902", "0.5254", "0.1184", "0.5254" ], [ "0.5156", "0.5186", "0.5415", "0.5186", "0.5415", "0.527", "0.5156", "0.527" ], [ "0.534", "0.517", "0.858", "0.5166", "0.858", "0.531", "0.534", "0.532" ], [ "0.11755", "0.527", "0.492", "0.528", "0.492", "0.542", "0.11755", "0.5415" ], [ "0.5156", "0.5337", "0.885", "0.5337", "0.885", "0.5474", "0.5156", "0.5474" ], [ "0.1184", "0.5444", "0.4895", "0.5444", "0.4895", "0.558", "0.1184", "0.558" ], [ "0.513", "0.55", "0.8833", "0.55", "0.8833", "0.5635", "0.513", "0.5635" ], [ "0.1184", "0.56", "0.4902", "0.56", "0.4902", "0.5737", "0.1184", "0.5737" ], [ "0.511", "0.564", "0.885", "0.564", "0.885", "0.582", "0.511", "0.582" ], [ "0.1184", "0.5767", "0.4902", "0.5767", "0.4902", "0.5903", "0.1184", "0.5903" ], [ "0.5146", "0.583", "0.8135", "0.583", "0.8135", "0.5933", "0.5146", "0.5933" ], [ "0.12177", "0.594", "0.4895", "0.594", "0.4895", "0.605", "0.12177", "0.605" ], [ "0.532", "0.597", "0.885", "0.597", "0.885", "0.611", "0.532", "0.611" ], [ "0.11755", "0.607", "0.4878", "0.607", "0.4878", "0.6216", "0.11755", "0.6216" ], [ "0.511", "0.6133", "0.8833", "0.6123", "0.8833", "0.627", "0.511", "0.6274" ], [ "0.1184", "0.6245", "0.4902", "0.6245", "0.4902", "0.638", "0.1184", "0.638" ], [ "0.513", "0.63", "0.8867", "0.63", "0.8867", "0.6445", "0.513", "0.6445" ], [ "0.1184", "0.641", "0.3267", "0.641", "0.3267", "0.655", "0.1184", "0.655" ], [ "0.5146", "0.6465", "0.882", "0.6465", "0.882", "0.66", "0.5146", "0.66" ], [ "0.12006", "0.6665", "0.403", "0.6665", "0.403", "0.6797", "0.12006", "0.6797" ], [ "0.513", "0.661", "0.8833", "0.661", "0.8833", "0.675", "0.513", "0.675" ], [ "0.5146", "0.6787", "0.885", "0.6787", "0.885", "0.6924", "0.5146", "0.6924" ], [ "0.12006", "0.685", "0.4895", "0.685", "0.4895", "0.6987", "0.12006", "0.6987" ], [ "0.5146", "0.695", "0.8833", "0.695", "0.8833", "0.708", "0.5146", "0.708" ], [ "0.1184", "0.7017", "0.4895", "0.7017", "0.4895", "0.7153", "0.1184", "0.7153" ], [ "0.513", "0.709", "0.8833", "0.7104", "0.8833", "0.7246", "0.513", "0.7236" ], [ "0.1184", "0.7173", "0.4878", "0.7173", "0.4878", "0.731", "0.1184", "0.731" ], [ "0.5146", "0.728", "0.6465", "0.728", "0.6465", "0.7383", "0.5146", "0.7383" ], [ "0.1184", "0.734", "0.4895", "0.734", "0.4895", "0.7476", "0.1184", "0.7476" ], [ "0.12177", "0.752", "0.4878", "0.752", "0.4878", "0.7627", "0.12177", "0.7627" ], [ "0.6826", "0.749", "0.705", "0.749", "0.705", "0.761", "0.6826", "0.761" ], [ "0.6055", "0.7627", "0.6396", "0.7695", "0.634", "0.782", "0.5996", "0.775" ], [ "0.706", "0.764", "0.796", "0.764", "0.796", "0.7793", "0.706", "0.7793" ], [ "0.8564", "0.7627", "0.8867", "0.7627", "0.8867", "0.7793", "0.8564", "0.7793" ], [ "0.679", "0.781", "0.708", "0.781", "0.708", "0.7935", "0.679", "0.7935" ], [ "0.12006", "0.7666", "0.4895", "0.7666", "0.4895", "0.78", "0.12006", "0.78" ], [ "0.1184", "0.782", "0.4895", "0.782", "0.4895", "0.796", "0.1184", "0.796" ], [ "0.11755", "0.799", "0.4895", "0.799", "0.4895", "0.8125", "0.11755", "0.8125" ], [ "0.513", "0.8027", "0.7886", "0.8027", "0.7886", "0.8164", "0.513", "0.8164" ], [ "0.1184", "0.814", "0.3838", "0.814", "0.3838", "0.8276", "0.1184", "0.8276" ], [ "0.1377", "0.8306", "0.4937", "0.8306", "0.4937", "0.844", "0.1377", "0.844" ], [ "0.513", "0.8276", "0.843", "0.8276", "0.843", "0.8413", "0.513", "0.8413" ], [ "0.1184", "0.847", "0.4937", "0.847", "0.4937", "0.861", "0.1184", "0.861" ], [ "0.513", "0.847", "0.8833", "0.847", "0.8833", "0.861", "0.513", "0.861" ], [ "0.1184", "0.8633", "0.4937", "0.8633", "0.4937", "0.877", "0.1184", "0.877" ], [ "0.513", "0.8633", "0.8867", "0.8633", "0.8867", "0.877", "0.513", "0.877" ], [ "0.1184", "0.88", "0.4954", "0.88", "0.4954", "0.8936", "0.1184", "0.8936" ], [ "0.5146", "0.8813", "0.8833", "0.8813", "0.8833", "0.891", "0.5146", "0.891" ], [ "0.1184", "0.8955", "0.4895", "0.8955", "0.4895", "0.909", "0.1184", "0.909" ], [ "0.513", "0.8955", "0.8833", "0.8955", "0.8833", "0.909", "0.513", "0.909" ] ]

[ "U, Well, I'm afraid my cooking isn't to your taste.", "UWell, I'm afraid my cooking isn't to your taste", "cooking isn't", "tast", "Uz Actull, I like it very much.", "Ilikeit", "Uz Actually I like it very much.", "I amglad", "Us. I'm glad you say that. Let me serve you more fish.", "Ue. OK, I wil he will be very glad to see you, too.", "you say - that", "Ua, Thanks. I didn't know you were good at cooking.", "Us. Why not bring your wife next time?", "me serve→ fish", "I didn'tknow", "Us. Why not bring your.wife next time?", "U.Thanks.I didn't know you wer", "egoodatcookinc", "you were→ good", "Ug, OK,I will. She wilbe very glad to see you, too", "U, I'm glad you say that. Let me serve you more fish.", "she be→ glad", "Original Context", "PermutedConte", "SVOTriplets", "Figure 2: Structural language modeling manipulations.", "multi-layer bidirectional Transformerfor learning", "of the semantic connections among utterances in", "contextualized representations,whichis defined as", "the context. The idea is similar to autoencoding", "(AE) which aims to reconstruct the original data", "H = FFN(MultiHead(K, Q, V),", "(1)", "from corrupted input (Yang et al., 2019). Given", "permuted dialogue contexts that comprise utter-", "whereK.O.Varepackedfromtheinputseguence", "ances in random orders, we maximize the expected", "representation X. As the common practice, we set", "log-likelihood of a sequence of the original ground-", "K = Q- V in the implementation.", "truth order.", "For the following part, we use H", "The goal of the utterance order restoration", "[hi, : : : , hn to denote the last-layer hidden states", "is to organize randomly shuffled utterances of", "of the input sequence.", "a conversation into a coherent dialogue context", "3.2 SPIDER Training Objectives", "We extract the hidden states of [EoU] from H", "as the representation of each utterance. For-", "To simulate the dialogue-like features, we propose", "mally, given an utterance sequence denoted as", "two pre-training objectives in addition to the origi-", "C", "=[Hur;Huzi...;Huk]with order o", "nal LM objectives: 1) utterance order restoration,", "[o1;02;...;ok], where K means the number of", "which predicts the order of the permuted utterances", "maximum positions to be predicted. We expect an", "in dialogue context; 2) sentence backbone regular-", "ordered context C* = [uor; uos;...; uo] is the", "ization, which regularizes the model to improve", "mostcoherentpermutationofutterances.", "thefactualcorrectnessofsummarizedsubject-verb-", "As predicting the permuted orders is a more chal-", "object triplets. The utterance manipulations are", "lenging optimization problem than NSP and SOP", "shown in Figure 2. The following subsections de-", "tasks due to the large searching space of permuta-", "scribe the objectives in turn.", "tions and causes slow convergence in preliminary", "3.2.1 Utterance Order Restoration", "experiments, we choose to only predict the order of", "the last few permuted utterances by a permutation", "Coherence is an essential aspect of conversation", "ratio to control the maximum number of permu-", "modeling. In a coherent discourse, utterances", "tations: K' = K * 8. The UOR training objective", "should respect specific orders of relations and logic", "is then formed as:", "The ordering of utterances in a dialogue context", "determinesthesemanticoftheconversation.There", "K'", "ILuor", "[ok log ou] ,", "(2)", "k=1", "fore, learning to order a set of disordered utterances", "in such a way that maximizes the discourse co-", "herence will have a critical impact in learning the", "where ok denotes the predicted order.", "representation of dialogue contexts.", "However, most previous studies focused on mod-", "3.2.2Sentence Backbone Regularization", "eling the semantic relevance between the context", "The sentence backbone regularization objective is", "and the response candidate. Here we introduce", "motivated to guide the model to distinguish the in-", "utterance-level position modeling, i.e., utterance or-", "ternal relation of thefact triplets that are extracted", "der restoration to encourage the model to be aware", "from each utterance, which would be helpful to" ]

{ "arxiv_id": "2105.10956", "language": "en", "page_idx": 3 }

[ [ "0.241", "0.0831", "0.763", "0.0831", "0.763", "0.0974", "0.241", "0.0974" ], [ "0.3418", "0.1027", "0.66", "0.1027", "0.66", "0.11755", "0.3418", "0.11755" ], [ "0.3494", "0.1426", "0.6533", "0.1401", "0.6533", "0.158", "0.3494", "0.1603" ], [ "0.3105", "0.1746", "0.6978", "0.1746", "0.6978", "0.1895", "0.3105", "0.1895" ], [ "0.241", "0.1882", "0.764", "0.1895", "0.764", "0.2073", "0.241", "0.206" ], [ "0.2544", "0.209", "0.756", "0.209", "0.756", "0.2227", "0.2544", "0.2227" ], [ "0.2644", "0.2632", "0.3435", "0.2654", "0.3435", "0.2808", "0.2644", "0.279" ], [ "0.511", "0.2654", "0.882", "0.2654", "0.882", "0.279", "0.511", "0.279" ], [ "0.513", "0.282", "0.882", "0.282", "0.882", "0.2957", "0.513", "0.2957" ], [ "0.1486", "0.2957", "0.4592", "0.2957", "0.4592", "0.31", "0.1486", "0.31" ], [ "0.511", "0.2976", "0.885", "0.2976", "0.885", "0.3113", "0.511", "0.3113" ], [ "0.02351", "0.306", "0.06635", "0.306", "0.0638", "0.4407", "0.02016", "0.4407" ], [ "0.1486", "0.3123", "0.4575", "0.3123", "0.4575", "0.3225", "0.1486", "0.3225" ], [ "0.5146", "0.3142", "0.885", "0.3142", "0.885", "0.328", "0.5146", "0.328" ], [ "0.147", "0.3247", "0.4592", "0.3247", "0.4592", "0.3384", "0.147", "0.3384" ], [ "0.513", "0.33", "0.8867", "0.33", "0.8867", "0.3445", "0.513", "0.3445" ], [ "0.1486", "0.339", "0.461", "0.339", "0.461", "0.3528", "0.1486", "0.3528" ], [ "0.5146", "0.3467", "0.8867", "0.3467", "0.8867", "0.3599", "0.5146", "0.3599" ], [ "0.1503", "0.3552", "0.4575", "0.3552", "0.4575", "0.3652", "0.1503", "0.3652" ], [ "0.147", "0.3665", "0.4592", "0.3665", "0.4592", "0.3801", "0.147", "0.3801" ], [ "0.513", "0.361", "0.885", "0.3623", "0.885", "0.3772", "0.513", "0.376" ], [ "0.147", "0.3813", "0.461", "0.3813", "0.461", "0.3943", "0.147", "0.3943" ], [ "0.513", "0.3777", "0.882", "0.3777", "0.882", "0.3926", "0.513", "0.3926" ], [ "0.1486", "0.3955", "0.4592", "0.3955", "0.4592", "0.4092", "0.1486", "0.4092" ], [ "0.5146", "0.3955", "0.818", "0.3955", "0.818", "0.4092", "0.5146", "0.4092" ], [ "0.147", "0.4104", "0.461", "0.4104", "0.461", "0.4233", "0.147", "0.4233" ], [ "0.532", "0.411", "0.885", "0.411", "0.885", "0.4246", "0.532", "0.4246" ], [ "0.1503", "0.4258", "0.4592", "0.4258", "0.4592", "0.4358", "0.1503", "0.4358" ], [ "0.513", "0.4275", "0.8833", "0.4275", "0.8833", "0.4412", "0.513", "0.4412" ], [ "0.02351", "0.434", "0.0638", "0.434", "0.0638", "0.5127", "0.02351", "0.5127" ], [ "0.147", "0.4382", "0.4592", "0.4382", "0.4592", "0.452", "0.147", "0.452" ], [ "0.5146", "0.4436", "0.8833", "0.4436", "0.8833", "0.4573", "0.5146", "0.4573" ], [ "0.1486", "0.4536", "0.4575", "0.4536", "0.4575", "0.4644", "0.1486", "0.4644" ], [ "0.513", "0.4602", "0.882", "0.4602", "0.882", "0.4739", "0.513", "0.4739" ], [ "0.1486", "0.4685", "0.4575", "0.4685", "0.4575", "0.4785", "0.1486", "0.4785" ], [ "0.147", "0.481", "0.461", "0.481", "0.461", "0.4946", "0.147", "0.4946" ], [ "0.513", "0.4756", "0.882", "0.4756", "0.882", "0.4893", "0.513", "0.4893" ], [ "0.147", "0.4954", "0.4592", "0.4954", "0.4592", "0.509", "0.147", "0.509" ], [ "0.511", "0.4912", "0.885", "0.4912", "0.885", "0.505", "0.511", "0.505" ], [ "0.147", "0.509", "0.4592", "0.509", "0.4592", "0.5225", "0.147", "0.5225" ], [ "0.5146", "0.5083", "0.885", "0.5083", "0.885", "0.5215", "0.5146", "0.5215" ], [ "0.1486", "0.524", "0.4592", "0.524", "0.4592", "0.5376", "0.1486", "0.5376" ], [ "0.01932", "0.513", "0.05878", "0.513", "0.0638", "0.7236", "0.02351", "0.7236" ], [ "0.513", "0.524", "0.8833", "0.5225", "0.8833", "0.5376", "0.513", "0.538" ], [ "0.1486", "0.539", "0.4592", "0.539", "0.4592", "0.55", "0.1486", "0.55" ], [ "0.513", "0.5405", "0.8867", "0.5405", "0.8867", "0.554", "0.513", "0.554" ], [ "0.147", "0.552", "0.461", "0.552", "0.461", "0.5654", "0.147", "0.5654" ], [ "0.5146", "0.5557", "0.885", "0.5557", "0.885", "0.5693", "0.5146", "0.5693" ], [ "0.147", "0.5664", "0.4592", "0.5664", "0.4592", "0.58", "0.147", "0.58" ], [ "0.511", "0.5723", "0.8867", "0.572", "0.8867", "0.586", "0.511", "0.5874" ], [ "0.147", "0.58", "0.4634", "0.58", "0.4634", "0.5933", "0.147", "0.5933" ], [ "0.147", "0.5933", "0.4207", "0.5933", "0.4207", "0.607", "0.147", "0.607" ], [ "0.5146", "0.589", "0.885", "0.589", "0.885", "0.6025", "0.5146", "0.6025" ], [ "0.511", "0.604", "0.8867", "0.605", "0.8867", "0.618", "0.511", "0.6177" ], [ "0.1368", "0.626", "0.2612", "0.626", "0.2612", "0.6396", "0.1368", "0.6396" ], [ "0.5146", "0.6206", "0.8867", "0.6206", "0.8867", "0.6343", "0.5146", "0.6343" ], [ "0.532", "0.637", "0.8833", "0.637", "0.8833", "0.651", "0.532", "0.651" ], [ "0.12006", "0.653", "0.4895", "0.653", "0.4895", "0.6665", "0.12006", "0.6665" ], [ "0.511", "0.652", "0.8833", "0.654", "0.8833", "0.6675", "0.511", "0.665" ], [ "0.12006", "0.6685", "0.4937", "0.6685", "0.4937", "0.6816", "0.12006", "0.6816" ], [ "0.513", "0.6694", "0.8833", "0.6694", "0.8833", "0.683", "0.513", "0.683" ], [ "0.12006", "0.685", "0.4895", "0.685", "0.4895", "0.6987", "0.12006", "0.6987" ], [ "0.513", "0.685", "0.8833", "0.685", "0.8833", "0.6987", "0.513", "0.6987" ], [ "0.1184", "0.6997", "0.4878", "0.6997", "0.4878", "0.7134", "0.1184", "0.7134" ], [ "0.511", "0.6997", "0.8833", "0.7007", "0.8833", "0.7153", "0.511", "0.7144" ], [ "0.12006", "0.7188", "0.4895", "0.7188", "0.4895", "0.7285", "0.12006", "0.7285" ], [ "0.5146", "0.7173", "0.8833", "0.7173", "0.8833", "0.731", "0.5146", "0.731" ], [ "0.12006", "0.7324", "0.4895", "0.7324", "0.4895", "0.745", "0.12006", "0.745" ], [ "0.513", "0.733", "0.8833", "0.733", "0.8833", "0.7476", "0.513", "0.7476" ], [ "0.1184", "0.75", "0.4895", "0.75", "0.4895", "0.763", "0.1184", "0.763" ], [ "0.513", "0.75", "0.882", "0.75", "0.882", "0.763", "0.513", "0.763" ], [ "0.1184", "0.7666", "0.4902", "0.7666", "0.4902", "0.78", "0.1184", "0.78" ], [ "0.513", "0.7666", "0.6733", "0.7666", "0.6733", "0.78", "0.513", "0.78" ], [ "0.1184", "0.782", "0.3435", "0.782", "0.3435", "0.796", "0.1184", "0.796" ], [ "0.531", "0.782", "0.885", "0.781", "0.885", "0.796", "0.531", "0.7964" ], [ "0.1351", "0.7964", "0.4895", "0.7974", "0.4895", "0.8125", "0.1351", "0.811" ], [ "0.5146", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.5146", "0.8125" ], [ "0.11755", "0.8135", "0.4902", "0.814", "0.4902", "0.829", "0.11755", "0.8276" ], [ "0.511", "0.814", "0.885", "0.814", "0.885", "0.8276", "0.511", "0.8276" ], [ "0.1184", "0.8306", "0.4954", "0.8306", "0.4954", "0.8457", "0.1184", "0.8457" ], [ "0.51", "0.8306", "0.8833", "0.8306", "0.8833", "0.844", "0.51", "0.844" ], [ "0.1184", "0.847", "0.4937", "0.847", "0.4937", "0.861", "0.1184", "0.861" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.1184", "0.8633", "0.4937", "0.8633", "0.4937", "0.877", "0.1184", "0.877" ], [ "0.513", "0.8633", "0.8833", "0.8633", "0.8833", "0.877", "0.513", "0.877" ], [ "0.11755", "0.8774", "0.4902", "0.879", "0.4902", "0.8936", "0.11755", "0.8926" ], [ "0.513", "0.879", "0.8833", "0.879", "0.8833", "0.8936", "0.513", "0.8936" ], [ "0.12006", "0.8955", "0.4895", "0.8955", "0.4895", "0.909", "0.12006", "0.909" ], [ "0.513", "0.8945", "0.848", "0.8936", "0.848", "0.9077", "0.513", "0.909" ] ]

[ "On Exposure Bias. Hallucination and Domain Shift", "in Neural Machine Translation", "Chaojun WanglRico Sennrich?,1", "1School of Informatics, University of Edinburgh", "2Department of Computational Linguistics, University of Zurich", "zippo_wang@foxmail.com, sennrich@cl.uzh.ch", "Abstract", "on previously predicted words,which would exac", "erbate error propagation. Previous work has sought", "The standard training algorithm in neural ma-", " to reduce exposure bias in training (Bengio et al.,", "", "chinetranslation(NMT)suffersfromexposure", "2015; Ranzato et al., 2016; Shen et al., 2016; Wise-", "bias, and alternative algorithms have been pro.", "man and Rush, 2016; Zhang et al., 2019). However,", "posed to mitigate this. However, the practical", "the relevance of error propagation is under debate:", "impact of exposurebiasis underdebate.In", "this paper, we link exposure bias to another", "Wu et al. (2018) argue that its role is overstated in", "well-known problem in NMT, namely the ten-", "literature, and that linguistic features explain some", "dency to generate hallucinations under domain", "of the accuracy drop at higher time steps.", "shift.In experiments on three datasets with", "Previous work has established a link between", "multipletestdomains,weshowthatexposure", "domain shift and hallucination in NMT (Koehn", "[T's]", "bias is partially to blame for hallucinations", "and Knowles, 2017; Muller et al., 2019). In this", "and that training with Minimum Risk Train", "paper, we will aim to also establish an empirical", "ing,which avoids exposure bias, can mitigate", "this. Our analysis explains why exposure bias", "link between hallucination and exposure bias. Such", "is more problematic under domain shift, and", " a link will deepen our understanding of the hallu-", "also links exposure bias to the beam search", "cination problem, but also has practical relevance,", "problem, i.e. performance deterioration with", "050", "e.g. to help predicting in which settings the use", "increasingbeam size.Ourresultsprovidea", "of sequence-level objectives is likely to be helpful.", "new justification for methods that reduce ex-", "We further empirically confirm the link between", "posure bias: even if they do not increase per.", "exposure bias and the ‘beam search problem', i.e.", "formance on in-domain test sets, they can in-", "crease model robustness to domain shift.", "the fact that translation quality does not increase", "consistentlywithbeamsize(KoehnandKnowles,", "Introduction", "2017; Ott et al., 2018; Stahlberg and Byrne, 2019).", "We base our experiments on German→English", "Neural Machine Translation (NMT) has advanced", "IWSLT'14, and two datasets used to investigate", "the state of the art in MT (Sutskever et al., 2014;", "domain robustness by Muller et al. (2019): a selec-", "Bahdanau et al., 2015; Vaswani et al., 2017), but is", "tion of corpora from OPUS (Lison and Tiedemann.", "susceptible to domain shift. Koehn and Knowles", "2016) for German→English, and a low-resource", "(2017)considerout-of-domaintranslationoneof", "German→Romansh scenario. We experiment with", "the key challenges in NMT. Such translations may", "Minimum Risk Training (MRT) (Och, 2003; Shen", "be fluent, but completely unrelated to the input", "et al., 2016), a training objective which inherently", "(hallucinations), and their misleading nature makes", "avoids exposure bias.", "them particularly problematic.", "Our experiments show that MRT indeed im-", "We hypothesise that exposure bias (Ranzato", "proves quality more in out-of-domain settings, and", "et al., 2016), a discrepancy between training and", "reduces the amount of hallucination. Our analysis", "inference, makes this problem worse. Specifi-", "of translation uncertainty also shows how the MLE", "cally, training with teacher forcing only exposes", "baseline over-estimates the probability of random", "the model to gold history, while previous predic-", "translations at all but the initial time steps, and how", "tions during inference may be erroneous. Thus, the", "MRT mitigates this problem. Finally, we show that", "model trained with teacher forcing may over-rely", "the beam search problem is reduced by MRT." ]

{ "arxiv_id": "2005.03642", "language": "en", "page_idx": 0 }

[ [ "0.12006", "0.0736", "0.361", "0.076", "0.361", "0.0903", "0.12006", "0.0885" ], [ "0.511", "0.076", "0.8833", "0.0748", "0.8833", "0.08905", "0.511", "0.0903" ], [ "0.5146", "0.09265", "0.885", "0.09265", "0.885", "0.1069", "0.5146", "0.1069" ], [ "0.1184", "0.0998", "0.4902", "0.0998", "0.4902", "0.1134", "0.1184", "0.1134" ], [ "0.513", "0.1081", "0.885", "0.1081", "0.885", "0.1217", "0.513", "0.1217" ], [ "0.11755", "0.1164", "0.4902", "0.1164", "0.4902", "0.13", "0.11755", "0.13" ], [ "0.513", "0.1223", "0.885", "0.1247", "0.885", "0.1395", "0.513", "0.1372" ], [ "0.11755", "0.1318", "0.2747", "0.1312", "0.2747", "0.1455", "0.11755", "0.1467" ], [ "0.5156", "0.1426", "0.8833", "0.1426", "0.8833", "0.1526", "0.5156", "0.1526" ], [ "0.2585", "0.1561", "0.2805", "0.1561", "0.2805", "0.1692", "0.2585", "0.1692" ], [ "0.513", "0.1567", "0.8867", "0.1567", "0.8867", "0.1704", "0.513", "0.1704" ], [ "0.1351", "0.1716", "0.2007", "0.1716", "0.2007", "0.1859", "0.1351", "0.1859" ], [ "0.2107", "0.1716", "0.2393", "0.1716", "0.2393", "0.1853", "0.2107", "0.1853" ], [ "0.2551", "0.1729", "0.2688", "0.1729", "0.2688", "0.1812", "0.2551", "0.1812" ], [ "0.293", "0.1716", "0.4895", "0.1716", "0.4895", "0.1859", "0.293", "0.1859" ], [ "0.513", "0.1729", "0.885", "0.1729", "0.885", "0.1859", "0.513", "0.1859" ], [ "0.1957", "0.1895", "0.2837", "0.187", "0.2837", "0.2019", "0.1957", "0.2037" ], [ "0.513", "0.1895", "0.885", "0.1895", "0.885", "0.2031", "0.513", "0.2031" ], [ "0.11755", "0.2102", "0.4902", "0.2114", "0.4902", "0.2257", "0.11755", "0.2245" ], [ "0.513", "0.2048", "0.8867", "0.2048", "0.8867", "0.2185", "0.513", "0.2185" ], [ "0.511", "0.2203", "0.885", "0.2197", "0.885", "0.234", "0.511", "0.2351" ], [ "0.11755", "0.2268", "0.4895", "0.2268", "0.4895", "0.2411", "0.11755", "0.2411" ], [ "0.11755", "0.2423", "0.492", "0.2434", "0.492", "0.2583", "0.11755", "0.257" ], [ "0.511", "0.2369", "0.885", "0.2369", "0.885", "0.2505", "0.511", "0.2505" ], [ "0.511", "0.253", "0.8833", "0.2517", "0.8833", "0.2666", "0.511", "0.2673" ], [ "0.1184", "0.26", "0.4902", "0.26", "0.4902", "0.275", "0.1184", "0.275" ], [ "0.513", "0.2695", "0.885", "0.2695", "0.885", "0.284", "0.513", "0.284" ], [ "0.1184", "0.2766", "0.4902", "0.2766", "0.4902", "0.2903", "0.1184", "0.2903" ], [ "0.513", "0.2861", "0.7886", "0.2861", "0.7886", "0.2998", "0.513", "0.2998" ], [ "0.1184", "0.2922", "0.4626", "0.2922", "0.4626", "0.3071", "0.1184", "0.3071" ], [ "0.531", "0.3018", "0.8833", "0.3018", "0.8833", "0.3154", "0.531", "0.3154" ], [ "0.1368", "0.308", "0.4902", "0.308", "0.4902", "0.3218", "0.1368", "0.3218" ], [ "0.1184", "0.3237", "0.4937", "0.3237", "0.4937", "0.3374", "0.1184", "0.3374" ], [ "0.513", "0.3184", "0.8833", "0.3184", "0.8833", "0.332", "0.513", "0.332" ], [ "0.513", "0.3342", "0.8135", "0.3342", "0.8135", "0.3474", "0.513", "0.3474" ], [ "0.1184", "0.3403", "0.4902", "0.3403", "0.4902", "0.354", "0.1184", "0.354" ], [ "0.11755", "0.3552", "0.492", "0.3557", "0.492", "0.3706", "0.11755", "0.3694" ], [ "0.513", "0.3623", "0.6113", "0.3623", "0.6113", "0.3772", "0.513", "0.3772" ], [ "0.1318", "0.3777", "0.492", "0.3777", "0.492", "0.3997", "0.1318", "0.3997" ], [ "0.509", "0.3843", "0.882", "0.3843", "0.882", "0.3984", "0.509", "0.3984" ], [ "0.1973", "0.4026", "0.3015", "0.4026", "0.3015", "0.4146", "0.1973", "0.4146" ], [ "0.511", "0.402", "0.8833", "0.402", "0.8833", "0.415", "0.511", "0.415" ], [ "0.11755", "0.4216", "0.4902", "0.4229", "0.4902", "0.437", "0.11755", "0.4358" ], [ "0.511", "0.4175", "0.882", "0.4175", "0.882", "0.4316", "0.511", "0.4316" ], [ "0.511", "0.434", "0.8833", "0.434", "0.8833", "0.4478", "0.511", "0.4478" ], [ "0.1184", "0.4407", "0.4895", "0.4407", "0.4895", "0.4536", "0.1184", "0.4536" ], [ "0.5146", "0.4495", "0.8833", "0.4495", "0.8833", "0.4644", "0.5146", "0.4644" ], [ "0.12006", "0.456", "0.4895", "0.456", "0.4895", "0.4702", "0.12006", "0.4702" ], [ "0.513", "0.466", "0.8833", "0.466", "0.8833", "0.4797", "0.513", "0.4797" ], [ "0.1184", "0.4727", "0.4895", "0.4727", "0.4895", "0.4863", "0.1184", "0.4863" ], [ "0.5146", "0.4822", "0.885", "0.4822", "0.885", "0.4954", "0.5146", "0.4954" ], [ "0.1184", "0.488", "0.4902", "0.488", "0.4902", "0.503", "0.1184", "0.503" ], [ "0.5146", "0.4988", "0.885", "0.4988", "0.885", "0.5127", "0.5146", "0.5127" ], [ "0.1184", "0.505", "0.4148", "0.505", "0.4148", "0.5186", "0.1184", "0.5186" ], [ "0.513", "0.514", "0.885", "0.514", "0.885", "0.528", "0.513", "0.528" ], [ "0.3064", "0.525", "0.3997", "0.524", "0.3997", "0.542", "0.3064", "0.5435" ], [ "0.1427", "0.533", "0.266", "0.535", "0.266", "0.5527", "0.1427", "0.5503" ], [ "0.513", "0.53", "0.885", "0.531", "0.885", "0.546", "0.513", "0.5444" ], [ "0.4668", "0.538", "0.4895", "0.538", "0.4895", "0.549", "0.4668", "0.549" ], [ "0.2688", "0.5464", "0.4382", "0.5415", "0.44", "0.56", "0.2703", "0.5654" ], [ "0.511", "0.546", "0.8867", "0.5474", "0.8867", "0.5625", "0.511", "0.56" ], [ "0.513", "0.5635", "0.8867", "0.5635", "0.8867", "0.5767", "0.513", "0.5767" ], [ "0.1184", "0.571", "0.4902", "0.571", "0.4902", "0.585", "0.1184", "0.585" ], [ "0.513", "0.579", "0.885", "0.579", "0.885", "0.593", "0.513", "0.593" ], [ "0.1184", "0.5874", "0.492", "0.5874", "0.492", "0.6016", "0.1184", "0.6016" ], [ "0.513", "0.5933", "0.8867", "0.5957", "0.8867", "0.61", "0.513", "0.608" ], [ "0.1184", "0.604", "0.4895", "0.604", "0.4895", "0.6177", "0.1184", "0.6177" ], [ "0.511", "0.61", "0.882", "0.61", "0.882", "0.6245", "0.511", "0.6245" ], [ "0.1184", "0.619", "0.4895", "0.619", "0.4895", "0.633", "0.1184", "0.633" ], [ "0.513", "0.627", "0.5728", "0.627", "0.5728", "0.641", "0.513", "0.641" ], [ "0.11584", "0.6343", "0.4768", "0.6353", "0.4768", "0.6494", "0.11584", "0.6484" ], [ "0.11755", "0.66", "0.2627", "0.662", "0.2627", "0.677", "0.11755", "0.675" ], [ "0.513", "0.6562", "0.644", "0.6562", "0.644", "0.6694", "0.513", "0.6694" ], [ "0.513", "0.677", "0.8867", "0.6777", "0.8867", "0.6924", "0.513", "0.6914" ], [ "0.1184", "0.686", "0.2007", "0.686", "0.2007", "0.7007", "0.1184", "0.7007" ], [ "0.513", "0.695", "0.885", "0.695", "0.885", "0.708", "0.513", "0.708" ], [ "0.11755", "0.707", "0.4937", "0.707", "0.4937", "0.7217", "0.11755", "0.7217" ], [ "0.513", "0.7104", "0.885", "0.7104", "0.885", "0.7236", "0.513", "0.7236" ], [ "0.12006", "0.7246", "0.4902", "0.7246", "0.4902", "0.7383", "0.12006", "0.7383" ], [ "0.513", "0.727", "0.8833", "0.727", "0.8833", "0.7407", "0.513", "0.7407" ], [ "0.11755", "0.7383", "0.4902", "0.7393", "0.4902", "0.7544", "0.11755", "0.753" ], [ "0.513", "0.7437", "0.885", "0.7437", "0.885", "0.7573", "0.513", "0.7573" ], [ "0.1184", "0.756", "0.4902", "0.756", "0.4902", "0.7695", "0.1184", "0.7695" ], [ "0.513", "0.759", "0.8833", "0.759", "0.8833", "0.7725", "0.513", "0.7725" ], [ "0.1184", "0.7725", "0.4895", "0.7725", "0.4895", "0.786", "0.1184", "0.786" ], [ "0.513", "0.7754", "0.8716", "0.7754", "0.8716", "0.789", "0.513", "0.789" ], [ "0.11755", "0.7876", "0.4895", "0.7876", "0.4895", "0.802", "0.11755", "0.802" ], [ "0.531", "0.7915", "0.8833", "0.7915", "0.8833", "0.8047", "0.531", "0.8047" ], [ "0.1184", "0.8047", "0.4785", "0.8047", "0.4785", "0.8184", "0.1184", "0.8184" ], [ "0.513", "0.808", "0.8867", "0.808", "0.8867", "0.822", "0.513", "0.822" ], [ "0.1351", "0.821", "0.4878", "0.821", "0.4878", "0.8345", "0.1351", "0.8345" ], [ "0.11755", "0.8374", "0.4895", "0.835", "0.4895", "0.8496", "0.11755", "0.8516" ], [ "0.532", "0.832", "0.869", "0.8345", "0.869", "0.8486", "0.532", "0.8467" ], [ "0.513", "0.847", "0.6094", "0.847", "0.6094", "0.8594", "0.513", "0.8594" ], [ "0.1351", "0.855", "0.4902", "0.857", "0.4902", "0.8745", "0.1351", "0.8726" ], [ "0.5366", "0.8594", "0.5786", "0.8594", "0.5786", "0.8706", "0.5366", "0.8706" ], [ "0.629", "0.858", "0.6978", "0.8555", "0.6978", "0.8706", "0.629", "0.8726" ], [ "0.7524", "0.86", "0.7734", "0.86", "0.7734", "0.8726", "0.7524", "0.8726" ], [ "0.8296", "0.86", "0.8867", "0.86", "0.8867", "0.8716", "0.8296", "0.8716" ], [ "0.1184", "0.8706", "0.44", "0.8706", "0.44", "0.8843", "0.1184", "0.8843" ], [ "0.5146", "0.8716", "0.7153", "0.8716", "0.7153", "0.8853", "0.5146", "0.8853" ], [ "0.1411", "0.8843", "0.4106", "0.8843", "0.4106", "0.898", "0.1411", "0.898" ], [ "0.513", "0.883", "0.881", "0.883", "0.881", "0.8965", "0.513", "0.8965" ], [ "0.12006", "0.898", "0.2761", "0.898", "0.2761", "0.9077", "0.12006", "0.9077" ], [ "0.5366", "0.8955", "0.876", "0.8955", "0.876", "0.909", "0.5366", "0.909" ] ]

[ "2Minimum Risk Training", "For DE-→EN, data comes from OPUS (Lison", "and Tiedemann, 2016), and is comprised of five", "The de-facto standard training objective in NMT", "domains: medical, IT, law, koran and subtitles", "is to minimize the negative log-likelihood C() of", "We use medical for training and development,", "the training data Dl:", "and report results on an in-domain test set and", "lyl", "the four other domains (out-of-domain; OOD).", "L(0) =", "", "", "-log P(yt|x, y<t;0) (1)", "German-→Romansh (DE→RM) is a low-resource", "(x,y)eD t=1", "language pair where robustness to domain shift is", "where x and y are the source and target sequence,", "of practical relevance. The training data is from the", "Allegra corpus (Scherrer and Cartoni, 2012) (law", "respectively, yt is the tth token in y, and y<t de-", "notes all previous tokens. MLE is typically per-", "domain) with 100 000 sentence pairs. The test do-", "main are blogs, using data from Convivenza3. We", "formed with teacher forcing, where y<t are ground-", "have access to 2000 sentences for development and", "truth labels in training, which creates a mismatch", "testing, respectively, in each domain.", "to inference, where y<t are model predictions.", "We tokenise and truecase data sets with", "Minimum Risk Training (MRT) is a sequence-", "level objective that avoids this problem. Specifi-", "Moses (Koehn et al., 2007), and use shared BPE", "with 32 000 units (Sennrich et al., 2016).", "cally, the objective function of MRT is the expected", "loss (risk)with respect to the posterior distribution:", "3.2Model", "R() = P(yx;0) (y,y) (2)", " We implement MRT in the Nematus toolkit (Sen", "(x,y)eDyey(x)", "nrich et al., 2017). All our experiments use", "in which the loss (y,y) indicates the discrep-", "the Transformer architecture (Vaswani et al.", "2017). Following Edunov et al. (2018), we use", "ancy between the gold translation y and the model", "1-BLEUsmooth (Lin and Och, 2004) as the MRT", "prediction y. Due to the intractable search space,", "loss. Models are pre-trained with the token-level", "the posterior distribution y(x) is approximated by", "objective MLE and then fine-tuned with MRT.", "a subspace S(x) by sampling a certain number of", "Hyperparameters mostly follow previous work", "candidate translations, and normalizing:", "(Edunov et al., 2018; Muller et al., 2019); for", "P (y|x; 0)%", "P(yx; 0,α) =", "MRT, we conduct limited hyperparameter search", "(3)", "Ey'es(x) P (yx; 0)°", "on the IWSLT'14 development set, including learn-", "ing rate, batch size, and the sharpness parameter", "where α is a hyperparameter to control the sharp-", "Q. We set the number of candidate translations for", "ness of the subspace. Based on preliminary results,", "MRT to 4 to balance effectiveness and efficiency.", "we use random sampling to generate candidate", "Detailed hyperparameters are reported in the Ap", "translations, and following Edunov et al. (2018), do", "pendix.", "not add the reference translation to the subspace.", "3Experiments", "3.3Evaluation", "For comparison to previous work, we report low-", "3.1 Data", "ercased, tokenised BLEU (Papineni et al., 2002)", "To verify the effectiveness of our MRT implemen-", "with multi-bleu.perl for IWSLT'14, and cased, deto-", "tation on top of a strong Transformer baseline", "kenised BLEU with SacreBLEU (Post, 2018) oth-", "(Vaswani et al., 2017), we first conduct experi-", "erwise. For settings with domain shift, we report", "ments on IWSLT'14 German→English (DE-→EN)", "average and standard deviation of 3 independent", "(Cettolo et al., 2014), which consists of 180000", "training runs to account for optimizer instability.", "sentence pairs. We follow previous work for data", "The manual evaluation was performed by two", "splits (Ranzato et al., 2016; Edunov et al., 2018).", "native speakers of German who completed bilin-", "For experiments with domain shift, we use data", "sets and preprocessing as Muller et al. (2019)?.", "3https://www.suedostschweiz.ch/blogs/", "convivenza", "'This is equivalent to maximizing the likelihood of the", "+Code", "available", "at", "https:", "data, hence Maximum Likelihood Estimation (MLE).", "//github.com/zippotju/", "2https://github.com/zurichNLP/", "Exposure-Bias-Hallucination-Domain-shift", "domain-robustness", "5 signature: BLEU+c.mixed+#.1+s.exp+tok.13a+v.1.4.2" ]

{ "arxiv_id": "2005.03642", "language": "en", "page_idx": 1 }

[ [ "0.2183", "0.0766", "0.3208", "0.0766", "0.3208", "0.0885", "0.2183", "0.0885" ], [ "0.3628", "0.0766", "0.4668", "0.0766", "0.4668", "0.0885", "0.3628", "0.0885" ], [ "0.513", "0.0766", "0.8833", "0.0766", "0.8833", "0.0903", "0.513", "0.0903" ], [ "0.2048", "0.0956", "0.2468", "0.0956", "0.2468", "0.10925", "0.2048", "0.10925" ], [ "0.2551", "0.0956", "0.297", "0.0956", "0.297", "0.10925", "0.2551", "0.10925" ], [ "0.3494", "0.0956", "0.388", "0.0956", "0.388", "0.1081", "0.3494", "0.1081" ], [ "0.3997", "0.0956", "0.4417", "0.0956", "0.4417", "0.10925", "0.3997", "0.10925" ], [ "0.513", "0.0932", "0.8867", "0.0932", "0.8867", "0.1069", "0.513", "0.1069" ], [ "0.315", "0.0974", "0.33", "0.0974", "0.33", "0.1069", "0.315", "0.1069" ], [ "0.12427", "0.0968", "0.1973", "0.0968", "0.1973", "0.1081", "0.12427", "0.1081" ], [ "0.4575", "0.0974", "0.4744", "0.0974", "0.4744", "0.1069", "0.4575", "0.1069" ], [ "0.12177", "0.111", "0.1755", "0.1134", "0.1738", "0.1277", "0.12006", "0.1259" ], [ "0.2083", "0.1134", "0.2418", "0.1134", "0.2418", "0.1259", "0.2083", "0.1259" ], [ "0.2585", "0.1134", "0.293", "0.1134", "0.293", "0.1259", "0.2585", "0.1259" ], [ "0.3064", "0.1134", "0.339", "0.1134", "0.339", "0.1259", "0.3064", "0.1259" ], [ "0.351", "0.1134", "0.3838", "0.1134", "0.3838", "0.1259", "0.351", "0.1259" ], [ "0.403", "0.1134", "0.4375", "0.1134", "0.4375", "0.1259", "0.403", "0.1259" ], [ "0.4492", "0.1134", "0.4836", "0.1134", "0.4836", "0.1247", "0.4492", "0.1247" ], [ "0.511", "0.10925", "0.5938", "0.1069", "0.5938", "0.1217", "0.511", "0.12354" ], [ "0.12177", "0.1247", "0.1873", "0.1271", "0.1873", "0.1395", "0.12177", "0.1372" ], [ "0.2083", "0.1259", "0.2418", "0.1259", "0.2418", "0.1383", "0.2083", "0.1383" ], [ "0.2585", "0.1259", "0.2913", "0.1259", "0.2913", "0.1383", "0.2585", "0.1383" ], [ "0.3064", "0.1259", "0.339", "0.1259", "0.339", "0.1383", "0.3064", "0.1383" ], [ "0.351", "0.1259", "0.3838", "0.1259", "0.3838", "0.1383", "0.351", "0.1383" ], [ "0.403", "0.1259", "0.4382", "0.1259", "0.4382", "0.1383", "0.403", "0.1383" ], [ "0.4492", "0.1259", "0.4836", "0.1259", "0.4836", "0.1383", "0.4492", "0.1383" ], [ "0.531", "0.1247", "0.8867", "0.1247", "0.8867", "0.1383", "0.531", "0.1383" ], [ "0.5146", "0.1414", "0.885", "0.1414", "0.885", "0.155", "0.5146", "0.155" ], [ "0.11584", "0.1526", "0.4902", "0.1538", "0.4902", "0.1687", "0.11584", "0.1675" ], [ "0.5146", "0.158", "0.8833", "0.158", "0.8833", "0.1716", "0.5146", "0.1716" ], [ "0.1184", "0.1692", "0.4014", "0.1692", "0.4014", "0.1829", "0.1184", "0.1829" ], [ "0.5146", "0.1733", "0.8833", "0.1733", "0.8833", "0.187", "0.5146", "0.187" ], [ "0.513", "0.1882", "0.8833", "0.1895", "0.8833", "0.2037", "0.513", "0.2031" ], [ "0.1335", "0.2048", "0.1823", "0.2048", "0.1823", "0.2162", "0.1335", "0.2162" ], [ "0.409", "0.2037", "0.4517", "0.2037", "0.4517", "0.2156", "0.409", "0.2156" ], [ "0.5146", "0.206", "0.8833", "0.206", "0.8833", "0.2203", "0.5146", "0.2203" ], [ "0.1335", "0.2203", "0.3728", "0.2203", "0.3728", "0.234", "0.1335", "0.234" ], [ "0.3955", "0.2203", "0.4265", "0.2203", "0.4265", "0.2328", "0.3955", "0.2328" ], [ "0.513", "0.2227", "0.885", "0.2227", "0.885", "0.2363", "0.513", "0.2363" ], [ "0.1368", "0.2328", "0.3745", "0.2328", "0.3745", "0.2465", "0.1368", "0.2465" ], [ "0.3938", "0.2322", "0.4668", "0.2322", "0.4668", "0.2465", "0.3938", "0.2465" ], [ "0.3938", "0.2434", "0.4265", "0.2452", "0.4248", "0.2583", "0.392", "0.2559" ], [ "0.511", "0.2382", "0.559", "0.2382", "0.559", "0.253", "0.511", "0.253" ], [ "0.1377", "0.2446", "0.3687", "0.2434", "0.3687", "0.257", "0.1377", "0.2583" ], [ "0.1351", "0.257", "0.3787", "0.257", "0.3787", "0.2708", "0.1351", "0.2708" ], [ "0.392", "0.2559", "0.4282", "0.2559", "0.4282", "0.2715", "0.392", "0.2715" ], [ "0.5156", "0.2695", "0.531", "0.2695", "0.531", "0.279", "0.5156", "0.279" ], [ "0.5366", "0.2673", "0.6206", "0.2673", "0.6206", "0.282", "0.5366", "0.282" ], [ "0.1301", "0.2737", "0.1755", "0.2737", "0.1755", "0.288", "0.1301", "0.288" ], [ "0.3955", "0.275", "0.4265", "0.275", "0.4265", "0.2874", "0.3955", "0.2874" ], [ "0.1335", "0.288", "0.1721", "0.288", "0.1721", "0.2998", "0.1335", "0.2998" ], [ "0.3955", "0.2861", "0.4668", "0.2861", "0.4668", "0.3005", "0.3955", "0.3005" ], [ "0.513", "0.2935", "0.8867", "0.2944", "0.8867", "0.3088", "0.513", "0.308" ], [ "0.11755", "0.3142", "0.4902", "0.3142", "0.4902", "0.328", "0.11755", "0.328" ], [ "0.513", "0.3123", "0.885", "0.3123", "0.885", "0.326", "0.513", "0.326" ], [ "0.1184", "0.33", "0.4475", "0.33", "0.4475", "0.3433", "0.1184", "0.3433" ], [ "0.5146", "0.328", "0.8833", "0.328", "0.8833", "0.3416", "0.5146", "0.3416" ], [ "0.5146", "0.3433", "0.885", "0.3433", "0.885", "0.357", "0.5146", "0.357" ], [ "0.513", "0.3599", "0.885", "0.3599", "0.885", "0.3735", "0.513", "0.3735" ], [ "0.1184", "0.3723", "0.4878", "0.3718", "0.4878", "0.386", "0.1184", "0.3872" ], [ "0.511", "0.3748", "0.8506", "0.3748", "0.8506", "0.3896", "0.511", "0.3896" ], [ "0.1184", "0.3884", "0.492", "0.3872", "0.492", "0.402", "0.1184", "0.4026" ], [ "0.11755", "0.4038", "0.492", "0.4038", "0.492", "0.4187", "0.11755", "0.4187" ], [ "0.511", "0.4026", "0.665", "0.4038", "0.665", "0.4187", "0.511", "0.4175" ], [ "0.1184", "0.4204", "0.4895", "0.4204", "0.4895", "0.434", "0.1184", "0.434" ], [ "0.513", "0.427", "0.885", "0.427", "0.885", "0.4412", "0.513", "0.4412" ], [ "0.1184", "0.4358", "0.4902", "0.4358", "0.4902", "0.4507", "0.1184", "0.4507" ], [ "0.513", "0.4436", "0.882", "0.4436", "0.882", "0.4573", "0.513", "0.4573" ], [ "0.1184", "0.4531", "0.4895", "0.4531", "0.4895", "0.466", "0.1184", "0.466" ], [ "0.511", "0.459", "0.8833", "0.459", "0.8833", "0.4727", "0.511", "0.4727" ], [ "0.12006", "0.4697", "0.4902", "0.4697", "0.4902", "0.4827", "0.12006", "0.4827" ], [ "0.513", "0.4744", "0.8867", "0.4744", "0.8867", "0.488", "0.513", "0.488" ], [ "0.1184", "0.484", "0.4895", "0.485", "0.4895", "0.4995", "0.1184", "0.4988" ], [ "0.513", "0.4912", "0.8867", "0.4912", "0.8867", "0.505", "0.513", "0.505" ], [ "0.1184", "0.4995", "0.4902", "0.5005", "0.4902", "0.5156", "0.1184", "0.514" ], [ "0.513", "0.509", "0.8833", "0.509", "0.8833", "0.5195", "0.513", "0.5195" ], [ "0.12006", "0.517", "0.4895", "0.517", "0.4895", "0.531", "0.12006", "0.531" ], [ "0.511", "0.5215", "0.885", "0.5225", "0.885", "0.5376", "0.511", "0.536" ], [ "0.11584", "0.532", "0.4307", "0.533", "0.4307", "0.5474", "0.11584", "0.5464" ], [ "0.5146", "0.5415", "0.6094", "0.5415", "0.6094", "0.552", "0.5146", "0.552" ], [ "0.531", "0.5557", "0.885", "0.5557", "0.885", "0.5693", "0.531", "0.5693" ], [ "0.12006", "0.564", "0.22", "0.564", "0.22", "0.576", "0.12006", "0.576" ], [ "0.513", "0.5723", "0.885", "0.5723", "0.885", "0.586", "0.513", "0.586" ], [ "0.11755", "0.585", "0.4902", "0.586", "0.4902", "0.601", "0.11755", "0.5996" ], [ "0.511", "0.5874", "0.8867", "0.5884", "0.8867", "0.6025", "0.511", "0.6016" ], [ "0.12006", "0.6025", "0.492", "0.6025", "0.492", "0.616", "0.12006", "0.616" ], [ "0.513", "0.605", "0.8867", "0.605", "0.8867", "0.618", "0.513", "0.618" ], [ "0.1184", "0.618", "0.4895", "0.618", "0.4895", "0.632", "0.1184", "0.632" ], [ "0.513", "0.6206", "0.8867", "0.6206", "0.8867", "0.6343", "0.513", "0.6343" ], [ "0.1184", "0.6343", "0.4895", "0.6343", "0.4895", "0.648", "0.1184", "0.648" ], [ "0.513", "0.637", "0.885", "0.637", "0.885", "0.651", "0.513", "0.651" ], [ "0.1184", "0.652", "0.3477", "0.652", "0.3477", "0.665", "0.1184", "0.665" ], [ "0.513", "0.653", "0.885", "0.653", "0.885", "0.6665", "0.513", "0.6665" ], [ "0.1368", "0.6685", "0.4895", "0.6685", "0.4895", "0.6816", "0.1368", "0.6816" ], [ "0.513", "0.6685", "0.885", "0.6685", "0.885", "0.6816", "0.513", "0.6816" ], [ "0.11755", "0.683", "0.4895", "0.684", "0.4895", "0.6987", "0.11755", "0.6978" ], [ "0.513", "0.685", "0.8833", "0.685", "0.8833", "0.6987", "0.513", "0.6987" ], [ "0.11755", "0.6997", "0.4878", "0.6997", "0.4878", "0.7134", "0.11755", "0.7134" ], [ "0.511", "0.7007", "0.8833", "0.7007", "0.8833", "0.7144", "0.511", "0.7144" ], [ "0.11584", "0.7163", "0.465", "0.7153", "0.465", "0.73", "0.11584", "0.731" ], [ "0.511", "0.7163", "0.882", "0.7163", "0.882", "0.73", "0.511", "0.73" ], [ "0.1368", "0.734", "0.4902", "0.734", "0.4902", "0.7476", "0.1368", "0.7476" ], [ "0.513", "0.734", "0.8833", "0.734", "0.8833", "0.7476", "0.513", "0.7476" ], [ "0.11755", "0.749", "0.4895", "0.749", "0.4895", "0.7627", "0.11755", "0.7627" ], [ "0.51", "0.7476", "0.885", "0.749", "0.885", "0.763", "0.51", "0.7627" ], [ "0.11755", "0.7656", "0.4895", "0.764", "0.4895", "0.7793", "0.11755", "0.78" ], [ "0.511", "0.7656", "0.885", "0.764", "0.885", "0.7793", "0.511", "0.78" ], [ "0.1184", "0.782", "0.4895", "0.781", "0.4895", "0.796", "0.1184", "0.7964" ], [ "0.532", "0.782", "0.8833", "0.782", "0.8833", "0.796", "0.532", "0.796" ], [ "0.12006", "0.799", "0.4902", "0.799", "0.4902", "0.8125", "0.12006", "0.8125" ], [ "0.513", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.814", "0.4902", "0.814", "0.4902", "0.8276", "0.1184", "0.8276" ], [ "0.513", "0.814", "0.8833", "0.814", "0.8833", "0.829", "0.513", "0.829" ], [ "0.1184", "0.8306", "0.4937", "0.8306", "0.4937", "0.844", "0.1184", "0.844" ], [ "0.511", "0.8306", "0.8833", "0.8306", "0.8833", "0.844", "0.511", "0.844" ], [ "0.1184", "0.8467", "0.4937", "0.8467", "0.4937", "0.861", "0.1184", "0.861" ], [ "0.511", "0.8467", "0.885", "0.8467", "0.885", "0.861", "0.511", "0.861" ], [ "0.11755", "0.8623", "0.4937", "0.8633", "0.4937", "0.8774", "0.11755", "0.877" ], [ "0.511", "0.8633", "0.8867", "0.8633", "0.8867", "0.877", "0.511", "0.877" ], [ "0.1184", "0.879", "0.4902", "0.879", "0.4902", "0.8926", "0.1184", "0.8926" ], [ "0.511", "0.88", "0.885", "0.88", "0.885", "0.8936", "0.511", "0.8936" ], [ "0.11755", "0.8945", "0.492", "0.8945", "0.492", "0.909", "0.11755", "0.909" ], [ "0.51", "0.8955", "0.8833", "0.8955", "0.8833", "0.909", "0.51", "0.909" ] ]

[ "inter-annotator", "intra-annotator", "domain performance, which is a first indicator that", "P(A)", "P(E)", "P(A)", "P(E)", "exposure bias may be more problematic under do-", "K", "annotation", "K", "fluency", "0.66", "0.38", "0.44", "0.87", "0.42", "0.77", "main shift.", "adequacy", "0.82", "0.61", "0.54", "0.93", "0.66", "0.79", "Our OOD results lag slightly behind those of", "Muller et al. (2019), but note that the techniques", "Table 1: Inter-annotator (N=250) and intra-annotator", "employed by them, namely reconstruction (Tu", "agreement (N=617) of manual evaluation.", "et al., 2017; Niu et al., 2019), subword regular-", "ization (Kudo, 2018), and noisy channel modelling", "system", "BLEU", "(Li and Jurafsky, 2016) are orthogonal to MRT. We", "ConvS2S (MLE) (Edunov et al., 2018)", "32.2", "leave the combination of these approaches to future", "ConvS2S (MRT) (Edunov et al., 2018)", "32.8 (+0.6)", "34.4", "work.", "Transformer (MLE) (Wu et al., 2019)", "DynamicConv (MLE) (Wu et al., 2019)", "35.2", "4.", "Analysis", "MLE", "34.7", "MRT", "35.2 (+0.5)", "BLEU results indicate that MRT can improve do-", "Table 2:Results for IWSLT'14 DE-→EN with MLE", "main robustness. In this section, we report on ad-", "and MRT (in brackets, improvement over MLE).", "ditional experiments to establish more direct links", "between exposure bias and domain robustness, hal-", "lucination, and the beam search problem. Experi-", "gual (German/English) high school or University", "ments are performed on DEEN OPUS data", "programs. We collected ~3600 annotations in total,", "spread over 12 configurations. We ask annotators", "4.1Hallucination", "to evaluate translations according to fluency and", "We manually evaluate the proportion of halluci-", "adequacy. For fluency, the annotator classifies a", "nated translations on out-of-domain and in-domain", "translation as fluent, partially fluent or not fluent;", "test sets. We follow the definition and evaluation", "for adequacy, as adequate, partially adequate or", "by Muller et al. (2019), considering a translation", "inadequate.We reportkappa coefficient (K) (Car-", "a hallucination if it is (partially) fluent, but un-", "letta, 1996) for inter-annotator and intra-annotator", "related in content tothe sourcetext (inadequate)", "agreement in Table l, and assess statistical signifi-", "We report the proportion of such hallucinations for", "cance with Fisher's exact test (two-tailed).", "each system.", "Results in Table 4 confirm that hallucinations", "3.4 Results", "are much more pronounced in out-of-domain test", "Table 2 shows results for IWSLT'14. We compare", "sets (33-35%) than in in-domain test sets (1-2%).", "to results by Edunov et al. (2018), who use a convo-", "MRT reduces the proportion of hallucinations on", "lutional architecture (Gehring et al., 2017), and Wu", "out-of-domain test sets (N=500 for each system;", "et al. (2019), who report results with Transfomer-", "reductions statistically significant at p < 0.05) and", "base and dynamic convolution.", "improves BLEU. Note that the two metrics do not", "With 34.7 BLEU, our baseline is competitive.", "correlate perfectly: MLE with label smoothing has", "We observe an improvement of 0.5 BLEU from", "higher BLEU (+1) than MRT based on MLE with-", "MRT, comparable to Edunov et al. (2018), although", " out label smoothing, but a similar proportion of", "we start from a stronger baseline (+2.5 BLEU)", "hallucinations. This indicates that label smooth.", "Table 3 shows results for data sets with do-", "ing increases translation quality in other aspects.", "main shift. To explore the effect of label smooth-", "while MRT has a clear effect on the number of", "ing (Szegedy et al., 2016), we train baselines with", "hallucinations, reducing it by up to 21% (relative).", "and without label smoothing. MLE with label", "A closer inspection of segments where the MLE", "smoothing performs better by itself, and we also", "system was found to hallucinate shows that some", "found MRT to be more effective on top of the ini-", "segments were scored higher in adequacy with", "tial model with label smoothing. For DE→EN,", "MRT, others lower in fluency.One example for", "MRT increases average OOD BLEU by 0.8 com-", "each case is shown in Table 5. Even the example", "pared to the MLE baseline with label smoothing;", "where MRT was considered disfluent and inade-", "for DE-→RM the improvement is 0.7 BLEU. We", " quate actually shows an attempt to cover the source", "note that MRT does not consistently improve in-", "sentence: the source word ‘Ableugner' (denier) is" ]

{ "arxiv_id": "2005.03642", "language": "en", "page_idx": 2 }

[ [ "0.4978", "0.076", "0.5615", "0.076", "0.5615", "0.0903", "0.4978", "0.0903" ], [ "0.6978", "0.076", "0.7676", "0.076", "0.7676", "0.0903", "0.6978", "0.0903" ], [ "0.4458", "0.0968", "0.513", "0.0968", "0.513", "0.1081", "0.4458", "0.1081" ], [ "0.5366", "0.0968", "0.623", "0.0968", "0.623", "0.1081", "0.5366", "0.1081" ], [ "0.6465", "0.0968", "0.714", "0.0968", "0.714", "0.1081", "0.6465", "0.1081" ], [ "0.738", "0.0968", "0.825", "0.0968", "0.825", "0.1081", "0.738", "0.1081" ], [ "0.4358", "0.11224", "0.471", "0.11224", "0.471", "0.1277", "0.4358", "0.1277" ], [ "0.1721", "0.0974", "0.2217", "0.0974", "0.2217", "0.1081", "0.1721", "0.1081" ], [ "0.5366", "0.11224", "0.5728", "0.11224", "0.5728", "0.1277", "0.5366", "0.1277" ], [ "0.636", "0.11224", "0.6724", "0.11224", "0.6724", "0.1277", "0.636", "0.1277" ], [ "0.739", "0.11224", "0.7734", "0.11224", "0.7734", "0.1277", "0.739", "0.1277" ], [ "0.4375", "0.1259", "0.4702", "0.1259", "0.4702", "0.1383", "0.4375", "0.1383" ], [ "0.1721", "0.1134", "0.3342", "0.1134", "0.3342", "0.1271", "0.1721", "0.1271" ], [ "0.538", "0.1259", "0.571", "0.1259", "0.571", "0.1383", "0.538", "0.1383" ], [ "0.638", "0.1259", "0.671", "0.1259", "0.671", "0.1383", "0.638", "0.1383" ], [ "0.739", "0.1259", "0.7715", "0.1259", "0.7715", "0.1383", "0.739", "0.1383" ], [ "0.5366", "0.136", "0.571", "0.136", "0.571", "0.152", "0.5366", "0.152" ], [ "0.1763", "0.1271", "0.336", "0.1271", "0.336", "0.1372", "0.1763", "0.1372" ], [ "0.1721", "0.1372", "0.4248", "0.1372", "0.4248", "0.1509", "0.1721", "0.1509" ], [ "0.638", "0.136", "0.6724", "0.136", "0.6724", "0.152", "0.638", "0.152" ], [ "0.739", "0.1372", "0.7715", "0.1372", "0.7715", "0.1497", "0.739", "0.1497" ], [ "0.1721", "0.1561", "0.2568", "0.1561", "0.2568", "0.1675", "0.1721", "0.1675" ], [ "0.4375", "0.1383", "0.4685", "0.1383", "0.4685", "0.1509", "0.4375", "0.1509" ], [ "0.4382", "0.1561", "0.519", "0.1561", "0.519", "0.1675", "0.4382", "0.1675" ], [ "0.546", "0.1561", "0.621", "0.1561", "0.621", "0.1675", "0.546", "0.1675" ], [ "0.6396", "0.1561", "0.721", "0.1561", "0.721", "0.1675", "0.6396", "0.1675" ], [ "0.742", "0.1561", "0.8213", "0.1561", "0.8213", "0.1675", "0.742", "0.1675" ], [ "0.1721", "0.1663", "0.22", "0.1663", "0.22", "0.1812", "0.1721", "0.1812" ], [ "0.538", "0.1663", "0.6206", "0.1663", "0.6206", "0.1776", "0.538", "0.1776" ], [ "0.6396", "0.1663", "0.721", "0.1663", "0.721", "0.1776", "0.6396", "0.1776" ], [ "0.742", "0.1675", "0.8213", "0.1675", "0.8213", "0.1787", "0.742", "0.1787" ], [ "0.538", "0.1787", "0.6187", "0.1787", "0.6187", "0.1895", "0.538", "0.1895" ], [ "0.4382", "0.1675", "0.519", "0.1675", "0.519", "0.1787", "0.4382", "0.1787" ], [ "0.1738", "0.1799", "0.2494", "0.1799", "0.2494", "0.1912", "0.1738", "0.1912" ], [ "0.4382", "0.1799", "0.519", "0.1799", "0.519", "0.1907", "0.4382", "0.1907" ], [ "0.6406", "0.1799", "0.721", "0.1799", "0.721", "0.1912", "0.6406", "0.1912" ], [ "0.742", "0.1787", "0.8213", "0.1787", "0.8213", "0.1895", "0.742", "0.1895" ], [ "0.1738", "0.1912", "0.2183", "0.1912", "0.2183", "0.2031", "0.1738", "0.2031" ], [ "0.4382", "0.1912", "0.519", "0.1912", "0.519", "0.2031", "0.4382", "0.2031" ], [ "0.54", "0.1912", "0.6206", "0.1912", "0.6206", "0.2031", "0.54", "0.2031" ], [ "0.6396", "0.1912", "0.721", "0.1912", "0.721", "0.2031", "0.6396", "0.2031" ], [ "0.742", "0.1912", "0.8228", "0.1912", "0.8228", "0.2031", "0.742", "0.2031" ], [ "0.1184", "0.2185", "0.885", "0.2185", "0.885", "0.2322", "0.1184", "0.2322" ], [ "0.1184", "0.2328", "0.867", "0.2328", "0.867", "0.2465", "0.1184", "0.2465" ], [ "0.2737", "0.275", "0.434", "0.275", "0.434", "0.288", "0.2737", "0.288" ], [ "0.513", "0.2737", "0.885", "0.2737", "0.885", "0.2874", "0.513", "0.2874" ], [ "0.3704", "0.2922", "0.444", "0.2922", "0.444", "0.3071", "0.3704", "0.3071" ], [ "0.511", "0.2893", "0.8833", "0.2893", "0.8833", "0.3027", "0.511", "0.3027" ], [ "0.266", "0.2944", "0.3586", "0.2944", "0.3586", "0.306", "0.266", "0.306" ], [ "0.1663", "0.2944", "0.2141", "0.2944", "0.2141", "0.306", "0.1663", "0.306" ], [ "0.1663", "0.31", "0.251", "0.31", "0.251", "0.3218", "0.1663", "0.3218" ], [ "0.3728", "0.31", "0.4434", "0.31", "0.4434", "0.3247", "0.3728", "0.3247" ], [ "0.5146", "0.308", "0.8833", "0.308", "0.8833", "0.3184", "0.5146", "0.3184" ], [ "0.2747", "0.3218", "0.3494", "0.3218", "0.3494", "0.3362", "0.2747", "0.3362" ], [ "0.2778", "0.3113", "0.3477", "0.3113", "0.3477", "0.3225", "0.2778", "0.3225" ], [ "0.168", "0.3237", "0.2107", "0.3237", "0.2107", "0.335", "0.168", "0.335" ], [ "0.513", "0.3225", "0.885", "0.3225", "0.885", "0.3362", "0.513", "0.3362" ], [ "0.168", "0.335", "0.2418", "0.335", "0.2418", "0.3467", "0.168", "0.3467" ], [ "0.2761", "0.335", "0.3477", "0.335", "0.3477", "0.3467", "0.2761", "0.3467" ], [ "0.3762", "0.335", "0.44", "0.335", "0.44", "0.3467", "0.3762", "0.3467" ], [ "0.2747", "0.3445", "0.3494", "0.3467", "0.3494", "0.361", "0.2747", "0.3594" ], [ "0.513", "0.339", "0.885", "0.339", "0.885", "0.3528", "0.513", "0.3528" ], [ "0.1663", "0.3457", "0.2141", "0.3457", "0.2141", "0.3599", "0.1663", "0.3599" ], [ "0.513", "0.3552", "0.885", "0.3552", "0.885", "0.3682", "0.513", "0.3682" ], [ "0.1184", "0.3735", "0.4902", "0.3735", "0.4902", "0.3872", "0.1184", "0.3872" ], [ "0.513", "0.3718", "0.8286", "0.3718", "0.8286", "0.3855", "0.513", "0.3855" ], [ "0.11755", "0.3884", "0.4702", "0.3872", "0.4702", "0.402", "0.11755", "0.4026" ], [ "0.529", "0.3872", "0.885", "0.3896", "0.885", "0.4038", "0.529", "0.402" ], [ "0.511", "0.4038", "0.885", "0.405", "0.885", "0.4192", "0.511", "0.4187" ], [ "0.2032", "0.4233", "0.3652", "0.4233", "0.3652", "0.437", "0.2032", "0.437" ], [ "0.513", "0.4216", "0.8867", "0.4216", "0.8867", "0.4353", "0.513", "0.4353" ], [ "0.2048", "0.4358", "0.3005", "0.4358", "0.3005", "0.4478", "0.2048", "0.4478" ], [ "0.126", "0.4246", "0.1721", "0.4246", "0.1721", "0.4358", "0.126", "0.4358" ], [ "0.1276", "0.437", "0.189", "0.437", "0.189", "0.4478", "0.1276", "0.4478" ], [ "0.5146", "0.437", "0.8867", "0.437", "0.8867", "0.4507", "0.5146", "0.4507" ], [ "0.1276", "0.4478", "0.1663", "0.4478", "0.1663", "0.4602", "0.1276", "0.4602" ], [ "0.2048", "0.449", "0.294", "0.449", "0.294", "0.4602", "0.2048", "0.4602" ], [ "0.1276", "0.459", "0.1628", "0.459", "0.1628", "0.4714", "0.1276", "0.4714" ], [ "0.513", "0.4536", "0.8867", "0.4536", "0.8867", "0.4673", "0.513", "0.4673" ], [ "0.2065", "0.4602", "0.3167", "0.4602", "0.3167", "0.4714", "0.2065", "0.4714" ], [ "0.511", "0.4685", "0.8833", "0.4685", "0.8833", "0.4827", "0.511", "0.4827" ], [ "0.1276", "0.478", "0.1721", "0.478", "0.1721", "0.4893", "0.1276", "0.4893" ], [ "0.2032", "0.4768", "0.4106", "0.4768", "0.4106", "0.4905", "0.2032", "0.4905" ], [ "0.2032", "0.488", "0.4895", "0.488", "0.4895", "0.502", "0.2032", "0.502" ], [ "0.5146", "0.4868", "0.882", "0.4868", "0.882", "0.4976", "0.5146", "0.4976" ], [ "0.2024", "0.4995", "0.3914", "0.502", "0.3914", "0.5156", "0.2024", "0.513" ], [ "0.1276", "0.4905", "0.189", "0.4905", "0.189", "0.5005", "0.1276", "0.5005" ], [ "0.1293", "0.5005", "0.1646", "0.5005", "0.1646", "0.513", "0.1293", "0.513" ], [ "0.513", "0.502", "0.885", "0.502", "0.885", "0.5156", "0.513", "0.5156" ], [ "0.1276", "0.513", "0.1646", "0.513", "0.1646", "0.525", "0.1276", "0.525" ], [ "0.2032", "0.513", "0.3687", "0.513", "0.3687", "0.527", "0.2032", "0.527" ], [ "0.513", "0.5186", "0.8833", "0.5186", "0.8833", "0.532", "0.513", "0.532" ], [ "0.513", "0.5337", "0.8833", "0.5337", "0.8833", "0.5474", "0.513", "0.5474" ], [ "0.1184", "0.5405", "0.4895", "0.5405", "0.4895", "0.554", "0.1184", "0.554" ], [ "0.11755", "0.554", "0.4878", "0.554", "0.4878", "0.568", "0.11755", "0.568" ], [ "0.5146", "0.552", "0.623", "0.552", "0.623", "0.5635", "0.5146", "0.5635" ], [ "0.11755", "0.5684", "0.4592", "0.5684", "0.4592", "0.582", "0.11755", "0.582" ], [ "0.529", "0.568", "0.882", "0.568", "0.882", "0.5806", "0.529", "0.5806" ], [ "0.513", "0.585", "0.8833", "0.585", "0.8833", "0.5986", "0.513", "0.5986" ], [ "0.5146", "0.601", "0.8833", "0.601", "0.8833", "0.6143", "0.5146", "0.6143" ], [ "0.1184", "0.6177", "0.4895", "0.6177", "0.4895", "0.6313", "0.1184", "0.6313" ], [ "0.513", "0.6177", "0.879", "0.6177", "0.879", "0.6274", "0.513", "0.6274" ], [ "0.1184", "0.6343", "0.4836", "0.6343", "0.4836", "0.648", "0.1184", "0.648" ], [ "0.5146", "0.633", "0.8833", "0.633", "0.8833", "0.6465", "0.5146", "0.6465" ], [ "0.511", "0.648", "0.885", "0.6484", "0.885", "0.663", "0.511", "0.662" ], [ "0.11755", "0.661", "0.3284", "0.663", "0.3284", "0.677", "0.11755", "0.675" ], [ "0.513", "0.665", "0.885", "0.665", "0.885", "0.6787", "0.513", "0.6787" ], [ "0.1184", "0.685", "0.4895", "0.685", "0.4895", "0.6987", "0.1184", "0.6987" ], [ "0.513", "0.681", "0.684", "0.681", "0.684", "0.695", "0.513", "0.695" ], [ "0.11755", "0.7007", "0.4895", "0.6997", "0.4895", "0.7144", "0.11755", "0.7153" ], [ "0.1184", "0.7163", "0.4895", "0.7163", "0.4895", "0.73", "0.1184", "0.73" ], [ "0.5156", "0.7163", "0.5557", "0.7163", "0.5557", "0.728", "0.5156", "0.728" ], [ "0.552", "0.7153", "0.711", "0.7153", "0.711", "0.7285", "0.552", "0.7285" ], [ "0.1184", "0.734", "0.486", "0.734", "0.486", "0.7446", "0.1184", "0.7446" ], [ "0.5146", "0.7407", "0.885", "0.7407", "0.885", "0.7544", "0.5146", "0.7544" ], [ "0.11755", "0.75", "0.4895", "0.75", "0.4895", "0.763", "0.11755", "0.763" ], [ "0.511", "0.756", "0.8833", "0.756", "0.8833", "0.7695", "0.511", "0.7695" ], [ "0.1184", "0.767", "0.4878", "0.767", "0.4878", "0.778", "0.1184", "0.778" ], [ "0.513", "0.774", "0.885", "0.774", "0.885", "0.7876", "0.513", "0.7876" ], [ "0.1184", "0.782", "0.4802", "0.782", "0.4802", "0.796", "0.1184", "0.796" ], [ "0.513", "0.789", "0.885", "0.789", "0.885", "0.8027", "0.513", "0.8027" ], [ "0.1368", "0.799", "0.4902", "0.799", "0.4902", "0.8125", "0.1368", "0.8125" ], [ "0.513", "0.8057", "0.885", "0.8057", "0.885", "0.8193", "0.513", "0.8193" ], [ "0.1184", "0.814", "0.4895", "0.814", "0.4895", "0.8276", "0.1184", "0.8276" ], [ "0.513", "0.822", "0.885", "0.822", "0.885", "0.835", "0.513", "0.835" ], [ "0.1184", "0.8306", "0.4902", "0.8306", "0.4902", "0.8457", "0.1184", "0.8457" ], [ "0.511", "0.8384", "0.8833", "0.8384", "0.8833", "0.8516", "0.511", "0.8516" ], [ "0.1184", "0.847", "0.4902", "0.847", "0.4902", "0.861", "0.1184", "0.861" ], [ "0.11755", "0.8633", "0.4902", "0.8633", "0.4902", "0.877", "0.11755", "0.877" ], [ "0.531", "0.868", "0.882", "0.869", "0.882", "0.8843", "0.531", "0.883" ], [ "0.11755", "0.88", "0.4895", "0.879", "0.4895", "0.8926", "0.11755", "0.8936" ], [ "0.5356", "0.883", "0.8716", "0.883", "0.8716", "0.8965", "0.5356", "0.8965" ], [ "0.11755", "0.8945", "0.4878", "0.8945", "0.4878", "0.9077", "0.11755", "0.9077" ], [ "0.534", "0.8945", "0.8228", "0.8955", "0.8228", "0.909", "0.534", "0.9077" ] ]

[ "DE-→EN", "DE-RM", "in-domain", "average OOD", "in-domain", "average OOD", "58.4", "system", "11.8", "45.2", "15.5", "61.5", "SMT (Muller et al., 2019)", "11.7", "52.5", "18.9", "13.1", "NMT (Mulleret al.,2019)", "NMT+RC+SR+NC (Muller et al., 2019)", "52.4", "20.7", "MLE w/o LS", "60.8", "58.3 (±0.53)", "9.7 (±0.25)", "52.2 (±0.19)", "15.8 (±0.39)", "+MRT", "10.2(±0.26)", "52.1.(±0.08)", "15.9 (±0.28)", "112 (±0.16)", "58.4 (±0.39)", "MLE w/ LS", "58.9 (±0.45)", "53.9 (±0.16)", "18.0(±0.17)", "+MRT", "58.8 (±0.36)", "12.0 (±0.29)", "53.9 (±0.12)", "18.7 (±0.09)", "Table 3: Average BLEU and standard deviation on in-domain and out-of-domain test sets for models trained on", "OPUS (DE-→EN) and Allegra (DE-→RM).RC: reconstruction; SR: subword regularization, NC: noisy channel.", "% hallucinations (BLEU)", "baseline assigns similar probabilities to tokens in", "in-domain", "the references and the distractors. Only for the first", "out-of-domain", "system", "MLE w/oLS", "2% (58.3)", "time stepsis there a clearpreferencefortherefer-", "29% (10.2)", "35%(9.7)", "+MRT", "ences over the (mostly random) distractors. This", "MLE w/ LS", "33% (11.2)", "1% (58.9)", "26% (12.0)", "shows that error propagation is a big risk: should", "+MRT", "the model make a wrong prediction initially, this is", "Table 4: Proportion of hallucinations and BLEU on out-", "unlikely to be penalised in later time steps.", "of-domain and in-domain test sets. DE→EN OPUS.", "MRT tends to increase the model's certainty at", "later time steps, but importantly, the increase is", "Wir haben ihn gefunden.", "sharper for the reference translations than for the", "We found him.", "source", "reference", "distractors. The direct comparison shows a widen-", "MLE", "Do not pass it.", "MRT", "ing gap in certainty between the reference and dis-", "We have found it.", "tractor sentences.7 In other words, producing a", "source", "So hore nicht auf die Ableugner.", "So hearken not to those who deny (the Truth).", "hallucination will incur a small penaltyat each", "Do not drive or use machines.", "reference", "MLE", "time step (compared to producing the reference),", "MRT", "Do not apply to dleugner.", "presumably due to a higher reliance on the source", "signal, lessening the risk of error propagation and", "Table 5: Out-of-domain translation examples. MLE", "hallucinates in both examples; MRT was rated more", "hallucinations.", "adequate in top example, less fluent in bottom one", "Our analysis shows similar trends on in-domain", "references. However, much higher probabilities are", "assigned to the first few tokens of the references", "mistranslated into dleugner'.We consider this", "than to the distractors. Hence. it is much less likely", "preferable to producing a complete hallucination.", "that a hallucination is kept in the beam, or will", "overtake a good translation in overall probability,", "4.2UncertaintyAnalysis", "reducing the practical impact of the model's over-", "Inspired by Ott et al. (2018), we analyse the", "reliance on its history.8", "model's uncertainty by computing the average prob-", "ability at each time step across a set of sentences", "4.3", "Beam Size Analysis", "Besides thereference translations.we also consider", "Figure 1 shows that with MLE, distractor sentences", "a set of 'distractor translations, which are random", "are assigned lower probabilities than the references", "sentencesfromthein-domaintestsetwhichmatch", "at the first few time steps, but are assigned similar,", "the corresponding reference translation in length.", "potentially even higher probabilities at later time", "In Figure I, we show out-of-domain results for", "steps. This establishes a connection between ex-", "an MLE model and multiple checkpoints of MRT", "posure bias and the beam search problem, i.e. the", "fine-tuning. The left two graphs show probabil-", "problem that increasing the search space can lead", "ities for references and distractors, respectively", "The right-most graph shows a direct comparison of", "6The uncertainty of the baseline is due to label smoothing.", "probabilities for references and distractors for the", "'For intermediate checkpoints, see Appendix, Figure 2.", "MLE baseline and the final MRT model. The MLE", "8Figures are shown in the Appendix (Figure 3)." ]

{ "arxiv_id": "2005.03642", "language": "en", "page_idx": 3 }

[ [ "0.2612", "0.0736", "0.3132", "0.0736", "0.3132", "0.08435", "0.2612", "0.08435" ], [ "0.4895", "0.0736", "0.538", "0.0736", "0.538", "0.08435", "0.4895", "0.08435" ], [ "0.6943", "0.0736", "0.7866", "0.0736", "0.7866", "0.08435", "0.6943", "0.08435" ], [ "0.1587", "0.0903", "0.168", "0.0903", "0.168", "0.0956", "0.1587", "0.0956" ], [ "0.4248", "0.0861", "0.4534", "0.0861", "0.4534", "0.0944", "0.4248", "0.0944" ], [ "0.409", "0.0932", "0.509", "0.0932", "0.509", "0.10394", "0.409", "0.10394" ], [ "0.157", "0.10394", "0.1704", "0.10394", "0.1704", "0.114", "0.157", "0.114" ], [ "0.4307", "0.10156", "0.5073", "0.1027", "0.5073", "0.1134", "0.4307", "0.11224" ], [ "0.1486", "0.1134", "0.1545", "0.1134", "0.1545", "0.12054", "0.1486", "0.12054" ], [ "0.434", "0.114", "0.4954", "0.114", "0.4954", "0.1217", "0.434", "0.1217" ], [ "0.6885", "0.1663", "0.7573", "0.1663", "0.7573", "0.177", "0.6885", "0.177" ], [ "0.1453", "0.11816", "0.1721", "0.11816", "0.1721", "0.1312", "0.1453", "0.1312" ], [ "0.147", "0.1687", "0.168", "0.1687", "0.168", "0.1812", "0.147", "0.1812" ], [ "0.69", "0.177", "0.756", "0.177", "0.756", "0.1841", "0.69", "0.1841" ], [ "0.157", "0.1853", "0.1696", "0.1853", "0.1696", "0.1948", "0.157", "0.1948" ], [ "0.6885", "0.1853", "0.8154", "0.1853", "0.8154", "0.1953", "0.6885", "0.1953" ], [ "0.157", "0.1996", "0.1738", "0.1996", "0.1738", "0.209", "0.157", "0.209" ], [ "0.687", "0.1953", "0.812", "0.1953", "0.812", "0.2031", "0.687", "0.2031" ], [ "0.287", "0.2114", "0.339", "0.2114", "0.339", "0.2156", "0.287", "0.2156" ], [ "0.513", "0.2114", "0.5356", "0.2114", "0.5356", "0.2156", "0.513", "0.2156" ], [ "0.5356", "0.2114", "0.5474", "0.2114", "0.5474", "0.2156", "0.5356", "0.2156" ], [ "0.763", "0.2114", "0.7734", "0.2114", "0.7734", "0.2156", "0.763", "0.2156" ], [ "0.7773", "0.212", "0.797", "0.212", "0.797", "0.2162", "0.7773", "0.2162" ], [ "0.8037", "0.2114", "0.8345", "0.2114", "0.8345", "0.2156", "0.8037", "0.2156" ], [ "0.266", "0.2185", "0.3074", "0.2185", "0.3074", "0.2257", "0.266", "0.2257" ], [ "0.4963", "0.2197", "0.532", "0.2197", "0.532", "0.2239", "0.4963", "0.2239" ], [ "0.721", "0.2185", "0.7617", "0.2185", "0.7617", "0.2257", "0.721", "0.2257" ], [ "0.1184", "0.2382", "0.882", "0.2382", "0.882", "0.2517", "0.1184", "0.2517" ], [ "0.1184", "0.253", "0.881", "0.253", "0.881", "0.2666", "0.1184", "0.2666" ], [ "0.1184", "0.2922", "0.4902", "0.2922", "0.4902", "0.3071", "0.1184", "0.3071" ], [ "0.511", "0.2915", "0.6514", "0.2922", "0.6514", "0.3071", "0.511", "0.306" ], [ "0.12006", "0.31", "0.4902", "0.31", "0.4902", "0.3237", "0.12006", "0.3237" ], [ "0.513", "0.3184", "0.882", "0.3184", "0.882", "0.3289", "0.513", "0.3289" ], [ "0.1184", "0.326", "0.4895", "0.326", "0.4895", "0.339", "0.1184", "0.339" ], [ "0.511", "0.3333", "0.882", "0.3333", "0.882", "0.3467", "0.511", "0.3467" ], [ "0.1184", "0.3403", "0.4895", "0.3403", "0.4895", "0.3552", "0.1184", "0.3552" ], [ "0.513", "0.3508", "0.8833", "0.3508", "0.8833", "0.364", "0.513", "0.364" ], [ "0.1184", "0.3582", "0.4895", "0.3582", "0.4895", "0.3718", "0.1184", "0.3718" ], [ "0.511", "0.3665", "0.8833", "0.3665", "0.8833", "0.3801", "0.511", "0.3801" ], [ "0.12006", "0.3748", "0.2032", "0.3748", "0.2032", "0.386", "0.12006", "0.386" ], [ "0.5146", "0.3843", "0.885", "0.3843", "0.885", "0.3943", "0.5146", "0.3943" ], [ "0.1351", "0.3914", "0.4895", "0.3926", "0.4895", "0.4067", "0.1351", "0.4062" ], [ "0.513", "0.3984", "0.885", "0.3984", "0.885", "0.412", "0.513", "0.412" ], [ "0.1184", "0.4092", "0.4895", "0.4092", "0.4895", "0.4233", "0.1184", "0.4233" ], [ "0.513", "0.415", "0.885", "0.415", "0.885", "0.4287", "0.513", "0.4287" ], [ "0.12006", "0.4258", "0.4895", "0.4258", "0.4895", "0.4407", "0.12006", "0.4407" ], [ "0.513", "0.4312", "0.885", "0.4312", "0.885", "0.4448", "0.513", "0.4448" ], [ "0.1184", "0.4412", "0.4902", "0.4412", "0.4902", "0.4548", "0.1184", "0.4548" ], [ "0.511", "0.4453", "0.8867", "0.4478", "0.8867", "0.4614", "0.511", "0.459" ], [ "0.12006", "0.4578", "0.4902", "0.4578", "0.4902", "0.4714", "0.12006", "0.4714" ], [ "0.511", "0.4631", "0.885", "0.4631", "0.885", "0.4768", "0.511", "0.4768" ], [ "0.1184", "0.4739", "0.4895", "0.4727", "0.4895", "0.4868", "0.1184", "0.488" ], [ "0.513", "0.4785", "0.885", "0.4785", "0.885", "0.4922", "0.513", "0.4922" ], [ "0.1184", "0.4905", "0.4895", "0.4905", "0.4895", "0.5044", "0.1184", "0.5044" ], [ "0.513", "0.4954", "0.885", "0.4954", "0.885", "0.509", "0.513", "0.509" ], [ "0.1184", "0.506", "0.4895", "0.506", "0.4895", "0.5195", "0.1184", "0.5195" ], [ "0.513", "0.5127", "0.785", "0.5127", "0.785", "0.5254", "0.513", "0.5254" ], [ "0.11755", "0.5215", "0.4895", "0.521", "0.4895", "0.535", "0.11755", "0.536" ], [ "0.531", "0.528", "0.8833", "0.528", "0.8833", "0.5415", "0.531", "0.5415" ], [ "0.1184", "0.538", "0.4902", "0.538", "0.4902", "0.552", "0.1184", "0.552" ], [ "0.513", "0.5435", "0.885", "0.5435", "0.885", "0.557", "0.513", "0.557" ], [ "0.1184", "0.5547", "0.4902", "0.5547", "0.4902", "0.5684", "0.1184", "0.5684" ], [ "0.5146", "0.56", "0.8833", "0.56", "0.8833", "0.5737", "0.5146", "0.5737" ], [ "0.1184", "0.571", "0.4902", "0.571", "0.4902", "0.5845", "0.1184", "0.5845" ], [ "0.5146", "0.576", "0.8833", "0.576", "0.8833", "0.589", "0.5146", "0.589" ], [ "0.1184", "0.586", "0.4902", "0.585", "0.4902", "0.5996", "0.1184", "0.601" ], [ "0.5146", "0.5933", "0.8833", "0.5933", "0.8833", "0.604", "0.5146", "0.604" ], [ "0.1184", "0.6016", "0.4902", "0.6025", "0.4902", "0.6177", "0.1184", "0.616" ], [ "0.513", "0.608", "0.885", "0.608", "0.885", "0.6216", "0.513", "0.6216" ], [ "0.12177", "0.6206", "0.4878", "0.6206", "0.4878", "0.6313", "0.12177", "0.6313" ], [ "0.513", "0.6235", "0.885", "0.6235", "0.885", "0.637", "0.513", "0.637" ], [ "0.1184", "0.6353", "0.4895", "0.6353", "0.4895", "0.6484", "0.1184", "0.6484" ], [ "0.513", "0.64", "0.8833", "0.64", "0.8833", "0.654", "0.513", "0.654" ], [ "0.1184", "0.652", "0.1814", "0.652", "0.1814", "0.663", "0.1184", "0.663" ], [ "0.513", "0.6562", "0.885", "0.6562", "0.885", "0.6694", "0.513", "0.6694" ], [ "0.513", "0.673", "0.8833", "0.673", "0.8833", "0.686", "0.513", "0.686" ], [ "0.2737", "0.6885", "0.434", "0.6885", "0.434", "0.7017", "0.2737", "0.7017" ], [ "0.511", "0.6885", "0.8833", "0.6885", "0.8833", "0.7017", "0.511", "0.7017" ], [ "0.2435", "0.707", "0.288", "0.707", "0.288", "0.7188", "0.2435", "0.7188" ], [ "0.3315", "0.707", "0.3762", "0.707", "0.3762", "0.7188", "0.3315", "0.7188" ], [ "0.4165", "0.707", "0.4668", "0.707", "0.4668", "0.7188", "0.4165", "0.7188" ], [ "0.513", "0.705", "0.885", "0.705", "0.885", "0.7188", "0.513", "0.7188" ], [ "0.511", "0.72", "0.6514", "0.7197", "0.6514", "0.734", "0.511", "0.7354" ], [ "0.1293", "0.708", "0.1763", "0.708", "0.1763", "0.7197", "0.1293", "0.7197" ], [ "0.1293", "0.7246", "0.2141", "0.7246", "0.2141", "0.7363", "0.1293", "0.7363" ], [ "0.236", "0.7256", "0.3015", "0.7256", "0.3015", "0.737", "0.236", "0.737" ], [ "0.3242", "0.7256", "0.388", "0.7256", "0.388", "0.737", "0.3242", "0.737" ], [ "0.4114", "0.7256", "0.4785", "0.7256", "0.4785", "0.7363", "0.4114", "0.7363" ], [ "0.1293", "0.737", "0.1755", "0.737", "0.1755", "0.749", "0.1293", "0.749" ], [ "0.236", "0.737", "0.3015", "0.737", "0.3015", "0.749", "0.236", "0.749" ], [ "0.318", "0.737", "0.3896", "0.737", "0.3896", "0.749", "0.318", "0.749" ], [ "0.4114", "0.737", "0.4768", "0.737", "0.4768", "0.749", "0.4114", "0.749" ], [ "0.1293", "0.75", "0.2065", "0.75", "0.2065", "0.761", "0.1293", "0.761" ], [ "0.2301", "0.75", "0.3032", "0.75", "0.3032", "0.761", "0.2301", "0.761" ], [ "0.318", "0.75", "0.3896", "0.75", "0.3896", "0.761", "0.318", "0.761" ], [ "0.4114", "0.75", "0.4768", "0.75", "0.4768", "0.761", "0.4114", "0.761" ], [ "0.5146", "0.7476", "0.6733", "0.7476", "0.6733", "0.7627", "0.5146", "0.7627" ], [ "0.1293", "0.7603", "0.1763", "0.7603", "0.1763", "0.775", "0.1293", "0.775" ], [ "0.2301", "0.761", "0.3015", "0.761", "0.3015", "0.7725", "0.2301", "0.7725" ], [ "0.3198", "0.761", "0.3896", "0.761", "0.3896", "0.7725", "0.3198", "0.7725" ], [ "0.4072", "0.761", "0.4768", "0.761", "0.4768", "0.7725", "0.4072", "0.7725" ], [ "0.513", "0.7715", "0.885", "0.7715", "0.885", "0.786", "0.513", "0.786" ], [ "0.11755", "0.788", "0.4902", "0.788", "0.4902", "0.802", "0.11755", "0.802" ], [ "0.511", "0.789", "0.8867", "0.788", "0.8867", "0.8027", "0.511", "0.804" ], [ "0.1184", "0.8027", "0.4878", "0.8027", "0.4878", "0.8164", "0.1184", "0.8164" ], [ "0.5146", "0.8047", "0.8867", "0.8047", "0.8867", "0.8193", "0.5146", "0.8193" ], [ "0.5146", "0.822", "0.885", "0.822", "0.885", "0.835", "0.5146", "0.835" ], [ "0.5146", "0.8374", "0.8867", "0.8374", "0.8867", "0.8516", "0.5146", "0.8516" ], [ "0.1377", "0.858", "0.4895", "0.8594", "0.4895", "0.8735", "0.1377", "0.8726" ], [ "0.5146", "0.855", "0.8833", "0.855", "0.8833", "0.868", "0.5146", "0.868" ], [ "0.11755", "0.8716", "0.4902", "0.8716", "0.4902", "0.8853", "0.11755", "0.8853" ], [ "0.513", "0.868", "0.8833", "0.8706", "0.8833", "0.8853", "0.513", "0.883" ], [ "0.1184", "0.8843", "0.4902", "0.8843", "0.4902", "0.898", "0.1184", "0.898" ], [ "0.5146", "0.8857", "0.681", "0.8857", "0.681", "0.8994", "0.5146", "0.8994" ], [ "0.12006", "0.898", "0.486", "0.898", "0.486", "0.9077", "0.12006", "0.9077" ] ]

[ "Reference", "Distractor", "Direct comparision", "0.", "MLE", "MRT100updates", "0.6", "MRT500updates", "1", "MRT3000upd", "MLE (reference)", "00.5", "00", "MLE(distractor)", "0.1", "MRT3000 updates (reference)", "0.0", "MRT3000updates(distractor", "10:121", "400", "12", "12", "14", "161", "timestep", "imeste", "cimestep", "Figure 1: Per-token probability of out-of-domain reference translations and in-domain distractors (first two graphs", "share legend). Rightmost plot shows direct comparison for MLE baseline and final MRT model. DE→EN OPUS", "to worse model performance.9 With larger beam", "5 Conclusions", "size, it is more likely that hallucinations survive", "Ourresults and analysis.showaconnectionbe", "pruning at the first few time steps, and with high", "tween the exposure bias due to MLE training with", "probabilities assigned to them at later time steps,", "teacher forcing and several well-known problems", "there is a chance that they become the top-scoring", "in neural machine translation, namely poor per", "translation.", "formanceunderdomainshift,hallucinatedtransla-", "We investigate whether the beam search problem", "tions, and deteriorating performance with increas-", "is mitigated by MRT. In Table 6, we report OOD", "ing beam size. We find that Minimum Risk Train-", "BLEU and the proportion of hallucinations with", "ing, which does not suffer from exposure bias, can", "beam sizes of 1, 4 and 50. While MRT does not", "be useful even when it does not increase perfor-", "eliminate the beam search problem, performance", "mance on an in-domain test set: it increases per-", "drops less steeply as beam size increases. With", "formance under domain shift, reduces the number", "beam size 4, our MRT models outperform the MLE", "of hallucinations substantially, and makes beam", "baseline by 0.5-0.8 BLEU; with beam size 50, this", "search with large beams more stable.", "difference grows to 0.6-1.5 BLEU. Our manual", "Our findings are pertinent to the academic de-", "evaluation (N=200 for each system for beam size 1", "bate how big of a problem exposure bias is in prac-", "and 50) shows that the proportion of hallucinations", "tice - we find that this can vary substantially de", "increases with beam size, and that MRT consis-", "pending on the dataset -, and they provide a new", "tently reduces the proportion by 11-21% (relative).", "justificationforsequence-leveltrainingobjectives", "For the system with label smoothing, the relative in-", "that reduce or eliminate exposure bias. Further-", "creaseinhallucinationswithincreasingbeamsize", "more, we believe that a better understanding of the", "is also smaller with MRT (+33%) than with MLE", "links between exposure bias and well-known trans-", "(+44%).", "lationproblems will helppractitioners decide when", "sequence-level training objectives are especially", "BLEU (% hallucinations)", "promising, for example in settings where the test", "k=1", "k=4", "k= 50", "domain is unknown, or where hallucinations are a", "common problem.", "system", "MLE w/o LS", "8.9 (28%)", "9.7 (35%)", "9.3 (37%)", "+MRT", "9.1 (24%)", "10.2 (29%)", "9.9 (33%)", "MLE w/ LS", "10.6 (27%)", "11.2 (33%)", "9.4 (39%)", "Acknowledgments", "+MRT", "11.3 (24%)", "12.0 (26%)", "10.9 (32%)", "Chaojun Wang was supported by the UK Engineer-", "Table 6: Average OOD BLEU and proportion of hallu-", "ing and Physical Sciences Research Council (EP-", "cinations with different beam sizes k. DE-→EN OPUS.", "SRC) fellowship grant EP/S001271/1 (MTStretch).", "Rico Sennrich acknowledges support of the Swiss", "National Science Foundation (MUTAMUR; no.", "The beam search problem has previously been linked to", "176727). This project has received support from", "length bias (Yang et al., 2018; Murray and Chiang, 2018) and", "Samsung Electronics Polska sp. z o.o. - Samsung", "the copy mode (Ott et al., 2018). We consider hallucinations", "R&D Institute Poland.", "anotherresultofusinglargesearchspaceswithMLEmodels." ]

{ "arxiv_id": "2005.03642", "language": "en", "page_idx": 4 }

[ [ "0.3284", "0.1206", "0.6445", "0.1206", "0.6445", "0.1351", "0.3284", "0.1351" ], [ "0.2484", "0.1686", "0.3816", "0.1692", "0.3816", "0.1843", "0.2484", "0.1831" ], [ "0.4207", "0.1686", "0.5933", "0.1692", "0.5933", "0.1843", "0.4207", "0.1831" ], [ "0.6406", "0.1692", "0.722", "0.1692", "0.722", "0.1843", "0.6406", "0.1843" ], [ "0.3325", "0.1895", "0.6396", "0.1895", "0.6396", "0.2026", "0.3325", "0.2026" ], [ "0.3022", "0.2052", "0.6665", "0.2058", "0.6665", "0.2203", "0.3022", "0.2197" ], [ "0.2378", "0.2556", "0.3137", "0.2556", "0.3137", "0.2708", "0.2378", "0.2708" ], [ "0.02124", "0.2632", "0.06046", "0.2632", "0.06046", "0.3674", "0.02124", "0.3674" ], [ "0.0825", "0.2874", "0.468", "0.2874", "0.468", "0.3018", "0.0825", "0.3018" ], [ "0.0809", "0.3037", "0.468", "0.3037", "0.468", "0.3176", "0.0809", "0.3176" ], [ "0.0809", "0.318", "0.4673", "0.318", "0.4673", "0.332", "0.0809", "0.332" ], [ "0.0801", "0.334", "0.468", "0.334", "0.468", "0.348", "0.0801", "0.348" ], [ "0.0801", "0.3484", "0.468", "0.3484", "0.468", "0.3623", "0.0801", "0.3623" ], [ "0.02533", "0.3623", "0.058", "0.3623", "0.058", "0.3965", "0.02533", "0.3965" ], [ "0.03595", "0.361", "0.04575", "0.361", "0.04575", "0.3662", "0.03595", "0.3662" ], [ "0.0801", "0.3643", "0.468", "0.3643", "0.468", "0.3774", "0.0801", "0.3774" ], [ "0.818", "0.3655", "0.8384", "0.3655", "0.8384", "0.3726", "0.818", "0.3726" ], [ "0.0801", "0.379", "0.468", "0.379", "0.468", "0.3926", "0.0801", "0.3926" ], [ "0.0809", "0.3933", "0.468", "0.3933", "0.468", "0.4072", "0.0809", "0.4072" ], [ "0.02533", "0.406", "0.05554", "0.406", "0.05554", "0.4507", "0.02533", "0.4507" ], [ "0.0801", "0.4092", "0.4673", "0.4092", "0.4673", "0.4224", "0.0801", "0.4224" ], [ "0.0801", "0.4248", "0.4697", "0.4248", "0.4697", "0.4382", "0.0801", "0.4382" ], [ "0.5034", "0.4204", "0.892", "0.4204", "0.892", "0.4343", "0.5034", "0.4343" ], [ "0.0784", "0.4395", "0.468", "0.4395", "0.468", "0.4526", "0.0784", "0.4526" ], [ "0.5024", "0.4363", "0.8896", "0.4343", "0.8896", "0.4475", "0.5024", "0.4495" ], [ "0.03186", "0.4475", "0.058", "0.4475", "0.058", "0.481", "0.03186", "0.481" ], [ "0.0784", "0.4526", "0.468", "0.4539", "0.468", "0.4685", "0.0784", "0.4673" ], [ "0.5024", "0.449", "0.892", "0.449", "0.892", "0.4622", "0.5024", "0.4622" ], [ "0.5024", "0.4622", "0.892", "0.4622", "0.892", "0.476", "0.5024", "0.476" ], [ "0.0809", "0.4697", "0.4673", "0.4697", "0.4673", "0.483", "0.0809", "0.483" ], [ "0.5024", "0.4766", "0.8594", "0.4766", "0.8594", "0.4905", "0.5024", "0.4905" ], [ "0.0801", "0.4849", "0.4673", "0.4849", "0.4673", "0.4988", "0.0801", "0.4988" ], [ "0.02287", "0.4956", "0.06046", "0.4956", "0.06046", "0.64", "0.02287", "0.64" ], [ "0.0801", "0.5", "0.468", "0.5", "0.468", "0.513", "0.0801", "0.513" ], [ "0.5024", "0.509", "0.892", "0.5103", "0.892", "0.5244", "0.5024", "0.5234" ], [ "0.0801", "0.515", "0.4656", "0.515", "0.4656", "0.529", "0.0801", "0.529" ], [ "0.0801", "0.53", "0.4697", "0.53", "0.4697", "0.5444", "0.0801", "0.5444" ], [ "0.5024", "0.5264", "0.8896", "0.5264", "0.8896", "0.5405", "0.5024", "0.5405" ], [ "0.0801", "0.5454", "0.2852", "0.5454", "0.2852", "0.5596", "0.0801", "0.5596" ], [ "0.5034", "0.5415", "0.892", "0.5415", "0.892", "0.5547", "0.5034", "0.5547" ], [ "0.5034", "0.558", "0.8896", "0.558", "0.8896", "0.5684", "0.5034", "0.5684" ], [ "0.5034", "0.5713", "0.892", "0.5713", "0.892", "0.5854", "0.5034", "0.5854" ], [ "0.0825", "0.589", "0.2075", "0.589", "0.2075", "0.604", "0.0825", "0.604" ], [ "0.5034", "0.586", "0.8906", "0.586", "0.8906", "0.5996", "0.5034", "0.5996" ], [ "0.5034", "0.6016", "0.892", "0.6016", "0.892", "0.6157", "0.5034", "0.6157" ], [ "0.0801", "0.6143", "0.4697", "0.6143", "0.4697", "0.6274", "0.0801", "0.6274" ], [ "0.5034", "0.616", "0.8906", "0.616", "0.8906", "0.63", "0.5034", "0.63" ], [ "0.0801", "0.63", "0.468", "0.63", "0.468", "0.643", "0.0801", "0.643" ], [ "0.5024", "0.631", "0.8906", "0.632", "0.8906", "0.646", "0.5024", "0.6445" ], [ "0.501", "0.646", "0.7134", "0.6465", "0.7134", "0.661", "0.501", "0.66" ], [ "0.02287", "0.634", "0.05637", "0.6333", "0.05963", "0.709", "0.02533", "0.7104" ], [ "0.0809", "0.6465", "0.4673", "0.6465", "0.4673", "0.657", "0.0809", "0.657" ], [ "0.0801", "0.66", "0.468", "0.66", "0.468", "0.674", "0.0801", "0.674" ], [ "0.5195", "0.662", "0.8105", "0.6636", "0.8105", "0.678", "0.5195", "0.677" ], [ "0.0801", "0.6753", "0.468", "0.6753", "0.468", "0.6895", "0.0801", "0.6895" ], [ "0.5", "0.6816", "0.8906", "0.6816", "0.8906", "0.696", "0.5", "0.696" ], [ "0.0801", "0.69", "0.4697", "0.69", "0.4697", "0.704", "0.0801", "0.704" ], [ "0.5073", "0.6978", "0.7793", "0.6978", "0.7793", "0.711", "0.5073", "0.711" ], [ "0.0801", "0.7056", "0.468", "0.705", "0.468", "0.7183", "0.0801", "0.7197" ], [ "0.0784", "0.72", "0.468", "0.7197", "0.468", "0.7344", "0.0784", "0.735" ], [ "0.5", "0.7183", "0.8896", "0.7183", "0.8896", "0.729", "0.5", "0.729" ], [ "0.0825", "0.7373", "0.4656", "0.7373", "0.4656", "0.7476", "0.0825", "0.7476" ], [ "0.509", "0.732", "0.892", "0.732", "0.892", "0.747", "0.509", "0.747" ], [ "0.0801", "0.7505", "0.468", "0.7505", "0.468", "0.7646", "0.0801", "0.7646" ], [ "0.509", "0.7476", "0.7915", "0.7476", "0.7915", "0.761", "0.509", "0.761" ], [ "0.0809", "0.7676", "0.4673", "0.7676", "0.4673", "0.778", "0.0809", "0.778" ], [ "0.4983", "0.7676", "0.8906", "0.7676", "0.8906", "0.781", "0.4983", "0.781" ], [ "0.0801", "0.781", "0.468", "0.781", "0.468", "0.7944", "0.0801", "0.7944" ], [ "0.5073", "0.7827", "0.8906", "0.781", "0.8906", "0.7944", "0.5073", "0.797" ], [ "0.0801", "0.7954", "0.468", "0.7954", "0.468", "0.8096", "0.0801", "0.8096" ], [ "0.509", "0.7974", "0.7466", "0.7974", "0.7466", "0.811", "0.509", "0.811" ], [ "0.0809", "0.8125", "0.4673", "0.8125", "0.4673", "0.8228", "0.0809", "0.8228" ], [ "0.4983", "0.8174", "0.8906", "0.8174", "0.8906", "0.8306", "0.4983", "0.8306" ], [ "0.0801", "0.827", "0.4714", "0.827", "0.4714", "0.8403", "0.0801", "0.8403" ], [ "0.5073", "0.8335", "0.8555", "0.8335", "0.8555", "0.8467", "0.5073", "0.8467" ], [ "0.0801", "0.8423", "0.1912", "0.8413", "0.1912", "0.853", "0.0801", "0.854" ], [ "0.0972", "0.857", "0.4714", "0.857", "0.4714", "0.8706", "0.0972", "0.8706" ], [ "0.4968", "0.851", "0.8906", "0.8516", "0.8906", "0.866", "0.4968", "0.8657" ], [ "0.0801", "0.872", "0.468", "0.872", "0.468", "0.886", "0.0801", "0.886" ], [ "0.509", "0.8687", "0.892", "0.8687", "0.892", "0.882", "0.509", "0.882" ], [ "0.0801", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0801", "0.901" ], [ "0.509", "0.8833", "0.777", "0.8833", "0.777", "0.8965", "0.509", "0.8965" ], [ "0.4836", "0.928", "0.4878", "0.928", "0.4878", "0.933", "0.4836", "0.933" ] ]

[ "Geometric Context from Videos", "S. Hussain Raza", "MatthiasGrundmann", "Irfan Essa", "Georgia Institute of Technology, Atlanta, GA, USA", "http://www.cc.gatech.edu/cpl/projects/videogeometriccontext", "Abstract", "S102 100", "We present a novel algorithm for estimating the broad 3D", "geometric structure of outdoor video scenes. Leveraging", "spatio-temporal video segmentation, we decompose a dy-", "namic scene captured by a video into geometric classes,", "based on predictions made by region-classifiers that are", "5", "", "trained on appearance and motion features. By examining", "Mix", "the homogeneity of the prediction, we combine predictions", "across multiple segmentation hierarchy levels alleviating", "LAD", "the need to determine the granularity a priori.We built", "a novel, extensive dataset on geometric context of video", "Figure 1: Video frames of an urban scene (left) and the predicted", "to evaluate our method, consisting of over 1o0 ground-", "geometric context of our system (right). We achieve high accu-", "S", "truth annotated outdoor videos with over 20,000 frames.", "racy leveraging motion and appearance features while achieving", "temporal consistency by relying on spatio-temporal regions across", "To further scale beyond this dataset, we propose a semi-", "various granularities. Please watch the accompanying video.", "supervised learning framework to expand the pool of la-", "", "beled data with high confidence predictions obtained from", "Building upon a hierarchical video-segmentation to achieve", "unlabeled data. Our system produces an accurate predic-", "tion of geometric context of video achieving 96% accuracy", "temporal consistency, we compute a wide variety of appear-", "across main geometric classes.", "ance, location, and motion features which are used to train", "classifierstopredictgeometriccontextin video.A signif-", "icant challenge for developing scene understanding system", "1. Introduction", "for videos is a need for an annotated video dataset available", "for training and evaluation. To this end, we have collected", "Holistic scene understanding requires an understanding of", "and annotated a video dataset with pixel level ground truth", "the broad 3D structure of the scene with all objects present.", "labels for over 20,000 frames across 100 videos covering a", "wide variety of scene examples.", "L:AIXI", "Oneimportant step towards this goal is topartition a scene", "into regions and label them relative to each other and within", "The primary contributions of this paper are:", "the scene geometry. Geometric classes can define the ba-", "·A scene description for video via geometric classes (96%", "sic 3D structure of a scene with respect to the camera, and", "accuracy across main geometric classes).", "suggest cues to identify horizontal surfaces and vertical ob-", "jects in the scene. Hoeim et al. [12] showed that such ge-", ".Exploitingmotionand temporalcausality/redundancy", "ometric context can be used to obtain a probabilistic rep-", "present in video by using motion features and aggregating", "resentation of the scene layout based on geometric classes,", "predictions across spatio-temporal regions.", "which in turn can be used to improve objectdetection.Tor-", "·A semi-supervised bootstrap learning framework for ex-", "ralba et al. [18] showed that global context plays an im-", "panding the pool of labeled data with highly confident pre-", "portant role in object detection. Recently, Divala et al. [5]", "dictions obtained on unlabeled data.", "showed that incorporating geometric context.not onlyim-", ".A novel dataset of 100 annotated videos (~20,000 frames)", "proves object detection but also makes misclassifications", "with pixel level labels, which will be made available.", "morereasonable", "In this paper, we propose a novel method to provide a", ".A thorough evaluation of our system by examining impor-", "high level description of a video scene by assigning geomet-", "tance of features, benefit of temporal redundancy and in-", "ric classes to spatio-temporal regions as shown in Figure 1.", "dependence of segmentation granularity.", "-" ]

{ "arxiv_id": "1510.07320", "language": "en", "page_idx": 0 }

[ [ "0.0801", "0.0903", "0.2181", "0.0903", "0.2181", "0.1048", "0.0801", "0.1048" ], [ "0.5024", "0.09155", "0.892", "0.09155", "0.892", "0.1061", "0.5024", "0.1061" ], [ "0.501", "0.1061", "0.564", "0.108", "0.564", "0.1225", "0.501", "0.1206" ], [ "0.0801", "0.11615", "0.468", "0.11615", "0.468", "0.13", "0.0801", "0.13" ], [ "0.0801", "0.1307", "0.4673", "0.1307", "0.4673", "0.1445", "0.0801", "0.1445" ], [ "0.5024", "0.1351", "0.7837", "0.1351", "0.7837", "0.1497", "0.5024", "0.1497" ], [ "0.0801", "0.1465", "0.4697", "0.1465", "0.4697", "0.1604", "0.0801", "0.1604" ], [ "0.0801", "0.1622", "0.4697", "0.1622", "0.4697", "0.1755", "0.0801", "0.1755" ], [ "0.5034", "0.1604", "0.892", "0.1604", "0.892", "0.1736", "0.5034", "0.1736" ], [ "0.0809", "0.1787", "0.4673", "0.1787", "0.4673", "0.1895", "0.0809", "0.1895" ], [ "0.5024", "0.1755", "0.892", "0.1755", "0.892", "0.1901", "0.5024", "0.1901" ], [ "0.0784", "0.1925", "0.4697", "0.1913", "0.4697", "0.2052", "0.0784", "0.2058" ], [ "0.501", "0.1895", "0.8906", "0.1901", "0.8906", "0.2052", "0.501", "0.2039" ], [ "0.0801", "0.207", "0.4697", "0.207", "0.4697", "0.2203", "0.0801", "0.2203" ], [ "0.501", "0.207", "0.892", "0.207", "0.892", "0.2203", "0.501", "0.2203" ], [ "0.0801", "0.2229", "0.4722", "0.2229", "0.4722", "0.2361", "0.0801", "0.2361" ], [ "0.501", "0.2216", "0.892", "0.2216", "0.892", "0.2349", "0.501", "0.2349" ], [ "0.0801", "0.2374", "0.4697", "0.2374", "0.4697", "0.2507", "0.0801", "0.2507" ], [ "0.5034", "0.2374", "0.8906", "0.2374", "0.8906", "0.2474", "0.5034", "0.2474" ], [ "0.0801", "0.252", "0.4697", "0.252", "0.4697", "0.2651", "0.0801", "0.2651" ], [ "0.5034", "0.2524", "0.888", "0.2524", "0.888", "0.2632", "0.5034", "0.2632" ], [ "0.0801", "0.2676", "0.2361", "0.2676", "0.2361", "0.281", "0.0801", "0.281" ], [ "0.5034", "0.2664", "0.8896", "0.2664", "0.8896", "0.2798", "0.5034", "0.2798" ], [ "0.10046", "0.2822", "0.468", "0.2822", "0.468", "0.2954", "0.10046", "0.2954" ], [ "0.5034", "0.2827", "0.8906", "0.2827", "0.8906", "0.2935", "0.5034", "0.2935" ], [ "0.0801", "0.2974", "0.4697", "0.2974", "0.4697", "0.3113", "0.0801", "0.3113" ], [ "0.5024", "0.2966", "0.892", "0.2966", "0.892", "0.31", "0.5024", "0.31" ], [ "0.0801", "0.3132", "0.4697", "0.3132", "0.4697", "0.327", "0.0801", "0.327" ], [ "0.5024", "0.3113", "0.8906", "0.3113", "0.8906", "0.3245", "0.5024", "0.3245" ], [ "0.0801", "0.3276", "0.4697", "0.3276", "0.4697", "0.3416", "0.0801", "0.3416" ], [ "0.5024", "0.327", "0.892", "0.327", "0.892", "0.3403", "0.5024", "0.3403" ], [ "0.0801", "0.3435", "0.468", "0.3435", "0.468", "0.3567", "0.0801", "0.3567" ], [ "0.5024", "0.3416", "0.892", "0.3416", "0.892", "0.3547", "0.5024", "0.3547" ], [ "0.0801", "0.358", "0.468", "0.358", "0.468", "0.3718", "0.0801", "0.3718" ], [ "0.5024", "0.3567", "0.892", "0.3567", "0.892", "0.3706", "0.5024", "0.3706" ], [ "0.0784", "0.3726", "0.4673", "0.3726", "0.4673", "0.3865", "0.0784", "0.3865" ], [ "0.5024", "0.3726", "0.8936", "0.3726", "0.8936", "0.3865", "0.5024", "0.3865" ], [ "0.0784", "0.3882", "0.468", "0.3882", "0.468", "0.4016", "0.0784", "0.4016" ], [ "0.5024", "0.387", "0.63", "0.387", "0.63", "0.401", "0.5024", "0.401" ], [ "0.0801", "0.4048", "0.4673", "0.4048", "0.4673", "0.4153", "0.0801", "0.4153" ], [ "0.5024", "0.4016", "0.8906", "0.4028", "0.8906", "0.4172", "0.5024", "0.4167" ], [ "0.0801", "0.4185", "0.4697", "0.4185", "0.4697", "0.432", "0.0801", "0.432" ], [ "0.5034", "0.4185", "0.8906", "0.4185", "0.8906", "0.432", "0.5034", "0.432" ], [ "0.0801", "0.433", "0.468", "0.433", "0.468", "0.4463", "0.0801", "0.4463" ], [ "0.5024", "0.4343", "0.892", "0.4343", "0.892", "0.4475", "0.5024", "0.4475" ], [ "0.0801", "0.449", "0.4697", "0.449", "0.4697", "0.4622", "0.0801", "0.4622" ], [ "0.5034", "0.4507", "0.8896", "0.4507", "0.8896", "0.461", "0.5034", "0.461" ], [ "0.0801", "0.464", "0.4697", "0.464", "0.4697", "0.4778", "0.0801", "0.4778" ], [ "0.501", "0.4634", "0.892", "0.464", "0.892", "0.4792", "0.501", "0.4778" ], [ "0.0801", "0.4792", "0.468", "0.4792", "0.468", "0.4924", "0.0801", "0.4924" ], [ "0.5024", "0.4792", "0.892", "0.4792", "0.892", "0.4924", "0.5024", "0.4924" ], [ "0.0809", "0.4956", "0.468", "0.4956", "0.468", "0.506", "0.0809", "0.506" ], [ "0.5024", "0.4944", "0.892", "0.4944", "0.892", "0.5083", "0.5024", "0.5083" ], [ "0.0801", "0.509", "0.286", "0.509", "0.286", "0.523", "0.0801", "0.523" ], [ "0.501", "0.5083", "0.892", "0.509", "0.892", "0.5234", "0.501", "0.523" ], [ "0.0989", "0.526", "0.4673", "0.526", "0.4673", "0.536", "0.0989", "0.536" ], [ "0.5024", "0.5234", "0.8906", "0.5234", "0.8906", "0.537", "0.5024", "0.537" ], [ "0.0801", "0.5405", "0.468", "0.5405", "0.468", "0.5537", "0.0801", "0.5537" ], [ "0.5024", "0.539", "0.892", "0.539", "0.892", "0.553", "0.5024", "0.553" ], [ "0.0801", "0.5547", "0.4697", "0.5547", "0.4697", "0.5684", "0.0801", "0.5684" ], [ "0.5024", "0.5547", "0.892", "0.5547", "0.892", "0.5684", "0.5024", "0.5684" ], [ "0.0809", "0.5713", "0.468", "0.5713", "0.468", "0.582", "0.0809", "0.582" ], [ "0.5024", "0.5693", "0.892", "0.5693", "0.892", "0.5835", "0.5024", "0.5835" ], [ "0.0801", "0.584", "0.468", "0.584", "0.468", "0.5977", "0.0801", "0.5977" ], [ "0.5024", "0.5854", "0.8906", "0.5854", "0.8906", "0.5986", "0.5024", "0.5986" ], [ "0.0801", "0.5996", "0.468", "0.5996", "0.468", "0.613", "0.0801", "0.613" ], [ "0.501", "0.5986", "0.8555", "0.5986", "0.8555", "0.6123", "0.501", "0.6123" ], [ "0.0801", "0.6157", "0.4697", "0.6157", "0.4697", "0.629", "0.0801", "0.629" ], [ "0.523", "0.6157", "0.8906", "0.6157", "0.8906", "0.629", "0.523", "0.629" ], [ "0.0801", "0.63", "0.4714", "0.63", "0.4714", "0.643", "0.0801", "0.643" ], [ "0.5024", "0.629", "0.8906", "0.63", "0.8906", "0.6445", "0.5024", "0.643" ], [ "0.0809", "0.6465", "0.4673", "0.6465", "0.4673", "0.657", "0.0809", "0.657" ], [ "0.5034", "0.6445", "0.892", "0.6445", "0.892", "0.6577", "0.5034", "0.6577" ], [ "0.0801", "0.66", "0.4697", "0.66", "0.4697", "0.674", "0.0801", "0.674" ], [ "0.5024", "0.66", "0.892", "0.66", "0.892", "0.674", "0.5024", "0.674" ], [ "0.0801", "0.6753", "0.468", "0.6753", "0.468", "0.6895", "0.0801", "0.6895" ], [ "0.5034", "0.675", "0.892", "0.675", "0.892", "0.6895", "0.5034", "0.6895" ], [ "0.0784", "0.69", "0.468", "0.69", "0.468", "0.704", "0.0784", "0.704" ], [ "0.5034", "0.69", "0.892", "0.69", "0.892", "0.704", "0.5034", "0.704" ], [ "0.0801", "0.7056", "0.4697", "0.7056", "0.4697", "0.7197", "0.0801", "0.7197" ], [ "0.5034", "0.707", "0.8906", "0.707", "0.8906", "0.7173", "0.5034", "0.7173" ], [ "0.0801", "0.72", "0.4534", "0.72", "0.4534", "0.735", "0.0801", "0.735" ], [ "0.5024", "0.7197", "0.8906", "0.7197", "0.8906", "0.733", "0.5024", "0.733" ], [ "0.0989", "0.735", "0.468", "0.735", "0.468", "0.7485", "0.0989", "0.7485" ], [ "0.5034", "0.7373", "0.8906", "0.7373", "0.8906", "0.7476", "0.5034", "0.7476" ], [ "0.0768", "0.7505", "0.468", "0.75", "0.468", "0.7646", "0.0768", "0.765" ], [ "0.5024", "0.7505", "0.8906", "0.7505", "0.8906", "0.765", "0.5024", "0.765" ], [ "0.0784", "0.765", "0.4673", "0.765", "0.4673", "0.779", "0.0784", "0.779" ], [ "0.5024", "0.7666", "0.892", "0.7666", "0.892", "0.78", "0.5024", "0.78" ], [ "0.0784", "0.781", "0.4697", "0.781", "0.4697", "0.7944", "0.0784", "0.7944" ], [ "0.5034", "0.7827", "0.8906", "0.7827", "0.8906", "0.7935", "0.5034", "0.7935" ], [ "0.0801", "0.7954", "0.4673", "0.7954", "0.4673", "0.8096", "0.0801", "0.8096" ], [ "0.5024", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5024", "0.81" ], [ "0.0989", "0.811", "0.4697", "0.811", "0.4697", "0.8247", "0.0989", "0.8247" ], [ "0.5024", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5024", "0.8247" ], [ "0.0801", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0801", "0.8403" ], [ "0.5024", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.5024", "0.8403" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.5024", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.5024", "0.855" ], [ "0.0784", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0784", "0.8706" ], [ "0.5024", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5024", "0.8706" ], [ "0.0768", "0.8706", "0.4697", "0.872", "0.4697", "0.886", "0.0768", "0.8853" ], [ "0.5024", "0.872", "0.8936", "0.872", "0.8936", "0.886", "0.5024", "0.886" ], [ "0.0801", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0801", "0.901" ], [ "0.501", "0.887", "0.6274", "0.886", "0.6274", "0.901", "0.501", "0.9014" ] ]

[ "2. Related Work", "and broadens the pool of available data by semi-supervised", "learning.", "Image based scene understanding methods[13, 9] can be di-", "rectly applied to individual video frames to generate a de-", "3. Dataset and Geometric Classes", "scription of the scene. However, these methods do not ex-", "ploit temporal information across neighboring frames. Fur-", "Existing Datasets: In our supervised learning setting, we", "ther,lackingtemporal consistency,they canresult in tem-", "require an annotated dataset supplying ground truth labels", "porally inconsistent labels across frames, which can im-", "for training and evaluation. While several datasets for ge-", "pact performance, as scene labels suddenly change between", "ometric scene understanding exists on still images [13, 9],", "frames.In addition, frame-based methods do not exploit", "our video-based scene analysis method demands an anno-", "temporal redundancy to process videos efficiently as pro-", "tated video dataset.However, existing datasets for video", "cessing each segment in video independently results in a", "sceneunderstanding only providelimitedground truthdata", "longer processing time.", "The CamVid dataset [2] provides pixel-level labels for 70]", "Temporal information can be incorporated using struc-", "non-consecutiveframes(aboutevery3Othframe,sampled", "ture from motion (SfM) [16, 1], which requires substan-", "at 1fps). The NYUScenes [6], and MPI-VehicleScenes [19]", "tial computation and might not generalize well to dynamic", "dataset consists of 74 and 156 annotated frames, respec-", "objects. SfM makes explicit assumptions about the scene,", "tively. Therefore, these datasets are not ideally suited for", "e.g. , mostly static with limited foreground motion, and re-", "comprehensive studies.To overcome this limitation, we", "quires basic camera priors. In addition, SfM reconstruc-", "provide a novel, pixel-level annotated dataset for geometric", "tion can fail due to lack of parallax, e.g. , walking forward.", "scene analysis of video, consisting of over 20,000 frames", "and rotation around camera center. We do not make any", "across 100 videos.", "assumptionabout the scenecontent,amount of foreground", "A video dataset for geometric scene understanding:", "motion, or the camera used. Another approach to achieve", "Our dataset consists of 160 outdoor videos, with annota-", "temporal consistency across frames is to use optical flow", "tions available for a subset of 100 videos. Some videos are", "between consecutive frames to estimate the neighborhood", "collectedfromYouTubeandothersarerecordedbyuswhile", "of each pixel and then combine past predictions to make", "walking or driving in an urban area. Video lengths range", "a final prediction [14]. This requires labeling every pixel", "from 60 to 400 frames and resolution varies from 320 × 480", "in every framein the video independently,which doesn't", "to 600 × 800, with varying aspect ratios. We partitioned", "leverage the causality in video.", "the datasets into three sets: 63 videos are used for training", "Our video scene understanding approach takes advan-", "and cross-validation (13,000 frames), 40 videos for inde-", "tage of spatio-temporal information by employing hierar-", "pendent testing via external-validation (7,000 frames), and", "chical video segmentation[10], which segments a video into", "60 videos are kept unlabeled (14,000 frames) and are later", "spatio-temporal regions.Further,weleverage causality in", "used for semi-supervised learning (Section 5.1). Videos in", "videos to efficiently label videos, achieving favorable com-", "the cross and external-validation sets are completely anno-", "plexity which is linear in the number of unique spatio-", "tated with ground truth labels (every frame and pixel).", "temporal segments in videos. Consequently, in contrast to", "Videos in our dataset contain entities such as sky", "image based or independent frame labeling, our system is", "ground, buildings, trees, and objects (cars, trains, humans).", "not directly affected by the total number of frames. Re-", "While many different partitions can be imagined, we fol-", "cently, Tighe et al. [17] applied their image label transfer to", "low [13, 11] and partition the video content into three main", "the video domain leveraging[10], by applying a max heuris-", "geometric classes:\"Sky\", “support\", and “vertical\"To", "tic across frames. In contrast, our approach performs geo-", "provide a more detailed description of the scene, we fur-", "metric labeling by leveraging multiple hierarchy level while", "ther divide the vertical class into three subclasses:\"Solid\"", "probabilistic aggregating labels over a temporal window.", "“porous\", and “object\".The solid vertical sub-class in-", "A significant hurdle in video scene understanding is the", "cludes solid,static objects resting on the ground, such as", " availability of a ground truth annotated dataset for train-", "buildings, boards, bridges, and rocks etc. The porous verti-", "ing. While several datasets exist for predicting geometric", "cal sub-class includes non-solid, static objects such as trees", "context in the image domain [13, 9], datasets for videos", "and foliage.Finally,movable objects,like humans, cars.", "[2, 6, 19] are currently limited in their scope. (see section 3)", "boats, and trains are included in the object class. Notice,", "Our video scene analysis method builds upon Hoeim", "that in contrast to [13, 11] we do not account for the ori-", "et al. \"s [13] image based approach, extending the im-", "entation of the vertical classes as their identity is likely", "age based approach to video.Our approach differs, in", "to change due to camera motion in video. Table 1 gives", "that it is taking advantage of spatio-temporal context, ex-", "an overview of the distribution of the classes in the cross-", "tends feature set being more suitable for video, leverages", "validation dataset, by showing the pixel area of each of the", "temporal redundancy while achieving temporal consistency", "geometric classes." ]

{ "arxiv_id": "1510.07320", "language": "en", "page_idx": 1 }

[ [ "0.10297", "0.125", "0.898", "0.125", "0.898", "0.1383", "0.10297", "0.1383" ], [ "0.152", "0.1718", "0.4944", "0.1718", "0.4944", "0.1863", "0.152", "0.1863" ], [ "0.576", "0.1704", "0.777", "0.1704", "0.777", "0.1881", "0.576", "0.1881" ], [ "0.232", "0.1875", "0.4143", "0.1875", "0.4143", "0.2008", "0.232", "0.2008" ], [ "0.562", "0.1875", "0.789", "0.1875", "0.789", "0.202", "0.562", "0.202" ], [ "0.2484", "0.1989", "0.398", "0.2008", "0.398", "0.2153", "0.2484", "0.2134" ], [ "0.615", "0.2008", "0.738", "0.2008", "0.738", "0.214", "0.615", "0.214" ], [ "0.259", "0.214", "0.384", "0.2134", "0.384", "0.2279", "0.259", "0.2291" ], [ "0.614", "0.214", "0.738", "0.214", "0.738", "0.2291", "0.614", "0.2291" ], [ "0.2133", "0.2247", "0.4346", "0.2267", "0.4346", "0.2443", "0.2133", "0.2424" ], [ "0.523", "0.2279", "0.83", "0.2279", "0.83", "0.2424", "0.523", "0.2424" ], [ "0.02287", "0.2683", "0.06046", "0.2683", "0.06046", "0.3726", "0.02287", "0.3726" ], [ "0.25", "0.2966", "0.3162", "0.2966", "0.3162", "0.3113", "0.25", "0.3113" ], [ "0.6807", "0.2986", "0.7524", "0.2986", "0.7524", "0.3093", "0.6807", "0.3093" ], [ "0.5557", "0.3093", "0.88", "0.3093", "0.88", "0.3193", "0.5557", "0.3193" ], [ "0.10455", "0.3208", "0.46", "0.3208", "0.46", "0.334", "0.10455", "0.334" ], [ "0.581", "0.318", "0.846", "0.318", "0.846", "0.3289", "0.581", "0.3289" ], [ "0.10297", "0.334", "0.4631", "0.334", "0.4631", "0.348", "0.10297", "0.348" ], [ "0.10455", "0.348", "0.4617", "0.348", "0.4617", "0.358", "0.10455", "0.358" ], [ "0.524", "0.351", "0.535", "0.348", "0.539", "0.356", "0.528", "0.3591" ], [ "0.7114", "0.3528", "0.721", "0.3528", "0.721", "0.3599", "0.7114", "0.3599" ], [ "0.10297", "0.3591", "0.4631", "0.3591", "0.4631", "0.3726", "0.10297", "0.3726" ], [ "0.8203", "0.3547", "0.8267", "0.3499", "0.8325", "0.3542", "0.826", "0.3591" ], [ "0.02696", "0.3662", "0.05554", "0.3662", "0.05554", "0.3984", "0.02696", "0.3984" ], [ "0.6216", "0.3662", "0.6855", "0.3674", "0.6855", "0.3757", "0.6216", "0.375" ], [ "0.71", "0.3662", "0.773", "0.3674", "0.773", "0.3757", "0.71", "0.375" ], [ "0.8384", "0.3662", "0.8857", "0.3662", "0.8857", "0.377", "0.8384", "0.377" ], [ "0.5254", "0.3674", "0.558", "0.3674", "0.558", "0.375", "0.5254", "0.375" ], [ "0.03595", "0.3674", "0.0474", "0.3674", "0.0474", "0.3706", "0.03595", "0.3706" ], [ "0.58", "0.3694", "0.596", "0.3694", "0.596", "0.3738", "0.58", "0.3738" ], [ "0.7715", "0.3694", "0.8022", "0.3694", "0.8022", "0.3738", "0.7715", "0.3738" ], [ "0.528", "0.3738", "0.5933", "0.3738", "0.5933", "0.3838", "0.528", "0.3838" ], [ "0.10297", "0.3718", "0.4631", "0.3718", "0.4631", "0.385", "0.10297", "0.385" ], [ "0.613", "0.3757", "0.6807", "0.3757", "0.6807", "0.3833", "0.613", "0.3833" ], [ "0.709", "0.3757", "0.734", "0.3738", "0.7344", "0.3806", "0.71", "0.3833" ], [ "0.7573", "0.3757", "0.806", "0.3757", "0.806", "0.3833", "0.7573", "0.3833" ], [ "0.822", "0.3757", "0.9004", "0.3757", "0.9004", "0.3833", "0.822", "0.3833" ], [ "0.609", "0.3801", "0.701", "0.3806", "0.701", "0.3926", "0.609", "0.3914" ], [ "0.527", "0.3806", "0.597", "0.3818", "0.597", "0.3933", "0.527", "0.3926" ], [ "0.7134", "0.3818", "0.8037", "0.3818", "0.8037", "0.3926", "0.7134", "0.3926" ], [ "0.839", "0.3838", "0.886", "0.3838", "0.886", "0.3914", "0.839", "0.3914" ], [ "0.54", "0.3914", "0.5557", "0.3914", "0.5557", "0.3997", "0.54", "0.3997" ], [ "0.10297", "0.3838", "0.4631", "0.3838", "0.4631", "0.3977", "0.10297", "0.3977" ], [ "0.7183", "0.3914", "0.7344", "0.3914", "0.7344", "0.3984", "0.7183", "0.3984" ], [ "0.732", "0.3926", "0.7715", "0.3984", "0.7686", "0.406", "0.7295", "0.3997" ], [ "0.786", "0.3926", "0.7974", "0.3926", "0.7974", "0.3965", "0.786", "0.3965" ], [ "0.8433", "0.3945", "0.902", "0.3965", "0.902", "0.404", "0.8433", "0.4016" ], [ "0.621", "0.3933", "0.664", "0.3933", "0.664", "0.4048", "0.621", "0.4048" ], [ "0.10297", "0.3965", "0.4631", "0.3965", "0.4631", "0.4104", "0.10297", "0.4104" ], [ "0.5645", "0.3958", "0.574", "0.3984", "0.5703", "0.4048", "0.5605", "0.4016" ], [ "0.8228", "0.3977", "0.8447", "0.3977", "0.8447", "0.4016", "0.8228", "0.4016" ], [ "0.02287", "0.404", "0.0621", "0.404", "0.0621", "0.486", "0.02287", "0.486" ], [ "0.5337", "0.404", "0.589", "0.404", "0.589", "0.414", "0.5337", "0.414" ], [ "0.6313", "0.404", "0.681", "0.404", "0.681", "0.414", "0.6313", "0.414" ], [ "0.7134", "0.404", "0.8013", "0.404", "0.8013", "0.414", "0.7134", "0.414" ], [ "0.822", "0.404", "0.899", "0.404", "0.899", "0.414", "0.822", "0.414" ], [ "0.527", "0.4192", "0.544", "0.4192", "0.544", "0.4268", "0.527", "0.4268" ], [ "0.10297", "0.4092", "0.4631", "0.4092", "0.4631", "0.4224", "0.10297", "0.4224" ], [ "0.10297", "0.4216", "0.4631", "0.4216", "0.4631", "0.435", "0.10297", "0.435" ], [ "0.619", "0.4192", "0.636", "0.4192", "0.636", "0.4268", "0.619", "0.4268" ], [ "0.558", "0.432", "0.5864", "0.432", "0.5864", "0.4363", "0.558", "0.4363" ], [ "0.634", "0.43", "0.6895", "0.432", "0.6895", "0.443", "0.634", "0.4414" ], [ "0.10455", "0.4363", "0.4617", "0.4363", "0.4617", "0.4463", "0.10455", "0.4463" ], [ "0.5415", "0.4363", "0.5557", "0.4363", "0.5557", "0.4443", "0.5415", "0.4443" ], [ "0.564", "0.4363", "0.576", "0.4395", "0.574", "0.4458", "0.5605", "0.4414" ], [ "0.574", "0.4382", "0.5923", "0.4382", "0.5923", "0.4458", "0.574", "0.4458" ], [ "0.734", "0.4363", "0.796", "0.4363", "0.796", "0.443", "0.734", "0.443" ], [ "0.10297", "0.449", "0.46", "0.449", "0.46", "0.459", "0.10297", "0.459" ], [ "0.6216", "0.4475", "0.648", "0.4475", "0.648", "0.455", "0.6216", "0.455" ], [ "0.6895", "0.449", "0.701", "0.449", "0.701", "0.4526", "0.6895", "0.4526" ], [ "0.7173", "0.4495", "0.8096", "0.4495", "0.8096", "0.4602", "0.7173", "0.4602" ], [ "0.619", "0.452", "0.705", "0.4526", "0.705", "0.464", "0.619", "0.4634" ], [ "0.527", "0.4507", "0.6045", "0.4507", "0.6045", "0.4583", "0.527", "0.4583" ], [ "0.10455", "0.461", "0.4617", "0.461", "0.4617", "0.4717", "0.10455", "0.4717" ], [ "0.527", "0.459", "0.6055", "0.459", "0.6055", "0.4666", "0.527", "0.4666" ], [ "0.7173", "0.461", "0.749", "0.461", "0.749", "0.464", "0.7173", "0.464" ], [ "0.759", "0.461", "0.7876", "0.461", "0.7876", "0.464", "0.759", "0.464" ], [ "0.7974", "0.461", "0.8037", "0.461", "0.8037", "0.464", "0.7974", "0.464" ], [ "0.6235", "0.4622", "0.6978", "0.4622", "0.6978", "0.473", "0.6235", "0.473" ], [ "0.5317", "0.4653", "0.601", "0.4653", "0.601", "0.476", "0.5317", "0.476" ], [ "0.638", "0.4685", "0.6807", "0.4685", "0.6807", "0.4792", "0.638", "0.4792" ], [ "0.74", "0.4653", "0.7817", "0.4634", "0.7837", "0.4734", "0.7417", "0.476" ], [ "0.10297", "0.4717", "0.4631", "0.4717", "0.4631", "0.4849", "0.10297", "0.4849" ], [ "0.5557", "0.4749", "0.574", "0.4749", "0.574", "0.4817", "0.5557", "0.4817" ], [ "0.10297", "0.4849", "0.4617", "0.4849", "0.4617", "0.4988", "0.10297", "0.4988" ], [ "0.02287", "0.4937", "0.06046", "0.4937", "0.06046", "0.712", "0.02287", "0.712" ], [ "0.5186", "0.4924", "0.9126", "0.4937", "0.9126", "0.5083", "0.5186", "0.507" ], [ "0.10455", "0.4988", "0.46", "0.4988", "0.46", "0.509", "0.10455", "0.509" ], [ "0.1013", "0.5103", "0.4617", "0.5103", "0.4617", "0.5234", "0.1013", "0.5234" ], [ "0.5195", "0.5103", "0.6045", "0.5103", "0.6045", "0.521", "0.5195", "0.521" ], [ "0.10455", "0.5234", "0.4617", "0.5234", "0.4617", "0.534", "0.10455", "0.534" ], [ "0.1013", "0.535", "0.4617", "0.535", "0.4617", "0.549", "0.1013", "0.549" ], [ "0.10297", "0.549", "0.4617", "0.549", "0.4617", "0.5596", "0.10297", "0.5596" ], [ "0.5215", "0.5537", "0.91", "0.5537", "0.91", "0.5674", "0.5215", "0.5674" ], [ "0.1013", "0.56", "0.3718", "0.56", "0.3718", "0.5737", "0.1013", "0.5737" ], [ "0.5195", "0.5674", "0.9126", "0.5674", "0.9126", "0.5806", "0.5195", "0.5806" ], [ "0.5215", "0.582", "0.914", "0.582", "0.914", "0.595", "0.5215", "0.595" ], [ "0.2378", "0.595", "0.3513", "0.595", "0.3513", "0.61", "0.2378", "0.61" ], [ "0.5215", "0.595", "0.7812", "0.595", "0.7812", "0.6094", "0.5215", "0.6094" ], [ "0.536", "0.6113", "0.91", "0.6113", "0.91", "0.625", "0.536", "0.625" ], [ "0.2157", "0.5977", "0.2296", "0.5977", "0.2296", "0.608", "0.2157", "0.608" ], [ "0.08496", "0.6157", "0.478", "0.616", "0.478", "0.631", "0.08496", "0.63" ], [ "0.0866", "0.631", "0.4795", "0.631", "0.4795", "0.6445", "0.0866", "0.6445" ], [ "0.5215", "0.6255", "0.9126", "0.6255", "0.9126", "0.6396", "0.5215", "0.6396" ], [ "0.0866", "0.6445", "0.4795", "0.6445", "0.4795", "0.6577", "0.0866", "0.6577" ], [ "0.5215", "0.6396", "0.914", "0.6396", "0.914", "0.653", "0.5215", "0.653" ], [ "0.0866", "0.6577", "0.478", "0.6577", "0.478", "0.672", "0.0866", "0.672" ], [ "0.5195", "0.653", "0.9126", "0.653", "0.9126", "0.6665", "0.5195", "0.6665" ], [ "0.0866", "0.6724", "0.478", "0.6724", "0.478", "0.686", "0.0866", "0.686" ], [ "0.5195", "0.6665", "0.914", "0.6665", "0.914", "0.6797", "0.5195", "0.6797" ], [ "0.5195", "0.6797", "0.914", "0.6797", "0.914", "0.6943", "0.5195", "0.6943" ], [ "0.0866", "0.686", "0.478", "0.686", "0.478", "0.6997", "0.0866", "0.6997" ], [ "0.5195", "0.6943", "0.911", "0.6943", "0.911", "0.7085", "0.5195", "0.7085" ], [ "0.0866", "0.7007", "0.478", "0.7007", "0.478", "0.714", "0.0866", "0.714" ], [ "0.08496", "0.714", "0.478", "0.714", "0.478", "0.728", "0.08496", "0.728" ], [ "0.5215", "0.7085", "0.91", "0.7085", "0.91", "0.7217", "0.5215", "0.7217" ], [ "0.5195", "0.7217", "0.9126", "0.7217", "0.9126", "0.7363", "0.5195", "0.7363" ], [ "0.0866", "0.728", "0.4795", "0.728", "0.4795", "0.741", "0.0866", "0.741" ], [ "0.5195", "0.7363", "0.9126", "0.7363", "0.9126", "0.75", "0.5195", "0.75" ], [ "0.0866", "0.7437", "0.3816", "0.7437", "0.3816", "0.754", "0.0866", "0.754" ], [ "0.1013", "0.756", "0.478", "0.756", "0.478", "0.7695", "0.1013", "0.7695" ], [ "0.5215", "0.75", "0.911", "0.75", "0.911", "0.763", "0.5215", "0.763" ], [ "0.5195", "0.763", "0.9126", "0.763", "0.9126", "0.7764", "0.5195", "0.7764" ], [ "0.0866", "0.7715", "0.4763", "0.7715", "0.4763", "0.782", "0.0866", "0.782" ], [ "0.5195", "0.778", "0.864", "0.778", "0.864", "0.7886", "0.5195", "0.7886" ], [ "0.0866", "0.784", "0.4756", "0.784", "0.4756", "0.7974", "0.0866", "0.7974" ], [ "0.0866", "0.7974", "0.478", "0.7974", "0.478", "0.8125", "0.0866", "0.8125" ], [ "0.536", "0.7915", "0.91", "0.7915", "0.91", "0.8047", "0.536", "0.8047" ], [ "0.5195", "0.8047", "0.9126", "0.8066", "0.9126", "0.8213", "0.5195", "0.8193" ], [ "0.08496", "0.811", "0.4795", "0.811", "0.4795", "0.8247", "0.08496", "0.8247" ], [ "0.08496", "0.8257", "0.478", "0.8257", "0.478", "0.839", "0.08496", "0.839" ], [ "0.5195", "0.821", "0.914", "0.821", "0.914", "0.834", "0.5195", "0.834" ], [ "0.08496", "0.839", "0.478", "0.839", "0.478", "0.853", "0.08496", "0.853" ], [ "0.5195", "0.834", "0.9126", "0.834", "0.9126", "0.8477", "0.5195", "0.8477" ], [ "0.5186", "0.8477", "0.911", "0.8486", "0.911", "0.8623", "0.5186", "0.8613" ], [ "0.0866", "0.863", "0.4795", "0.863", "0.4795", "0.877", "0.0866", "0.877" ], [ "0.5215", "0.8623", "0.914", "0.8623", "0.914", "0.8755", "0.5215", "0.8755" ], [ "0.0866", "0.877", "0.3718", "0.877", "0.3718", "0.89", "0.0866", "0.89" ], [ "0.5195", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5195", "0.8896" ] ]

[ "InterpretableRumorDetection in Microblogs by Attending to UserInteractions", "Ling Min Serena Khoo, Hai Leong Chieu", "Zhong Qian, Jing Jiang", "DSONational Laboratories", "Singapore Management University", "12 Science Park Drive", "80 Stamford Road", "Singapore 118225", "Singapore 178902", "[klingmin,chaileon) @dso.org.sg", "qianzhongqz@163.com,jingjiang@smu.edu.sg", " e", "Abstract", "Source Claim:", "walmartdonates s10,o00to support darren wilson and theon going", "We address rumor detection by learning to differentiate be-", "racist policemurders#ferguson#boycottwalmart URL", "tween the community's response to real and fake claims in", "microblogs.Existing state-of-the-artmodels are based on tree", "R,", "R_3", "models that model conversational trees. However, in social", "9", "2", "@TefPoeWalmal", "ethanksWalmart", ".Well I will b", "@TefPo", "", "uohi", "tdonate", "wouldn't surpris", "media, a user posting a reply might be replying to the entire", "donatesS10ktoD", "$10.00", "efendWilsor", "paying off that credit", "Wilson?!No way this is", "me, is there proof", "after beaten in struggle", "ardandneve", "of th", "thread rather than to a specific user. We propose a post-level", "for", "gueEerqus", "nar", "congowherethal", "Link?Fer", "attention model (PLAN) to model long distance interactions", "de", "hon", "[T's]", "couldn't find.", "#sayitaintso\"", "http:/f.co/g3Jku9YuGt", "condones murderl!'", "R_1_", "between tweets with the multi-head attention mechanism in a", "transformer network. We investigated variants of this model:", "R_2_1", "vburtus", "\"@davidgrelle", "(1)a structure aware self-attention model (StA-PLAN)that", "?70", "n", "omar", "@SAGandAFTRA", "incorporates tree structure information in thetransformer net-", "@TefP", "W", "not of #MikeBrown.Lie", "can't support this due", "googled itto I'mthe", "work, and (2) a hierarchical token and post-level attention", "type to jump all over", "dehunke", "rstwee", "", "to laws regulAting", "what's on social", "public cos", "shooting.\"", "model (StA-HiTPLAN) that learns a sentence representation", "media", "with token-level self-attention. To the best of our knowledge,", "1A29901100Z:AIXE", "Figure 1: Sample of a thread from Twitter15 resulting from", "we arethefirstto evaluateourmodels on tworumordetection", "data sets: the PHEME data set as well as the Twitter15 and", "a fake claim.", "Twitterl6 data sets.Weshow that ourbest models outperform", "current state-of-the-art models for both data sets. Moreover,", "theattention mechanismallowsustoexplain rumordetection", "Hence, the interaction between users as well as the con-", "predictions at both token-level and post-level.", "tent shared could be captured for fake news detection. We", "discuss briefly two state-of-the-art models (Ma, Gao, and", "Introduction", "Wong 2018b; Kumar and Carley 2019).", "Ma et al. (2018b) organized the source claim and its re-", "1.", "The spread of fake news can have far reaching and devas-", "tating effects. $130 billion in stock value was wiped out in", "sponding tweets in a tree structure as shown in Figure 1.", "minutes after a false Associated Press's tweet claimed that", "Each tweet is represented as a node; the top node would", "Barack Obama was injured following an explosion in 2013", "be the source claim and the children of a node are tweets", "(Rapoza 2017). Some researchers even believe that fake", "that have responded to it directly. They modeled the spread", "of information in a propagation tree using recursive neural", "news has affected the outcome of the 2016 United States", "networks. Signals from different nodes are aggregated re-", "presidential election (Gunther, Beck, and Nisbet 2018). The", "severity of the impact of fake news warrants the need for an", "cursively in either a bottom-up or a top-down manner. In-", "formation is propagated from the child node to the parent", "effective and automated means of detecting fake news and", "node in the bottom-up model and vice versa in a top-down", "hashencespurredmuchresearchinthisarea.", "Our work focuses on the detection of fake claims using", "model. Similarly, Kumar et al. (2019) organized conversa-", "tion threads with a tree structure and explored several vari-", "community response to such claims.This area of research", "ants ofbranchand treeLSTMsforrumourdetection", "exploits the collective wisdom of the community by ap-", "plying natural language processing to comments directed", "Both papers used tree models with the intention of mod-", "elling structural information present in the conversation", "towards a claim (see Figure 1). The key principle behind", " such works is that users on social media would share opin-", "thread. Information is propagated either from the parent to", "ions, conjectures and evidences for inaccurate information.", "the child or vice versa in tree models. However, the thread", "structure in social media conversations might be unclear", "Copyright @ 2020, Association for the Advancement of Artificial", "Each user is often able to observe all the replies in different", "Intelligence (www.aaai.org).All rights reserved.", "branches of the conversation. A user debunking a fake news" ]

{ "arxiv_id": "2001.10667", "language": "en", "page_idx": 0 }

[ [ "0.0866", "0.06946", "0.4756", "0.06946", "0.4756", "0.0833", "0.0866", "0.0833" ], [ "0.5376", "0.06946", "0.9126", "0.06946", "0.9126", "0.0833", "0.5376", "0.0833" ], [ "0.0866", "0.084", "0.478", "0.084", "0.478", "0.09784", "0.0866", "0.09784" ], [ "0.5195", "0.084", "0.911", "0.084", "0.911", "0.09784", "0.5195", "0.09784" ], [ "0.08496", "0.09784", "0.4795", "0.09784", "0.4795", "0.1111", "0.08496", "0.1111" ], [ "0.5195", "0.09784", "0.914", "0.0966", "0.914", "0.1111", "0.5195", "0.11237" ], [ "0.08496", "0.11237", "0.4795", "0.11237", "0.4795", "0.1256", "0.08496", "0.1256" ], [ "0.5215", "0.113", "0.91", "0.113", "0.91", "0.1237", "0.5215", "0.1237" ], [ "0.0866", "0.1256", "0.4795", "0.1256", "0.4795", "0.1395", "0.0866", "0.1395" ], [ "0.5195", "0.1256", "0.911", "0.1256", "0.911", "0.1395", "0.5195", "0.1395" ], [ "0.0866", "0.1395", "0.4763", "0.1395", "0.4763", "0.1528", "0.0866", "0.1528" ], [ "0.5195", "0.1395", "0.9126", "0.1395", "0.9126", "0.1528", "0.5195", "0.1528" ], [ "0.08826", "0.1528", "0.4763", "0.1528", "0.4763", "0.1666", "0.08826", "0.1666" ], [ "0.5186", "0.1528", "0.9126", "0.1515", "0.9126", "0.1654", "0.5186", "0.1666" ], [ "0.08826", "0.1666", "0.478", "0.1666", "0.478", "0.1799", "0.08826", "0.1799" ], [ "0.5215", "0.1666", "0.911", "0.1666", "0.911", "0.1799", "0.5215", "0.1799" ], [ "0.08826", "0.1818", "0.478", "0.1818", "0.478", "0.1925", "0.08826", "0.1925" ], [ "0.5195", "0.1812", "0.9126", "0.1812", "0.9126", "0.1945", "0.5195", "0.1945" ], [ "0.08826", "0.1957", "0.4763", "0.1957", "0.4763", "0.2058", "0.08826", "0.2058" ], [ "0.5195", "0.1945", "0.7075", "0.1945", "0.7075", "0.2084", "0.5195", "0.2084" ], [ "0.08417", "0.207", "0.478", "0.2084", "0.478", "0.2229", "0.08417", "0.2216" ], [ "0.5337", "0.207", "0.9126", "0.2084", "0.9126", "0.2229", "0.5337", "0.2216" ], [ "0.08496", "0.2229", "0.4763", "0.2229", "0.4763", "0.2361", "0.08496", "0.2361" ], [ "0.5215", "0.2229", "0.9126", "0.2229", "0.9126", "0.2361", "0.5215", "0.2361" ], [ "0.0866", "0.2374", "0.4763", "0.2374", "0.4763", "0.2474", "0.0866", "0.2474" ], [ "0.523", "0.238", "0.9126", "0.238", "0.9126", "0.2488", "0.523", "0.2488" ], [ "0.0866", "0.2507", "0.4795", "0.2507", "0.4795", "0.2644", "0.0866", "0.2644" ], [ "0.5195", "0.2507", "0.9126", "0.2507", "0.9126", "0.2644", "0.5195", "0.2644" ], [ "0.08496", "0.2632", "0.2524", "0.2644", "0.2524", "0.2778", "0.08496", "0.2766" ], [ "0.523", "0.2664", "0.911", "0.2664", "0.911", "0.2766", "0.523", "0.2766" ], [ "0.10297", "0.2778", "0.4763", "0.2778", "0.4763", "0.2917", "0.10297", "0.2917" ], [ "0.5215", "0.2798", "0.8613", "0.2798", "0.8613", "0.2903", "0.5215", "0.2903" ], [ "0.0866", "0.2917", "0.478", "0.2917", "0.478", "0.305", "0.0866", "0.305" ], [ "0.535", "0.2917", "0.911", "0.2922", "0.911", "0.307", "0.535", "0.3062" ], [ "0.08496", "0.3037", "0.478", "0.305", "0.478", "0.3193", "0.08496", "0.318" ], [ "0.5215", "0.307", "0.9126", "0.307", "0.9126", "0.3208", "0.5215", "0.3208" ], [ "0.0866", "0.3208", "0.4763", "0.3208", "0.4763", "0.3308", "0.0866", "0.3308" ], [ "0.5215", "0.3208", "0.9126", "0.3208", "0.9126", "0.334", "0.5215", "0.334" ], [ "0.08496", "0.3333", "0.4795", "0.3333", "0.4795", "0.3467", "0.08496", "0.3467" ], [ "0.5215", "0.3352", "0.9126", "0.3352", "0.9126", "0.3484", "0.5215", "0.3484" ], [ "0.08496", "0.3467", "0.4795", "0.3467", "0.4795", "0.3599", "0.08496", "0.3599" ], [ "0.5195", "0.3484", "0.9126", "0.3484", "0.9126", "0.3623", "0.5195", "0.3623" ], [ "0.0866", "0.3623", "0.4763", "0.3623", "0.4763", "0.3726", "0.0866", "0.3726" ], [ "0.5215", "0.3623", "0.911", "0.3623", "0.911", "0.3757", "0.5215", "0.3757" ], [ "0.08496", "0.375", "0.4795", "0.375", "0.4795", "0.3882", "0.08496", "0.3882" ], [ "0.5215", "0.3757", "0.91", "0.3757", "0.91", "0.3901", "0.5215", "0.3901" ], [ "0.08496", "0.3882", "0.4795", "0.3882", "0.4795", "0.4016", "0.08496", "0.4016" ], [ "0.5195", "0.3894", "0.9126", "0.3894", "0.9126", "0.4028", "0.5195", "0.4028" ], [ "0.08496", "0.4016", "0.4756", "0.4016", "0.4756", "0.4153", "0.08496", "0.4153" ], [ "0.523", "0.4048", "0.9126", "0.4048", "0.9126", "0.4153", "0.523", "0.4153" ], [ "0.0866", "0.4167", "0.4763", "0.4167", "0.4763", "0.43", "0.0866", "0.43" ], [ "0.5195", "0.4172", "0.914", "0.4172", "0.914", "0.4312", "0.5195", "0.4312" ], [ "0.0866", "0.43", "0.4795", "0.43", "0.4795", "0.443", "0.0866", "0.443" ], [ "0.5215", "0.4312", "0.911", "0.4312", "0.911", "0.4443", "0.5215", "0.4443" ], [ "0.08826", "0.4458", "0.406", "0.4458", "0.406", "0.4558", "0.08826", "0.4558" ], [ "0.523", "0.4463", "0.911", "0.4463", "0.911", "0.457", "0.523", "0.457" ], [ "0.5195", "0.459", "0.846", "0.459", "0.846", "0.473", "0.5195", "0.473" ], [ "0.0866", "0.4717", "0.4084", "0.4717", "0.4084", "0.4849", "0.0866", "0.4849" ], [ "0.5376", "0.4734", "0.91", "0.4734", "0.91", "0.4873", "0.5376", "0.4873" ], [ "0.08496", "0.486", "0.4739", "0.488", "0.4739", "0.502", "0.08496", "0.5" ], [ "0.5215", "0.4873", "0.911", "0.4873", "0.911", "0.5005", "0.5215", "0.5005" ], [ "0.10297", "0.502", "0.478", "0.502", "0.478", "0.515", "0.10297", "0.515" ], [ "0.5215", "0.502", "0.911", "0.502", "0.911", "0.515", "0.5215", "0.515" ], [ "0.10455", "0.5166", "0.4795", "0.5166", "0.4795", "0.53", "0.10455", "0.53" ], [ "0.5195", "0.515", "0.9126", "0.515", "0.9126", "0.529", "0.5195", "0.529" ], [ "0.1013", "0.53", "0.3406", "0.53", "0.3406", "0.5444", "0.1013", "0.5444" ], [ "0.5195", "0.529", "0.9126", "0.528", "0.9126", "0.5415", "0.5195", "0.5425" ], [ "0.0866", "0.547", "0.4714", "0.547", "0.4714", "0.561", "0.0866", "0.561" ], [ "0.5215", "0.5425", "0.9126", "0.5425", "0.9126", "0.5566", "0.5215", "0.5566" ], [ "0.1013", "0.5596", "0.4763", "0.56", "0.4763", "0.5747", "0.1013", "0.5737" ], [ "0.5195", "0.5566", "0.914", "0.5566", "0.914", "0.571", "0.5195", "0.571" ], [ "0.10297", "0.5747", "0.4763", "0.5747", "0.4763", "0.5884", "0.10297", "0.5884" ], [ "0.5215", "0.5693", "0.911", "0.571", "0.911", "0.584", "0.5215", "0.5835" ], [ "0.10297", "0.589", "0.4763", "0.589", "0.4763", "0.603", "0.10297", "0.603" ], [ "0.5195", "0.5835", "0.9126", "0.584", "0.9126", "0.5986", "0.5195", "0.5977" ], [ "0.5195", "0.5977", "0.9126", "0.5977", "0.9126", "0.6113", "0.5195", "0.6113" ], [ "0.08826", "0.605", "0.4756", "0.605", "0.4756", "0.619", "0.08826", "0.619" ], [ "0.5195", "0.6123", "0.9126", "0.6123", "0.9126", "0.6255", "0.5195", "0.6255" ], [ "0.1054", "0.622", "0.3162", "0.622", "0.3162", "0.632", "0.1054", "0.632" ], [ "0.5195", "0.6255", "0.9126", "0.6255", "0.9126", "0.6396", "0.5195", "0.6396" ], [ "0.1054", "0.637", "0.478", "0.637", "0.478", "0.6514", "0.1054", "0.6514" ], [ "0.5215", "0.6416", "0.911", "0.6416", "0.911", "0.6514", "0.5215", "0.6514" ], [ "0.0866", "0.651", "0.478", "0.651", "0.478", "0.664", "0.0866", "0.664" ], [ "0.523", "0.655", "0.91", "0.655", "0.91", "0.6655", "0.523", "0.6655" ], [ "0.08496", "0.6655", "0.4795", "0.6655", "0.4795", "0.6787", "0.08496", "0.6787" ], [ "0.5215", "0.6665", "0.9126", "0.6665", "0.9126", "0.6797", "0.5215", "0.6797" ], [ "0.08496", "0.6787", "0.4795", "0.6787", "0.4795", "0.6924", "0.08496", "0.6924" ], [ "0.5215", "0.681", "0.91", "0.681", "0.91", "0.6943", "0.5215", "0.6943" ], [ "0.0866", "0.6934", "0.1708", "0.6934", "0.1708", "0.705", "0.0866", "0.705" ], [ "0.5195", "0.6943", "0.914", "0.6943", "0.914", "0.7085", "0.5195", "0.7085" ], [ "0.5195", "0.7085", "0.9126", "0.7085", "0.9126", "0.7217", "0.5195", "0.7217" ], [ "0.2075", "0.7183", "0.3555", "0.7183", "0.3555", "0.733", "0.2075", "0.733" ], [ "0.5215", "0.7246", "0.91", "0.7246", "0.91", "0.735", "0.5215", "0.735" ], [ "0.08826", "0.7363", "0.478", "0.7363", "0.478", "0.7505", "0.08826", "0.7505" ], [ "0.523", "0.738", "0.9087", "0.738", "0.9087", "0.7485", "0.523", "0.7485" ], [ "0.08496", "0.75", "0.478", "0.7505", "0.478", "0.765", "0.08496", "0.7646" ], [ "0.5215", "0.7505", "0.911", "0.7505", "0.911", "0.7646", "0.5215", "0.7646" ], [ "0.08496", "0.765", "0.4795", "0.765", "0.4795", "0.779", "0.08496", "0.779" ], [ "0.5215", "0.7646", "0.914", "0.7646", "0.914", "0.778", "0.5215", "0.778" ], [ "0.0866", "0.779", "0.4763", "0.779", "0.4763", "0.7925", "0.0866", "0.7925" ], [ "0.5195", "0.778", "0.794", "0.778", "0.794", "0.7915", "0.5195", "0.7915" ], [ "0.08496", "0.7925", "0.4795", "0.7925", "0.4795", "0.806", "0.08496", "0.806" ], [ "0.535", "0.7915", "0.9126", "0.7915", "0.9126", "0.8047", "0.535", "0.8047" ], [ "0.08496", "0.806", "0.478", "0.806", "0.478", "0.8193", "0.08496", "0.8193" ], [ "0.5215", "0.806", "0.9126", "0.806", "0.9126", "0.8193", "0.5215", "0.8193" ], [ "0.0866", "0.8213", "0.4756", "0.8213", "0.4756", "0.832", "0.0866", "0.832" ], [ "0.5195", "0.8193", "0.9126", "0.8184", "0.9126", "0.8335", "0.5195", "0.834" ], [ "0.08496", "0.834", "0.478", "0.834", "0.478", "0.8477", "0.08496", "0.8477" ], [ "0.5195", "0.834", "0.9126", "0.834", "0.9126", "0.8477", "0.5195", "0.8477" ], [ "0.0866", "0.8477", "0.4763", "0.8477", "0.4763", "0.8613", "0.0866", "0.8613" ], [ "0.5215", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5215", "0.8613" ], [ "0.0866", "0.8623", "0.4795", "0.8623", "0.4795", "0.8755", "0.0866", "0.8755" ], [ "0.5186", "0.86", "0.91", "0.8613", "0.91", "0.8755", "0.5186", "0.875" ], [ "0.0866", "0.877", "0.165", "0.877", "0.165", "0.888", "0.0866", "0.888" ], [ "0.5195", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5195", "0.8896" ] ]

[ "may not be directed solely at the person he is replying to .", "Content Information: Early work on deceptive content", "the content created could also be applicable to other tweets", "detection studied the use of linguistic cues such as per-", "in the thread. Tree models do not model interactions between", "centage of pronouns, word length, verb quantity, and word", "nodes from other branches explicitly and is a key limitation", "classes(Fuller,Biros,and Wilson 2009:Mihalcea and Strap-", "when modelling social media conversations. For example, in", "parava 2009; Ott et al. 2011; Rubin et al. 2016) on fake re-", "Figure I, tweet R_1 and its replying tweet, R_1.1, have ex-", "views, witness accounts, and satire. The detection of fake", "pressed doubt regarding the factuality of the source claim.", "news using linguistic features have also been studied in", "Tweets R 2.1 and R_3_1 have provided conclusive evidence", "(O'Brien et al. 2018; Wang 2017). Such analysis of de-", "that debunk the source claims as fake. Though R2_1 and", "ceptive content relies on linguistic features which might be", "R_3_1arechildnodesofR_2andR_3respectively,theycould", "unique to domains or topics.", "provide important information to all the other nodes along", "Source and Social Network: Another group of work", "the tree, such as R_1_1 and R_1. Therefore, we should con-", "studied the source of fake news, and its social network.", "siderinteractionsbetweenalltweets,not justthosebetween", "Wang et al.(2017)foundthat adding source information to", "parent and their child, for better aggregation of information", "the content improves fake news classification accuracy. Hu", "from the replying tweets.", "etal.(2013)foundthataccountscreatedtospreadfakenews", "In this paper, we propose to overcome some limitations", "tendtohavedifferentsocialnetworkcharacteristics.", "of tree models in modelling social media conversations as", "Fact Checking: Fact checking websites such as politifact.", "identified from our analysis.More specifically, we flattened", "com and snopes.com rely on manual verification to debunk", "thetree structure and arranged all tweets in a chronological", "fake news, but are unable to match up the rate at which", "order. We propose a post-level attention model (PLAN) that", "fake news are being generated (Funke 2019).Automated fact", "allows all possible pairwise interaction between tweets with", "checking aims to check claims against trustworthy sources", "the self-attention mechanism.To combine the strengths", "such as Wikipedia (Ciampaglia et al. 2015). More recently", "of both self-attention mechanism and tree models, we", "Thorne et al. (2018) proposed the FEVER shared task to ver-", "incorporated structural information and performed structure", "ify an input claim against a database of 5 million Wikipedia", "aware self-attention at a post-level (StA-PLAN). Lastly", "documents,and classifyeachclaimintooneofthreeclasses:", "we designed a structure aware hierarchical token and and", "Supports, Refute or Not Enough Info. Fact checking is a", "post-level attention network (StA-HiTPLAN) to learn more", "more principled approach to fake news detection. However", "complexsentencerepresentationsforeachtweet.", "italso requiresan established corpus of verifiedfacts and", "may not work for new claims with little evidence.", "The contributions of this paper are the following:", "Community Response: Detecting fake news from com-", ". We utilize the attention weights from our model to pro", "munity response is most closely related to our paper. Re-", "vide both token-level and post-level explanations behind", "searchers have worked on automatically predicting the ve-", "the model's prediction. To the best of our knowledge, we", "racity of a claim by building classifiers that leverages on the", "are the first paper that has done this.", "comments and replies to social media posts, as well as the", ". We compare against previous works on two data sets", "propagation pattern (Enayet and El-Beltagy 2017; Castillo,", "PHEME 5 events (Kochkina, Liakata, and Zubiaga 2018)", "Mendoza, and Poblete 201l; Ma, Gao, and Wong 2017;", "and Twitter15 and Twitter16 (Ma, Gao, and Wong 2017).", "2018b).Ma et al. (2018a) adopted a multi-task learning ap-", "Previous works only evaluated on one of the two data sets.", "proach to build a classifier that learns stance aware features", "for rumour detection. Similarly, Li et al. (2019) adopted a", ". Our proposed models could outperform current state-of-", "multi-task learning approach for his models and included", "the-artmodelsforbothdatasets.", "user information in his models. Chen et al. (2017) proposed", "The rest of the paper is organized as follows. In Section", "topool outdistinctivefeaturesthatcapturescontextual varia-", "2, we examine some related work. We define our problem", "tionsofposts overtime.Beyond linguisticfeatures,otherre-", "statement in Section 3. We present our models in Section 4", "searchers have also looked at the demographics of the users", "and results in Section 5. We then conclude with future work", "(Yang et al. 2012; Li, Zhang, and Si 2019) or interaction pe-", "in Section6.", "riodicity (Kwon et al. 2013) to determine the credibility of", "users. Yang et al. (2012) collected the user characteristics", "2Related Work", "who were engagedinspreading fakenewsand used only", "Existing approaches on automatically differentiating fake", "user characteristics to builda fake news detectorby clas.", "from real claims leverage on a variety of features: (i) the", "sifying the propagation path. Li et al. (2019) used a com-", "content of the claim, (i) the bias and social network of", "bination of user information and content features to train a", "the source of the claim, (ii) fact checking with trustwor-", "LSTM with multi-task learning objective.", "thy sources (e.g., Wikipedia), and (iv) community response", "In this paper, we work on detecting rumor and fake news", "to the claims. Our work falls into the last class of detecting", "solely from the post and comments. Unlike Li, Zhang, and", "rumors orfakeclaimsfromcommunityresponse.Inthissec-", "Si; Yang et al. (2019; 2012), we do not use the identities", "tion, we give a brief review of each class of work, focusing", "of user accounts. Most closely related to our work is Ma", "on works that detect fake claims using community response", "et al. (2016; 2017; 2018b) and Kumar et al. (2019). Ma et", "For a more detailed survey, we refer the reader to (Sharma", "al. (2016; 2017; 2018b) used recurrent neural network, prop", "etal.2019).", "agation tree kernels and recursive neural trees to model posts" ]

{ "arxiv_id": "2001.10667", "language": "en", "page_idx": 1 }

[ [ "0.08826", "0.07135", "0.478", "0.07135", "0.478", "0.0821", "0.08826", "0.0821" ], [ "0.5376", "0.06946", "0.9126", "0.06946", "0.9126", "0.0833", "0.5376", "0.0833" ], [ "0.08496", "0.084", "0.4795", "0.084", "0.4795", "0.09784", "0.08496", "0.09784" ], [ "0.5195", "0.084", "0.91", "0.084", "0.91", "0.09784", "0.5195", "0.09784" ], [ "0.0866", "0.09784", "0.4795", "0.09784", "0.4795", "0.1111", "0.0866", "0.1111" ], [ "0.5215", "0.09973", "0.91", "0.09973", "0.91", "0.10986", "0.5215", "0.10986" ], [ "0.08905", "0.113", "0.4756", "0.113", "0.4756", "0.1237", "0.08905", "0.1237" ], [ "0.5215", "0.113", "0.911", "0.113", "0.911", "0.1237", "0.5215", "0.1237" ], [ "0.0866", "0.1256", "0.478", "0.1256", "0.478", "0.1395", "0.0866", "0.1395" ], [ "0.5195", "0.1256", "0.914", "0.1256", "0.914", "0.1395", "0.5195", "0.1395" ], [ "0.08496", "0.1395", "0.4453", "0.1383", "0.4453", "0.1528", "0.08496", "0.154" ], [ "0.5195", "0.1395", "0.9126", "0.1395", "0.9126", "0.1528", "0.5195", "0.1528" ], [ "0.5195", "0.1528", "0.9126", "0.1528", "0.9126", "0.1666", "0.5195", "0.1666" ], [ "0.1896", "0.1666", "0.3718", "0.1686", "0.3718", "0.1863", "0.1896", "0.1843" ], [ "0.5186", "0.1654", "0.9126", "0.1666", "0.9126", "0.1812", "0.5186", "0.1799" ], [ "0.5215", "0.1812", "0.914", "0.1812", "0.914", "0.1945", "0.5215", "0.1945" ], [ "0.08496", "0.1901", "0.478", "0.1901", "0.478", "0.2039", "0.08496", "0.2039" ], [ "0.5186", "0.1932", "0.91", "0.1945", "0.91", "0.2084", "0.5186", "0.207" ], [ "0.08496", "0.2052", "0.478", "0.2052", "0.478", "0.2184", "0.08496", "0.2184" ], [ "0.5186", "0.209", "0.9126", "0.2084", "0.9126", "0.2216", "0.5186", "0.2229" ], [ "0.0866", "0.2197", "0.4763", "0.2197", "0.4763", "0.2329", "0.0866", "0.2329" ], [ "0.5195", "0.2229", "0.6724", "0.2229", "0.6724", "0.2361", "0.5195", "0.2361" ], [ "0.08496", "0.2329", "0.4043", "0.2329", "0.4043", "0.2468", "0.08496", "0.2468" ], [ "0.5215", "0.2468", "0.739", "0.2468", "0.739", "0.26", "0.5215", "0.26" ], [ "0.1805", "0.2551", "0.3506", "0.2568", "0.3506", "0.2727", "0.1805", "0.2708" ], [ "0.5215", "0.2632", "0.9126", "0.2632", "0.9126", "0.2778", "0.5215", "0.2778" ], [ "0.0866", "0.2798", "0.4756", "0.2798", "0.4756", "0.2942", "0.0866", "0.2942" ], [ "0.5215", "0.2798", "0.9126", "0.2798", "0.9126", "0.2903", "0.5215", "0.2903" ], [ "0.08496", "0.2954", "0.4192", "0.2954", "0.4192", "0.3062", "0.08496", "0.3062" ], [ "0.5195", "0.2922", "0.9126", "0.2922", "0.9126", "0.3062", "0.5195", "0.3062" ], [ "0.10297", "0.31", "0.4763", "0.31", "0.4763", "0.324", "0.10297", "0.324" ], [ "0.5195", "0.3062", "0.914", "0.3062", "0.914", "0.3193", "0.5195", "0.3193" ], [ "0.0866", "0.324", "0.478", "0.324", "0.478", "0.3372", "0.0866", "0.3372" ], [ "0.5215", "0.3213", "0.9126", "0.3213", "0.9126", "0.332", "0.5215", "0.332" ], [ "0.08826", "0.339", "0.478", "0.339", "0.478", "0.3499", "0.08826", "0.3499" ], [ "0.5195", "0.3333", "0.911", "0.3333", "0.911", "0.3467", "0.5195", "0.3467" ], [ "0.0866", "0.3516", "0.478", "0.3516", "0.478", "0.3655", "0.0866", "0.3655" ], [ "0.5195", "0.348", "0.9126", "0.348", "0.9126", "0.361", "0.5195", "0.361" ], [ "0.08496", "0.3655", "0.478", "0.3655", "0.478", "0.379", "0.08496", "0.379" ], [ "0.523", "0.3623", "0.911", "0.3623", "0.911", "0.3726", "0.523", "0.3726" ], [ "0.08496", "0.379", "0.4763", "0.379", "0.4763", "0.3926", "0.08496", "0.3926" ], [ "0.5186", "0.3738", "0.911", "0.375", "0.911", "0.3882", "0.5186", "0.387" ], [ "0.0866", "0.3933", "0.4756", "0.3933", "0.4756", "0.4072", "0.0866", "0.4072" ], [ "0.5215", "0.3894", "0.9126", "0.3894", "0.9126", "0.4028", "0.5215", "0.4028" ], [ "0.0866", "0.4072", "0.4795", "0.4072", "0.4795", "0.4204", "0.0866", "0.4204" ], [ "0.5195", "0.4028", "0.91", "0.4028", "0.91", "0.4167", "0.5195", "0.4167" ], [ "0.08496", "0.4204", "0.478", "0.4204", "0.478", "0.4343", "0.08496", "0.4343" ], [ "0.5215", "0.4172", "0.9126", "0.4172", "0.9126", "0.4312", "0.5215", "0.4312" ], [ "0.0866", "0.4343", "0.478", "0.4343", "0.478", "0.449", "0.0866", "0.449" ], [ "0.5215", "0.4312", "0.9126", "0.4312", "0.9126", "0.4443", "0.5215", "0.4443" ], [ "0.08496", "0.449", "0.478", "0.449", "0.478", "0.4622", "0.08496", "0.4622" ], [ "0.523", "0.4458", "0.9087", "0.4458", "0.9087", "0.4558", "0.523", "0.4558" ], [ "0.08496", "0.4622", "0.478", "0.4622", "0.478", "0.476", "0.08496", "0.476" ], [ "0.5215", "0.459", "0.9126", "0.459", "0.9126", "0.473", "0.5215", "0.473" ], [ "0.08496", "0.476", "0.4673", "0.476", "0.4673", "0.4905", "0.08496", "0.4905" ], [ "0.5195", "0.4702", "0.9126", "0.473", "0.9126", "0.4873", "0.5195", "0.4849" ], [ "0.10297", "0.4905", "0.478", "0.4905", "0.478", "0.504", "0.10297", "0.504" ], [ "0.5195", "0.486", "0.903", "0.486", "0.903", "0.5", "0.5195", "0.5" ], [ "0.0866", "0.506", "0.478", "0.506", "0.478", "0.5166", "0.0866", "0.5166" ], [ "0.0866", "0.518", "0.4763", "0.518", "0.4763", "0.532", "0.0866", "0.532" ], [ "0.8203", "0.515", "0.8555", "0.515", "0.8555", "0.532", "0.8203", "0.532" ], [ "0.5474", "0.5244", "0.818", "0.5244", "0.818", "0.539", "0.5474", "0.539" ], [ "0.892", "0.526", "0.911", "0.526", "0.911", "0.5386", "0.892", "0.5386" ], [ "0.0866", "0.532", "0.478", "0.532", "0.478", "0.5454", "0.0866", "0.5454" ], [ "0.8364", "0.5405", "0.847", "0.5405", "0.847", "0.5444", "0.8364", "0.5444" ], [ "0.0866", "0.547", "0.3652", "0.547", "0.3652", "0.56", "0.0866", "0.56" ], [ "0.536", "0.5503", "0.914", "0.5503", "0.914", "0.5645", "0.536", "0.5645" ], [ "0.08496", "0.565", "0.478", "0.5664", "0.478", "0.5806", "0.08496", "0.58" ], [ "0.5195", "0.5645", "0.8677", "0.5645", "0.8677", "0.5776", "0.5195", "0.5776" ], [ "0.10297", "0.5806", "0.3406", "0.5806", "0.3406", "0.5947", "0.10297", "0.5947" ], [ "0.7036", "0.5835", "0.721", "0.5835", "0.721", "0.5923", "0.7036", "0.5923" ], [ "0.6978", "0.5923", "0.767", "0.5947", "0.767", "0.61", "0.6978", "0.608" ], [ "0.8906", "0.5923", "0.914", "0.5923", "0.914", "0.607", "0.8906", "0.607" ], [ "0.6665", "0.5967", "0.6846", "0.5967", "0.6846", "0.605", "0.6665", "0.605" ], [ "0.0866", "0.5947", "0.468", "0.5947", "0.468", "0.608", "0.0866", "0.608" ], [ "0.681", "0.5967", "0.6953", "0.5967", "0.6953", "0.606", "0.681", "0.606" ], [ "0.697", "0.6113", "0.7236", "0.608", "0.726", "0.6187", "0.699", "0.622" ], [ "0.2181", "0.6255", "0.2467", "0.6255", "0.2467", "0.638", "0.2181", "0.638" ], [ "0.2402", "0.6255", "0.3489", "0.6255", "0.3489", "0.6396", "0.2402", "0.6396" ], [ "0.5215", "0.625", "0.9126", "0.625", "0.9126", "0.6396", "0.5215", "0.6396" ], [ "0.5215", "0.6396", "0.819", "0.6396", "0.819", "0.653", "0.5215", "0.653" ], [ "0.08496", "0.646", "0.4795", "0.6445", "0.4795", "0.659", "0.08496", "0.66" ], [ "0.536", "0.654", "0.91", "0.654", "0.91", "0.6675", "0.536", "0.6675" ], [ "0.0866", "0.66", "0.4795", "0.66", "0.4795", "0.674", "0.0866", "0.674" ], [ "0.5195", "0.6675", "0.9126", "0.6675", "0.9126", "0.681", "0.5195", "0.681" ], [ "0.08496", "0.675", "0.478", "0.675", "0.478", "0.688", "0.08496", "0.688" ], [ "0.5195", "0.681", "0.9126", "0.681", "0.9126", "0.6943", "0.5195", "0.6943" ], [ "0.08496", "0.688", "0.4436", "0.688", "0.4436", "0.702", "0.08496", "0.702" ], [ "0.5215", "0.6963", "0.911", "0.6963", "0.911", "0.707", "0.5215", "0.707" ], [ "0.5195", "0.7085", "0.911", "0.709", "0.911", "0.7227", "0.5195", "0.7217" ], [ "0.0866", "0.7163", "0.339", "0.7163", "0.339", "0.73", "0.0866", "0.73" ], [ "0.5195", "0.7227", "0.9126", "0.7227", "0.9126", "0.7363", "0.5195", "0.7363" ], [ "0.08496", "0.735", "0.478", "0.7363", "0.478", "0.7505", "0.08496", "0.75" ], [ "0.5195", "0.7373", "0.914", "0.7373", "0.914", "0.7505", "0.5195", "0.7505" ], [ "0.08417", "0.75", "0.478", "0.7505", "0.478", "0.765", "0.08417", "0.7646" ], [ "0.5195", "0.7505", "0.91", "0.7505", "0.91", "0.7646", "0.5195", "0.7646" ], [ "0.08496", "0.7646", "0.4763", "0.7646", "0.4763", "0.778", "0.08496", "0.778" ], [ "0.5215", "0.7646", "0.9126", "0.7646", "0.9126", "0.778", "0.5215", "0.778" ], [ "0.08496", "0.778", "0.478", "0.779", "0.478", "0.7925", "0.08496", "0.7915" ], [ "0.5215", "0.78", "0.9087", "0.78", "0.9087", "0.7905", "0.5215", "0.7905" ], [ "0.08496", "0.7925", "0.4795", "0.7925", "0.4795", "0.806", "0.08496", "0.806" ], [ "0.523", "0.7935", "0.9126", "0.7935", "0.9126", "0.8037", "0.523", "0.8037" ], [ "0.0866", "0.806", "0.478", "0.806", "0.478", "0.8193", "0.0866", "0.8193" ], [ "0.5215", "0.8066", "0.6167", "0.8066", "0.6167", "0.8184", "0.5215", "0.8184" ], [ "0.08496", "0.821", "0.478", "0.821", "0.478", "0.834", "0.08496", "0.834" ], [ "0.0866", "0.834", "0.478", "0.834", "0.478", "0.8477", "0.0866", "0.8477" ], [ "0.5215", "0.83", "0.853", "0.83", "0.853", "0.8447", "0.5215", "0.8447" ], [ "0.08826", "0.8496", "0.4763", "0.8496", "0.4763", "0.86", "0.08826", "0.86" ], [ "0.5215", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5215", "0.8613" ], [ "0.08496", "0.8623", "0.4795", "0.8623", "0.4795", "0.8755", "0.08496", "0.8755" ], [ "0.5215", "0.863", "0.91", "0.863", "0.91", "0.8735", "0.5215", "0.8735" ], [ "0.08496", "0.8755", "0.2695", "0.8755", "0.2695", "0.8896", "0.08496", "0.8896" ], [ "0.5195", "0.8755", "0.914", "0.8755", "0.914", "0.8896", "0.5195", "0.8896" ] ]

[ "in the same thread either sequentially or witha tree structure", "However, as we will see in the data statistics in Table 1,", "to predict whether the sequence pertains to rumour or real", "the trees in the data sets are very shallow, with the major-", "news. Kumar et al. (2019) used various variants of LSTM", "ity of comments replying directly to the source tweet in-", "(Branch LSTM,Tree LSTM and Binarized Constituency", "stead of replying to other tweets. We found that in social", "Tree LSTM). In this paper, instead of recursive tree models,", "media, as often the entire thread is visible, a user replying", "we propose a transformer network for rumor detection.", "to the root post might be continuing the conversation of ear-", "lier users and not specifically writing a reply to the root post.", "3RumorDetection", "Tree models that do not explicitly model every possible pair-", "wise interaction between tweets are therefore sub-optimal in", "In this section, we first define our problem statement. We", "modelling social media conversations. Therefore, we pro-", "address the problem of predicting the veracity of a claim", "pose to overcome the limitations with our described models", "given all of its responding tweets and the interactions be-", "in subsequent sections.", "tween these tweets. We define each thread to be:", "4.2Transformer Networks", "X=[ri,a2,a3,.,an],", "The attention mechanism in transformer networks enable ef-", "where i is the source tweet, c, the ith tweet in chronologi-", "fective modelingof long-range dependencies (Vaswaniet", "cal order. and n the number of tweets in the thread.", "al. 2017). This is an advantage for rumor detection since", "Besides the textual information, there is also structural in-", "there could be many replying tweets in a conversation and", "formation in the conversation tree which could be exploited", "it is vital to be able to model the interactions between", "(Ma, Gao, and Wong 2018b; Wu, Yang, and Zhu 2015).In", "the tweets effectively. We briefly discuss the multi-head at-", "the tree-structured models, a pair of tweets ri and are", "tention (MHA) layer in the transformer network and refer", "only related if either &, is replying to &, or vice versa. in all", "thereaderto(Vaswani et al.2017)formoredetails.Each", "of our proposed models, we allow any post to attend to any", "MHAlayer in the transformer network is made up of a self-", "other post in the same thread. In our structure aware mod-", "attention sublayer and fully connected feed-forward sub-", "els, we label the relations between any pair of tweets i and", "layer. In the self-attention sublayer, a query and a set of key-", ", with a relation label, R(i, ) e (parent, child, before, af-", "value pair is mapped to an output. The output is a weighted", "ter, self). The value of R(i, j) is obtained by applying the", "sum of the values, where the weight for each value is deter-", "following set of rules in sequence: parent: if r is directly", "minedbythecompatibilitybetweenthequeryandthecorre", "replying i, child: if r, is directly replying ci, before: if ci", "sponding key. The compatibility, Qij, is the attention weight", "comes before &j, after: if ci comes after j, self: if i = j.", "between a query from position i and a key from position j,", "The rumor detection task reduces to learning to predict", "and is calculated with simple scaled dot-product attention:", "each (X,R)toitsrumor class y.We conductedexperiments", "on two rumor detection data sets, namely the Twitter15 and", "qiki", "Qij = Compatibility(qi,k)= softmax", "(1)", "Twitterl6 data, and the PHEME 5 data. The classes are dif-", "", "ferent for the data sets we are working on:", "The output at each position, zi, would then be a weighted", ".for Twitter15 and Twitterl6:y E [non-rumor, false-", "sum of the compatibility and value at other positions.", "rumor, true-rumor, unverifiedj, and", "n", "Laijvj,", "(2)", "2;", ". for PHEME, y E (false-rumor, true-rumor, unverified).", "三", "j=1", "4", "Approaches", "where Qij E [0, 1] is the attention weight from Equation 1.", "A higher value indicates higher compatibility.", "For completeness, we first provide a brief description of", "In order to allow the model to jointly attend to infor-", "recursive neural networks and the attention mechanism in", "mation from different representation subspaces at different", "transformer networks. We then describe the proposed mod-", "positions, Vaswani et al. (2017) introduced the concept of", "els that is primarily based on the attention mechanism.", "multi-head attention.Queries,keys and values are projected", "htimeswithdifferent learned linearprojections.Eachpro-", "4.1Recursive Neural Networks", "jected versions of queries, keys and values would perform", "Ma et al. (2018b) applied tree-structured recursive neural", "the attention function (Equation 2) in parallel, yielding h", "networks (RvNN) to the rumor detection problem: each", "output values. These are concatenated and once again pro-", "thread is represented as a tree where the root is the claim.", "jected, generating the final values. The final layers would", "and each comment is a child to the post it is responding to.", "thenbepassedthroughafullyconnectedfeed-forwardsub", "In RvNN, the input of each node in the tree is a vector of", "layerconsistingoftwolinearlayerswithaRELUactivation", "tf-idf values of the words in the post. Ma et al. (2018b) stud-", "between them.", "ied two models, a bottom-up and a top-down tree models.", "Kumar et al. (2019) applied a similar mechanism.The for-", "4.3 Post-Level Attention Network (PLAN)", "mulationisappealingas weexpectinformationfromdeny-", "The architecture of our post-level attention network (PLAN)", "ing and questioning comments to be propagated and used for", "isshown inFigure 2a.Weflattened the structure ofthe con-", "the classification of rumors.", "versation tree and arranged the tweets chronologically in a" ]

{ "arxiv_id": "2001.10667", "language": "en", "page_idx": 2 }

[ [ "0.1323", "0.0707", "0.1887", "0.0707", "0.1887", "0.0852", "0.1323", "0.0852" ], [ "0.2263", "0.07135", "0.286", "0.07135", "0.286", "0.0833", "0.2263", "0.0833" ], [ "0.3015", "0.07135", "0.3628", "0.07135", "0.3628", "0.0833", "0.3015", "0.0833" ], [ "0.3774", "0.07135", "0.4314", "0.07135", "0.4314", "0.0833", "0.3774", "0.0833" ], [ "0.5195", "0.06946", "0.9126", "0.06946", "0.9126", "0.0833", "0.5195", "0.0833" ], [ "0.2402", "0.0884", "0.2712", "0.0884", "0.2712", "0.1048", "0.2402", "0.1048" ], [ "0.3882", "0.0884", "0.4207", "0.0884", "0.4207", "0.1048", "0.3882", "0.1048" ], [ "0.5195", "0.084", "0.9126", "0.084", "0.9126", "0.09784", "0.5195", "0.09784" ], [ "0.3105", "0.089", "0.3464", "0.084", "0.3506", "0.1029", "0.3145", "0.108" ], [ "0.134", "0.089", "0.2035", "0.089", "0.2035", "0.1042", "0.134", "0.1042" ], [ "0.1316", "0.101", "0.2075", "0.1029", "0.2075", "0.11743", "0.1316", "0.11554" ], [ "0.2385", "0.101", "0.2712", "0.101", "0.2712", "0.11743", "0.2385", "0.11743" ], [ "0.3145", "0.101", "0.3489", "0.101", "0.3489", "0.11743", "0.3145", "0.11743" ], [ "0.3882", "0.1016", "0.4207", "0.1016", "0.4207", "0.11743", "0.3882", "0.11743" ], [ "0.5195", "0.09784", "0.911", "0.09784", "0.911", "0.1111", "0.5195", "0.1111" ], [ "0.3162", "0.11237", "0.3489", "0.11554", "0.3464", "0.1307", "0.3137", "0.1276" ], [ "0.5195", "0.11237", "0.9126", "0.11237", "0.9126", "0.1256", "0.5195", "0.1256" ], [ "0.3882", "0.11426", "0.4207", "0.11426", "0.4207", "0.13", "0.3882", "0.13" ], [ "0.134", "0.11615", "0.2075", "0.11615", "0.2075", "0.1276", "0.134", "0.1276" ], [ "0.2402", "0.113", "0.2712", "0.113", "0.2712", "0.13", "0.2402", "0.13" ], [ "0.3179", "0.127", "0.3464", "0.127", "0.3464", "0.1433", "0.3179", "0.1433" ], [ "0.3882", "0.127", "0.4192", "0.127", "0.4192", "0.1427", "0.3882", "0.1427" ], [ "0.5195", "0.1256", "0.9126", "0.1256", "0.9126", "0.1395", "0.5195", "0.1395" ], [ "0.2418", "0.1276", "0.2695", "0.1276", "0.2695", "0.1427", "0.2418", "0.1427" ], [ "0.134", "0.1288", "0.1691", "0.1288", "0.1691", "0.1414", "0.134", "0.1414" ], [ "0.2418", "0.1395", "0.2695", "0.1395", "0.2695", "0.1559", "0.2418", "0.1559" ], [ "0.3882", "0.1395", "0.4224", "0.1395", "0.4224", "0.1547", "0.3882", "0.1547" ], [ "0.5195", "0.1395", "0.9126", "0.1395", "0.9126", "0.1528", "0.5195", "0.1528" ], [ "0.3186", "0.1401", "0.3447", "0.1401", "0.3447", "0.154", "0.3186", "0.154" ], [ "0.134", "0.1414", "0.1666", "0.1414", "0.1666", "0.154", "0.134", "0.154" ], [ "0.3186", "0.1528", "0.3472", "0.1528", "0.3472", "0.1674", "0.3186", "0.1674" ], [ "0.517", "0.1515", "0.9126", "0.1528", "0.9126", "0.1666", "0.517", "0.1654" ], [ "0.2427", "0.154", "0.2688", "0.154", "0.2688", "0.1674", "0.2427", "0.1674" ], [ "0.134", "0.154", "0.1993", "0.154", "0.1993", "0.1654", "0.134", "0.1654" ], [ "0.3906", "0.154", "0.4167", "0.154", "0.4167", "0.1674", "0.3906", "0.1674" ], [ "0.2427", "0.1654", "0.2688", "0.1654", "0.2688", "0.1787", "0.2427", "0.1787" ], [ "0.3137", "0.1666", "0.343", "0.1622", "0.3472", "0.1787", "0.3179", "0.1818" ], [ "0.5186", "0.1654", "0.914", "0.1654", "0.914", "0.1799", "0.5186", "0.1799" ], [ "0.2385", "0.1768", "0.2737", "0.1768", "0.2737", "0.1925", "0.2385", "0.1925" ], [ "0.134", "0.1674", "0.2035", "0.1674", "0.2035", "0.1787", "0.134", "0.1787" ], [ "0.134", "0.1787", "0.2009", "0.1787", "0.2009", "0.1895", "0.134", "0.1895" ], [ "0.3162", "0.1787", "0.3472", "0.1787", "0.3472", "0.1913", "0.3162", "0.1913" ], [ "0.3882", "0.178", "0.4207", "0.178", "0.4207", "0.1901", "0.3882", "0.1901" ], [ "0.5186", "0.1812", "0.6436", "0.1812", "0.6436", "0.1945", "0.5186", "0.1945" ], [ "0.232", "0.1895", "0.2778", "0.1895", "0.2778", "0.2039", "0.232", "0.2039" ], [ "0.3064", "0.1881", "0.3545", "0.1901", "0.353", "0.2058", "0.3057", "0.2039" ], [ "0.38", "0.1895", "0.4272", "0.1895", "0.4272", "0.2052", "0.38", "0.2052" ], [ "0.5215", "0.1989", "0.898", "0.1989", "0.898", "0.2122", "0.5215", "0.2122" ], [ "0.1356", "0.1913", "0.2091", "0.1913", "0.2091", "0.2026", "0.1356", "0.2026" ], [ "0.5195", "0.2108", "0.911", "0.2122", "0.911", "0.2267", "0.5195", "0.226" ], [ "0.08905", "0.2172", "0.4714", "0.2184", "0.4714", "0.2329", "0.08905", "0.2317" ], [ "0.5376", "0.2279", "0.6685", "0.2279", "0.6685", "0.238", "0.5376", "0.238" ], [ "0.5195", "0.2393", "0.911", "0.2393", "0.911", "0.254", "0.5195", "0.254" ], [ "0.08496", "0.26", "0.478", "0.26", "0.478", "0.2734", "0.08496", "0.2734" ], [ "0.535", "0.2551", "0.7075", "0.254", "0.7075", "0.2651", "0.535", "0.2664" ], [ "0.08496", "0.2747", "0.478", "0.2747", "0.478", "0.2886", "0.08496", "0.2886" ], [ "0.5215", "0.2715", "0.911", "0.2715", "0.911", "0.2854", "0.5215", "0.2854" ], [ "0.08417", "0.2886", "0.326", "0.2874", "0.326", "0.3018", "0.08417", "0.303" ], [ "0.5186", "0.2854", "0.914", "0.286", "0.914", "0.2998", "0.5186", "0.2986" ], [ "0.5195", "0.2998", "0.914", "0.2998", "0.914", "0.3132", "0.5195", "0.3132" ], [ "0.1666", "0.3113", "0.2418", "0.315", "0.2402", "0.3308", "0.165", "0.327" ], [ "0.3423", "0.307", "0.3718", "0.307", "0.3718", "0.3193", "0.3423", "0.3193" ], [ "0.455", "0.3125", "0.4805", "0.3125", "0.4805", "0.33", "0.455", "0.33" ], [ "0.5215", "0.3145", "0.911", "0.3145", "0.911", "0.3245", "0.5215", "0.3245" ], [ "0.255", "0.3208", "0.2615", "0.3208", "0.2615", "0.324", "0.255", "0.324" ], [ "0.2793", "0.3176", "0.304", "0.3176", "0.304", "0.3257", "0.2793", "0.3257" ], [ "0.308", "0.3213", "0.4102", "0.3213", "0.4102", "0.336", "0.308", "0.336" ], [ "0.5215", "0.327", "0.9126", "0.327", "0.9126", "0.3403", "0.5215", "0.3403" ], [ "0.1503", "0.339", "0.2418", "0.3447", "0.2402", "0.3599", "0.1487", "0.3547" ], [ "0.343", "0.3352", "0.3718", "0.3352", "0.3718", "0.348", "0.343", "0.348" ], [ "0.2778", "0.3423", "0.3064", "0.3447", "0.3057", "0.356", "0.2769", "0.3542" ], [ "0.4575", "0.3423", "0.4795", "0.3423", "0.4795", "0.3567", "0.4575", "0.3567" ], [ "0.5195", "0.3403", "0.9126", "0.3416", "0.9126", "0.3547", "0.5195", "0.3542" ], [ "0.2524", "0.348", "0.2615", "0.3467", "0.2615", "0.351", "0.2532", "0.3516" ], [ "0.3064", "0.351", "0.411", "0.348", "0.411", "0.3623", "0.3064", "0.3655" ], [ "0.5195", "0.3547", "0.914", "0.3547", "0.914", "0.3687", "0.5195", "0.3687" ], [ "0.0866", "0.3687", "0.4795", "0.3687", "0.4795", "0.3818", "0.0866", "0.3818" ], [ "0.5195", "0.3687", "0.9126", "0.3687", "0.9126", "0.3818", "0.5195", "0.3818" ], [ "0.0866", "0.3833", "0.478", "0.3833", "0.478", "0.3977", "0.0866", "0.3977" ], [ "0.5215", "0.3833", "0.9126", "0.3833", "0.9126", "0.3965", "0.5215", "0.3965" ], [ "0.08496", "0.3965", "0.4004", "0.3965", "0.4004", "0.4104", "0.08496", "0.4104" ], [ "0.5195", "0.3965", "0.9126", "0.3965", "0.9126", "0.4104", "0.5195", "0.4104" ], [ "0.5215", "0.4124", "0.911", "0.4124", "0.911", "0.4224", "0.5215", "0.4224" ], [ "0.1584", "0.4204", "0.4028", "0.4216", "0.4028", "0.4395", "0.1584", "0.4382" ], [ "0.5215", "0.4248", "0.9126", "0.4248", "0.9126", "0.4382", "0.5215", "0.4382" ], [ "0.0866", "0.4426", "0.4795", "0.4426", "0.4795", "0.4558", "0.0866", "0.4558" ], [ "0.5215", "0.4395", "0.9126", "0.4395", "0.9126", "0.4495", "0.5215", "0.4495" ], [ "0.08496", "0.457", "0.478", "0.457", "0.478", "0.4702", "0.08496", "0.4702" ], [ "0.5195", "0.452", "0.914", "0.452", "0.914", "0.4653", "0.5195", "0.4653" ], [ "0.08496", "0.4702", "0.4656", "0.4702", "0.4656", "0.4841", "0.08496", "0.4841" ], [ "0.5195", "0.4666", "0.914", "0.4666", "0.914", "0.4797", "0.5195", "0.4797" ], [ "0.5195", "0.4797", "0.744", "0.4797", "0.744", "0.4937", "0.5195", "0.4937" ], [ "0.08496", "0.4937", "0.3179", "0.4968", "0.3179", "0.511", "0.08496", "0.5083" ], [ "0.535", "0.4924", "0.914", "0.4912", "0.914", "0.506", "0.535", "0.507" ], [ "0.08496", "0.512", "0.478", "0.513", "0.478", "0.528", "0.08496", "0.5264" ], [ "0.5195", "0.507", "0.911", "0.507", "0.911", "0.521", "0.5195", "0.521" ], [ "0.08417", "0.5264", "0.4763", "0.528", "0.4763", "0.5415", "0.08417", "0.5405" ], [ "0.5215", "0.523", "0.911", "0.523", "0.911", "0.5327", "0.5215", "0.5327" ], [ "0.5195", "0.535", "0.911", "0.535", "0.911", "0.549", "0.5195", "0.549" ], [ "0.08496", "0.5415", "0.4795", "0.5415", "0.4795", "0.5547", "0.08496", "0.5547" ], [ "0.5195", "0.549", "0.911", "0.549", "0.911", "0.5625", "0.5195", "0.5625" ], [ "0.08496", "0.5547", "0.478", "0.5547", "0.478", "0.5684", "0.08496", "0.5684" ], [ "0.5215", "0.5645", "0.6216", "0.5645", "0.6216", "0.5747", "0.5215", "0.5747" ], [ "0.0866", "0.571", "0.4763", "0.571", "0.4763", "0.5806", "0.0866", "0.5806" ], [ "0.08496", "0.5835", "0.4795", "0.5835", "0.4795", "0.5967", "0.08496", "0.5967" ], [ "0.5195", "0.587", "0.7183", "0.587", "0.7183", "0.601", "0.5195", "0.601" ], [ "0.08496", "0.5967", "0.4795", "0.5967", "0.4795", "0.61", "0.08496", "0.61" ], [ "0.5195", "0.605", "0.9126", "0.605", "0.9126", "0.6187", "0.5195", "0.6187" ], [ "0.0866", "0.6123", "0.478", "0.6123", "0.478", "0.6226", "0.0866", "0.6226" ], [ "0.08496", "0.625", "0.4795", "0.625", "0.4795", "0.638", "0.08496", "0.638" ], [ "0.5186", "0.619", "0.9126", "0.6187", "0.9126", "0.632", "0.5186", "0.6333" ], [ "0.08496", "0.638", "0.4795", "0.638", "0.4795", "0.6514", "0.08496", "0.6514" ], [ "0.5215", "0.6353", "0.9126", "0.6353", "0.9126", "0.646", "0.5215", "0.646" ], [ "0.5195", "0.648", "0.9126", "0.648", "0.9126", "0.661", "0.5195", "0.661" ], [ "0.0866", "0.654", "0.4763", "0.654", "0.4763", "0.664", "0.0866", "0.664" ], [ "0.08826", "0.6675", "0.4763", "0.6675", "0.4763", "0.678", "0.08826", "0.678" ], [ "0.5215", "0.662", "0.9126", "0.662", "0.9126", "0.6724", "0.5215", "0.6724" ], [ "0.08496", "0.6797", "0.4763", "0.6797", "0.4763", "0.6934", "0.08496", "0.6934" ], [ "0.5195", "0.675", "0.914", "0.675", "0.914", "0.688", "0.5195", "0.688" ], [ "0.08496", "0.6934", "0.4795", "0.6934", "0.4795", "0.707", "0.08496", "0.707" ], [ "0.5195", "0.6895", "0.9126", "0.6895", "0.9126", "0.7026", "0.5195", "0.7026" ], [ "0.5195", "0.702", "0.7173", "0.702", "0.7173", "0.7153", "0.5195", "0.7153" ], [ "0.08496", "0.7085", "0.4795", "0.7085", "0.4795", "0.7217", "0.08496", "0.7217" ], [ "0.08496", "0.7217", "0.4795", "0.7217", "0.4795", "0.735", "0.08496", "0.735" ], [ "0.5186", "0.7217", "0.9043", "0.7227", "0.9043", "0.7373", "0.5186", "0.7363" ], [ "0.08496", "0.735", "0.478", "0.735", "0.478", "0.7485", "0.08496", "0.7485" ], [ "0.5195", "0.7363", "0.914", "0.7363", "0.914", "0.75", "0.5195", "0.75" ], [ "0.08496", "0.7485", "0.4763", "0.7485", "0.4763", "0.762", "0.08496", "0.762" ], [ "0.5215", "0.752", "0.911", "0.752", "0.911", "0.765", "0.5215", "0.765" ], [ "0.08496", "0.762", "0.4756", "0.762", "0.4756", "0.776", "0.08496", "0.776" ], [ "0.5186", "0.7646", "0.911", "0.765", "0.911", "0.779", "0.5186", "0.778" ], [ "0.5195", "0.779", "0.9126", "0.779", "0.9126", "0.7925", "0.5195", "0.7925" ], [ "0.0866", "0.7886", "0.281", "0.7886", "0.281", "0.803", "0.0866", "0.803" ], [ "0.5195", "0.7925", "0.9126", "0.7925", "0.9126", "0.806", "0.5195", "0.806" ], [ "0.0866", "0.8047", "0.4763", "0.8047", "0.4763", "0.8193", "0.0866", "0.8193" ], [ "0.5215", "0.806", "0.914", "0.806", "0.914", "0.8193", "0.5215", "0.8193" ], [ "0.08496", "0.821", "0.4763", "0.821", "0.4763", "0.834", "0.08496", "0.834" ], [ "0.5215", "0.8213", "0.9126", "0.8213", "0.9126", "0.832", "0.5215", "0.832" ], [ "0.08496", "0.834", "0.4795", "0.834", "0.4795", "0.8477", "0.08496", "0.8477" ], [ "0.5195", "0.834", "0.914", "0.834", "0.914", "0.8477", "0.5195", "0.8477" ], [ "0.0866", "0.8477", "0.478", "0.8477", "0.478", "0.8613", "0.0866", "0.8613" ], [ "0.5215", "0.8496", "0.91", "0.8496", "0.91", "0.86", "0.5215", "0.86" ], [ "0.08417", "0.8613", "0.4795", "0.8623", "0.4795", "0.8755", "0.08417", "0.875" ], [ "0.5215", "0.863", "0.91", "0.863", "0.91", "0.8735", "0.5215", "0.8735" ], [ "0.08496", "0.8755", "0.4795", "0.8755", "0.4795", "0.8896", "0.08496", "0.8896" ], [ "0.5186", "0.8755", "0.914", "0.8755", "0.914", "0.8896", "0.5186", "0.8896" ] ]

[ " Data set", "Twitter15", "Twitter16", "PHEME", "PHEME and Twitter15 and Twitter16. Our model dimension", "2.80", "3.12", "is 300 and the dimension of intermediate output is 600. We", "2.77", "Tree-depth", "Num leaves", "34.2", "31.9", "10.3", "used 12 post-level MHA layers and 2 token-level MHA lay-", "37.9", "ers. For training of the model, we used the ADAM optimizer", "14.9", "Num tweets", "40.6", "172", "393", "with 6000 warm start-up steps. We used an initial learning", "334", "False", "350", "1008", "rate of 0.01 with 0.3 dropout. We used a batch size of 32", "189", "True", "190", "for PLAN and StA-PLAN, and 16 for StA-HiTPLAN due", "358", "Unverified", "571", "371", "205", "to memory limitations of the GPUs used. We compare the", "1413", "Non-rumor", "Total trees", "756", "1972", "following models:", "57,368", "27,652", "29,383", " PLAN: The post-level attention network in Section 4.3.", "Total tweets", ". StA-PLAN: The structure aware post-level attention net-", "Table I: Average tree depths, number of leaves and tweets", "workinSection4.4.", ": StA-HiTPLAN: The structure aware hierarchical atten-", "encode each time bin. The time delay embedding would be", "tion model in Section 4.5.", "added to the sentence embedding of tweet. The time delay", "We also investigate these models with and without the time-", "embedding, TDE, for each tweet is:", "delayembeddingas described in Section4.6.However,as", "time delay information did not improve upon PLAN and", "TDEpos,2i", "pos", "(8)", "StA-PLANforTwitterl5andTwitterl6.wedidnotrunex-", "-", "sin", "100002i/dmodel", "periments for StA-HiTPLAN + time delay for Twitter15 and", "TDEpos,2i+1", "pos", "cos", "(9)", "Twitter16.For Twitter15 and Twitter 16, we compared our", "-", "100002i/dmodel", "proposed models with RvNN models proposed by Ma et", "where pos represents the time bin each tweet fall into and", "al. (2018b). As we are using different preprocessing steps", "pos [0, 100), i refers to the dimension and dmodel refers", "(described in 5.1) for Twitter15 and Twitter16, we retrained", "to the total number of dimensions of the model.", "the RvNN models with our implementation. We report both", "the originalresults and resultsfrom ourimplementation.For", "5Experiments and Results", "PHEME, we compare against the LSTM models proposed", "We evaluate our model based on two rumour detection data", "byKumaret al.(2019).As therewere several combination", "sets - (i) Twitter15 and Twitter16 and (ii) PHEME 5 events", "of embeddings and models in (Kumar and Carley 2019), we", "data set. We show the statistics of the data sets in Table 1.", "compare our results with the best results reported for each", "variant of LSTM model proposed.", "5.1 Data and Pre-processing", "We summarize our experimental results in Tables 2 and", "For the PHEME 5 data set, we follow the experimental set-", "3. For Twitter 15 and Twitter 16, all our models outper-", "ting of Kumar and Carley (2019) in using event wise cross-", "form tree recursive neural networks,and our best models", "outperform by 14.2% for Twitter15 and 6.8% for Twitter16.", "validation. For the Twitter15 and Twitter 16 data sets, there", "For PHEME, our best model outperforms previous work by", "is a large proportion of retweets in each claim: 89% for Twit-", "1.6% F1-score.", "ter15and90%forTwitter16.Asweassumethatretweets do", "not contribute new information to the model, we removed", "5.3Discussion of results", "all retweets for Twitter15 and Twitter16. After the removal", "In the rest of this section, we analyze the performance of our", "ofretweets,weobservethatasmallnumberofclaimswould", "be left with only the source tweet. Since the principle behind", "proposed models on Twitter15,Twitter16 and PHEME.We", "our methodology is that we could exploit crowd signals for", "found conflicting conclusions forthe differentdatasets.In", "particular, although all of the datasets are similar in nature,", "rumordetection,claims withoutanyrepliesshould then be", "\"unverified\".Hence,we amended thelabels for such claims", "thestate-of-the-artperformanceofPHEMEisfarworsethan", "to be “unverified\" in the training data (1.49% of Twitter15", "that of Twitter15 and Twitter16. As such, we also provide", "8 False, 10 True and 3 Non-Rumour. 1.46% of Twitter16 - 9", "analysis suggesting possible explanations for this disparity", "in results for the two datasets.", "False, 2 True and O Non-Rumour). In order for our results to", "be comparable with previous work, we excluded such claims", "Structural Information: We proposed two methods", "from our testing set. We used the original splits released", "StA-PLAN and time-delay embeddings that aim to capture", "by (Ma, Gao, and Wong 2018b) to split our data. We show", "structural and chronological information among posts. StA-", "the statistics of the data sets after pre-processing in Table 1.", "PLAN is the best performerforTwitterl5but did not out-", "perform PLAN for Twitterl6, though results were not sub-", "5.2Experimental Setup", "stantially different. The reason structural aware model only", "In all experiments, we used the GLOVE 30Od (Pennington,", "works for Twitter15 might be because the Twitter15 data set", "Socher, and Manning 2014) embedding to represent each to-", "is substantially bigger (interms of number of tweets,see", "ken in a tweet. (Preliminary experiments with BERT (Devlin", "Table 1) than both PHEME and Twitter16. A big data set", "et al. 2018) did not improve results, and were computation-", "mightbenecessarytoexploitthecomplicated structuralin-", "ally far more expensive thanGLOVE).We used the same", "formation in the structure aware model.Time-delay infor-", "set of hyper parameters in all of our experiments for both", "mation was useful for PHEME, where all proposed models" ]

{ "arxiv_id": "2001.10667", "language": "en", "page_idx": 4 }

[ [ "0.12427", "0.0831", "0.8774", "0.0831", "0.8774", "0.0974", "0.12427", "0.0974" ], [ "0.2896", "0.10156", "0.7095", "0.1027", "0.7095", "0.11755", "0.2896", "0.1164" ], [ "0.1453", "0.1484", "0.8623", "0.1484", "0.8623", "0.1633", "0.1453", "0.1633" ], [ "0.4055", "0.1663", "0.602", "0.1663", "0.602", "0.1799", "0.4055", "0.1799" ], [ "0.4417", "0.1818", "0.567", "0.1818", "0.567", "0.1965", "0.4417", "0.1965" ], [ "0.2065", "0.199", "0.804", "0.199", "0.804", "0.2131", "0.2065", "0.2131" ], [ "0.266", "0.2642", "0.3435", "0.2642", "0.3435", "0.279", "0.266", "0.279" ], [ "0.5146", "0.2666", "0.882", "0.2666", "0.882", "0.2798", "0.5146", "0.2798" ], [ "0.5146", "0.282", "0.882", "0.282", "0.882", "0.2957", "0.5146", "0.2957" ], [ "0.1486", "0.2903", "0.4592", "0.2903", "0.4592", "0.304", "0.1486", "0.304" ], [ "0.5146", "0.2986", "0.881", "0.2986", "0.881", "0.3088", "0.5146", "0.3088" ], [ "0.02184", "0.31", "0.05878", "0.31", "0.06213", "0.4382", "0.02519", "0.4382" ], [ "0.147", "0.3047", "0.461", "0.3047", "0.461", "0.3184", "0.147", "0.3184" ], [ "0.1453", "0.3184", "0.4592", "0.3196", "0.4592", "0.3333", "0.1453", "0.332" ], [ "0.513", "0.3142", "0.8833", "0.3142", "0.8833", "0.3289", "0.513", "0.3289" ], [ "0.147", "0.3342", "0.4575", "0.3342", "0.4575", "0.3474", "0.147", "0.3474" ], [ "0.5146", "0.33", "0.6094", "0.33", "0.6094", "0.3416", "0.5146", "0.3416" ], [ "0.147", "0.3474", "0.461", "0.3474", "0.461", "0.361", "0.147", "0.361" ], [ "0.531", "0.3467", "0.885", "0.3467", "0.885", "0.3599", "0.531", "0.3599" ], [ "0.147", "0.3623", "0.461", "0.3623", "0.461", "0.376", "0.147", "0.376" ], [ "0.513", "0.3635", "0.8867", "0.3635", "0.8867", "0.3772", "0.513", "0.3772" ], [ "0.147", "0.376", "0.4592", "0.376", "0.4592", "0.3896", "0.147", "0.3896" ], [ "0.511", "0.3777", "0.882", "0.379", "0.882", "0.3938", "0.511", "0.3926" ], [ "0.147", "0.3901", "0.4592", "0.3901", "0.4592", "0.4038", "0.147", "0.4038" ], [ "0.5146", "0.3955", "0.885", "0.3955", "0.885", "0.4092", "0.5146", "0.4092" ], [ "0.1486", "0.405", "0.4592", "0.405", "0.4592", "0.4187", "0.1486", "0.4187" ], [ "0.513", "0.411", "0.881", "0.411", "0.881", "0.4246", "0.513", "0.4246" ], [ "0.147", "0.4187", "0.4592", "0.4187", "0.4592", "0.4316", "0.147", "0.4316" ], [ "0.1503", "0.434", "0.4568", "0.434", "0.4568", "0.4448", "0.1503", "0.4448" ], [ "0.513", "0.4287", "0.879", "0.4287", "0.879", "0.4395", "0.513", "0.4395" ], [ "0.02519", "0.437", "0.06213", "0.437", "0.06213", "0.517", "0.02519", "0.517" ], [ "0.1486", "0.4478", "0.461", "0.4478", "0.461", "0.4614", "0.1486", "0.4614" ], [ "0.5146", "0.4453", "0.882", "0.4453", "0.882", "0.456", "0.5146", "0.456" ], [ "0.1486", "0.4614", "0.4592", "0.4614", "0.4592", "0.4744", "0.1486", "0.4744" ], [ "0.5146", "0.4614", "0.881", "0.4614", "0.881", "0.4714", "0.5146", "0.4714" ], [ "0.1486", "0.4756", "0.4592", "0.4756", "0.4592", "0.4893", "0.1486", "0.4893" ], [ "0.5146", "0.4756", "0.882", "0.4756", "0.882", "0.4905", "0.5146", "0.4905" ], [ "0.1486", "0.4912", "0.4592", "0.4912", "0.4592", "0.502", "0.1486", "0.502" ], [ "0.513", "0.4905", "0.882", "0.4912", "0.882", "0.506", "0.513", "0.505" ], [ "0.1453", "0.5044", "0.4592", "0.503", "0.4592", "0.517", "0.1453", "0.5186" ], [ "0.513", "0.5073", "0.879", "0.5073", "0.879", "0.5215", "0.513", "0.5215" ], [ "0.02184", "0.5186", "0.05792", "0.5186", "0.06046", "0.728", "0.02519", "0.728" ], [ "0.147", "0.5186", "0.4592", "0.5186", "0.4592", "0.532", "0.147", "0.532" ], [ "0.5146", "0.524", "0.8833", "0.524", "0.8833", "0.538", "0.5146", "0.538" ], [ "0.1453", "0.532", "0.3167", "0.531", "0.3167", "0.546", "0.1453", "0.5464" ], [ "0.5146", "0.5415", "0.5728", "0.5415", "0.5728", "0.5527", "0.5146", "0.5527" ], [ "0.1276", "0.557", "0.2612", "0.557", "0.2612", "0.572", "0.1276", "0.572" ], [ "0.531", "0.5547", "0.8833", "0.5547", "0.8833", "0.5684", "0.531", "0.5684" ], [ "0.126", "0.561", "0.1301", "0.561", "0.1301", "0.5664", "0.126", "0.5664" ], [ "0.5146", "0.5723", "0.8833", "0.5723", "0.8833", "0.586", "0.5146", "0.586" ], [ "0.12006", "0.583", "0.4895", "0.583", "0.4895", "0.5967", "0.12006", "0.5967" ], [ "0.511", "0.5874", "0.8833", "0.5884", "0.8833", "0.6025", "0.511", "0.6016" ], [ "0.12177", "0.601", "0.4895", "0.601", "0.4895", "0.611", "0.12177", "0.611" ], [ "0.5146", "0.604", "0.885", "0.604", "0.885", "0.6177", "0.5146", "0.6177" ], [ "0.12006", "0.615", "0.4895", "0.615", "0.4895", "0.629", "0.12006", "0.629" ], [ "0.5146", "0.6206", "0.885", "0.6206", "0.885", "0.6343", "0.5146", "0.6343" ], [ "0.1184", "0.63", "0.4878", "0.63", "0.4878", "0.6445", "0.1184", "0.6445" ], [ "0.513", "0.6353", "0.882", "0.6357", "0.882", "0.651", "0.513", "0.6494" ], [ "0.1184", "0.6465", "0.4895", "0.6455", "0.4895", "0.66", "0.1184", "0.661" ], [ "0.513", "0.653", "0.7886", "0.653", "0.7886", "0.6665", "0.513", "0.6665" ], [ "0.1184", "0.663", "0.4895", "0.663", "0.4895", "0.6777", "0.1184", "0.6777" ], [ "0.529", "0.6694", "0.882", "0.6694", "0.882", "0.683", "0.529", "0.683" ], [ "0.11755", "0.6787", "0.486", "0.6787", "0.486", "0.6924", "0.11755", "0.6924" ], [ "0.513", "0.684", "0.881", "0.684", "0.881", "0.6978", "0.513", "0.6978" ], [ "0.1184", "0.6953", "0.4878", "0.6934", "0.4878", "0.708", "0.1184", "0.7104" ], [ "0.5146", "0.7017", "0.882", "0.7017", "0.882", "0.7153", "0.5146", "0.7153" ], [ "0.1184", "0.712", "0.4895", "0.712", "0.4895", "0.7256", "0.1184", "0.7256" ], [ "0.5146", "0.7173", "0.882", "0.7173", "0.882", "0.7324", "0.5146", "0.7324" ], [ "0.1184", "0.728", "0.4895", "0.727", "0.4895", "0.741", "0.1184", "0.742" ], [ "0.5146", "0.7354", "0.881", "0.7354", "0.881", "0.745", "0.5146", "0.745" ], [ "0.12006", "0.7446", "0.486", "0.7446", "0.486", "0.7583", "0.12006", "0.7583" ], [ "0.511", "0.749", "0.881", "0.749", "0.881", "0.7627", "0.511", "0.7627" ], [ "0.12006", "0.7603", "0.4878", "0.7603", "0.4878", "0.774", "0.12006", "0.774" ], [ "0.513", "0.7656", "0.882", "0.7656", "0.882", "0.7793", "0.513", "0.7793" ], [ "0.12006", "0.777", "0.4895", "0.777", "0.4895", "0.7905", "0.12006", "0.7905" ], [ "0.513", "0.782", "0.8833", "0.782", "0.8833", "0.796", "0.513", "0.796" ], [ "0.1184", "0.792", "0.3105", "0.792", "0.3105", "0.8057", "0.1184", "0.8057" ], [ "0.5146", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.5146", "0.8125" ], [ "0.1351", "0.807", "0.4902", "0.808", "0.4902", "0.8223", "0.1351", "0.822" ], [ "0.5146", "0.814", "0.882", "0.814", "0.882", "0.8276", "0.5146", "0.8276" ], [ "0.12006", "0.8247", "0.4895", "0.8247", "0.4895", "0.8384", "0.12006", "0.8384" ], [ "0.513", "0.8306", "0.8833", "0.8306", "0.8833", "0.844", "0.513", "0.844" ], [ "0.12006", "0.8403", "0.4902", "0.8403", "0.4902", "0.854", "0.12006", "0.854" ], [ "0.511", "0.8467", "0.881", "0.8467", "0.881", "0.86", "0.511", "0.86" ], [ "0.1184", "0.857", "0.4895", "0.857", "0.4895", "0.8706", "0.1184", "0.8706" ], [ "0.1184", "0.8735", "0.4895", "0.8735", "0.4895", "0.887", "0.1184", "0.887" ], [ "0.532", "0.868", "0.8486", "0.8706", "0.8486", "0.8853", "0.532", "0.883" ], [ "0.511", "0.8813", "0.6245", "0.8843", "0.6245", "0.898", "0.511", "0.8945" ], [ "0.1411", "0.8955", "0.3105", "0.8955", "0.3105", "0.909", "0.1411", "0.909" ], [ "0.534", "0.8945", "0.7773", "0.8955", "0.7773", "0.909", "0.534", "0.9077" ] ]

[ "Identification of Tasks, Datasets. Evaluation Metrics. and Numeric Scores", "for Scientific Leaderboards Construction", "Yufang Hou, Charles Jochim, Martin Gleize, Francesca Bonin and Debasis Ganguly", "IBM Research - Ireland", "Dublin, Ireland", "(yhoulcharlesjImgleizelfboninldebasgalj@ie.ibm.com", "Abstract", "dedicated to analysis of reproducibility in experi-", "ments (Ferro et al., 2018), or by detailed compar-", "While the fast-paced inception of novel tasks", "ativeanalvsis of experimentalresults reportedon", "610 n ", "and new datasets helps foster active research", "in a community towards interesting directions,", "the same dataset in published papers (Armstrong", "keeping track of the abundance of research ac-", "et al.. 2009).", "tivity in different areas on different datasets is", "A useful output of this meta-analysis is often", "likely to become increasingly difficult. The", "a summary of the results of a comparable set of", "community could greatly benefit from an auto-", "experiments (in terms of the tasks they are ap-", "matic system able to summarize scientific re-", "plied on, the datasets on which they are tested", "sults, e.g., in the form of a leaderboard. In", "and the metrics used for evaluation) in a tabu-", "this paper we build two datasets and develop", "aframework(TDMS-IE)aimedatautomati-", "larform,commonlyreferredtoasaleaderboard", "[T'S]", "cally extracting task, dataset, metric and score", "Such a meta-analysis summary in the form of a", "from NLP papers, towards the automatic con-", "leaderboardispotentiallyusefultoresearchersfor", "struction of leaderboards. Experiments show", "the purpose of (1) choosing the appropriate exist-", "that ourmodeloutperforms several baselines", "ing literature for fair comparisons against a newly", "by a large margin. Our model is a first step", "proposed method; and (2) selecting strong base", "1A069061", "towards automatic leaderboard construction,", "lines, which the new method should be compared", "e.g., in the NLP domain.", "against.", "Introduction", "Although recently there has been some effort", "", "to manually keep an account of progress on vari-", "Recent years have witnessed a significant in-", "ous research fields in the form of leaderboards, ei-", "creaseinthenumberoflaboratory-basedevalua-", "ther by individual researchers? or in a moderated", "tion benchmarks in many of scientific disciplines,", "crowd-sourced environment by organizations3, it", "e.g., in the year 2018 alone, 140,616 papers were", "is likely to become increasingly difficult and time-", "submitted to the pre-print repository arXivl and", "consuming over the passage of time.", "among them, 3,710 papers are under the Com-", "In this paper, we develop a model to auto-", "puter Science - Computation and Language cat-", "matically identify tasks,datasets, evaluation met-", "egory. This massive increase in evaluation bench-", "rics, and to extract the corresponding best numeric", "marks (e.g., in the form of shared tasks) is par-", "scores from experimental scientific papers. An il-", "ticularly true for an empirical field such as NLP,", "lustrativeexampleisshowninFigurel:giventhe", "which strongly encourages the research commu-", " sample paper shown on the left, which carries out", "nity to develop a set of publicly available bench-", "research work on three different tasks (i.e., coref-", "mark tasks, datasets and tools so as to reinforce", "erence resolution, named entity recognition, and", "reproducibleexperiments.", "entity linking), the system is supposed to extract", "Researchers have realized the importance of", "the corresponding Task-Dataset-Metric-Score tu-", "conducting meta-analysis of a number of compa-", "ples as shown on the right part in Figure l. It is", "rable publications, i.e., the ones which use similar,", "noteworthy that we aim to identify a set of pre-", "if not identical, experimental settings, from shared", "tasks and proceedings, as shown by special issues", "2https://github.com/sebastianruder/", "NLP-progress", "lhttps://arxiv.org/", "3https://paperswithcode.com" ]

{ "arxiv_id": "1906.09317", "language": "en", "page_idx": 0 }

[ [ "0.6265", "0.0736", "0.681", "0.0736", "0.681", "0.08435", "0.6265", "0.08435" ], [ "0.6978", "0.0736", "0.754", "0.0736", "0.754", "0.08435", "0.6978", "0.08435" ], [ "0.7715", "0.0736", "0.8345", "0.0736", "0.8345", "0.08435", "0.7715", "0.08435" ], [ "0.638", "0.0861", "0.671", "0.0861", "0.671", "0.0974", "0.638", "0.0974" ], [ "0.711", "0.0861", "0.744", "0.0861", "0.744", "0.0974", "0.711", "0.0974" ], [ "0.7866", "0.0861", "0.818", "0.0861", "0.818", "0.0986", "0.7866", "0.0986" ], [ "0.1696", "0.0968", "0.4324", "0.0968", "0.4324", "0.10986", "0.1696", "0.10986" ], [ "0.1696", "0.0873", "0.2551", "0.0873", "0.2551", "0.0974", "0.1696", "0.0974" ], [ "0.638", "0.0968", "0.669", "0.0968", "0.669", "0.10925", "0.638", "0.10925" ], [ "0.711", "0.0968", "0.742", "0.0968", "0.742", "0.10925", "0.711", "0.10925" ], [ "0.7866", "0.0968", "0.818", "0.0968", "0.818", "0.10925", "0.7866", "0.10925" ], [ "0.1696", "0.10925", "0.4375", "0.10925", "0.4375", "0.1223", "0.1696", "0.1223" ], [ "0.638", "0.10925", "0.671", "0.10925", "0.671", "0.1217", "0.638", "0.1217" ], [ "0.711", "0.10925", "0.744", "0.10925", "0.744", "0.1217", "0.711", "0.1217" ], [ "0.785", "0.10925", "0.818", "0.10925", "0.818", "0.1217", "0.785", "0.1217" ], [ "0.711", "0.12054", "0.742", "0.12054", "0.742", "0.133", "0.711", "0.133" ], [ "0.7866", "0.12054", "0.818", "0.12054", "0.818", "0.133", "0.7866", "0.133" ], [ "0.638", "0.1217", "0.671", "0.1217", "0.671", "0.133", "0.638", "0.133" ], [ "0.1704", "0.1223", "0.4592", "0.1223", "0.4592", "0.133", "0.1704", "0.133" ], [ "0.1663", "0.1318", "0.6113", "0.1318", "0.6113", "0.1455", "0.1663", "0.1455" ], [ "0.638", "0.133", "0.671", "0.133", "0.671", "0.1455", "0.638", "0.1455" ], [ "0.7095", "0.133", "0.742", "0.133", "0.742", "0.1443", "0.7095", "0.1443" ], [ "0.7866", "0.1312", "0.8193", "0.133", "0.818", "0.1467", "0.785", "0.1443" ], [ "0.1604", "0.1633", "0.8403", "0.1633", "0.8403", "0.1776", "0.1604", "0.1776" ], [ "0.1184", "0.2037", "0.4902", "0.2037", "0.4902", "0.2173", "0.1184", "0.2173" ], [ "0.513", "0.2037", "0.882", "0.2037", "0.882", "0.2173", "0.513", "0.2173" ], [ "0.1184", "0.2203", "0.4937", "0.2203", "0.4937", "0.234", "0.1184", "0.234" ], [ "0.511", "0.2203", "0.882", "0.2203", "0.882", "0.234", "0.511", "0.234" ], [ "0.1184", "0.2363", "0.486", "0.2363", "0.486", "0.2494", "0.1184", "0.2494" ], [ "0.5156", "0.2517", "0.529", "0.2517", "0.529", "0.26", "0.5156", "0.26" ], [ "0.5425", "0.2494", "0.729", "0.2494", "0.729", "0.2642", "0.5425", "0.2642" ], [ "0.1184", "0.2673", "0.4878", "0.2673", "0.4878", "0.2808", "0.1184", "0.2808" ], [ "0.1184", "0.253", "0.4895", "0.2505", "0.4895", "0.2642", "0.1184", "0.2666" ], [ "0.513", "0.275", "0.885", "0.2756", "0.885", "0.2903", "0.513", "0.2893" ], [ "0.12006", "0.2852", "0.4895", "0.2852", "0.4895", "0.2986", "0.12006", "0.2986" ], [ "0.513", "0.2922", "0.8833", "0.2922", "0.8833", "0.306", "0.513", "0.306" ], [ "0.12006", "0.3018", "0.4878", "0.3018", "0.4878", "0.3154", "0.12006", "0.3154" ], [ "0.513", "0.3088", "0.8833", "0.3088", "0.8833", "0.3225", "0.513", "0.3225" ], [ "0.12006", "0.3176", "0.4878", "0.3176", "0.4878", "0.3308", "0.12006", "0.3308" ], [ "0.513", "0.3247", "0.8833", "0.3247", "0.8833", "0.339", "0.513", "0.339" ], [ "0.12006", "0.335", "0.4895", "0.335", "0.4895", "0.3457", "0.12006", "0.3457" ], [ "0.511", "0.3403", "0.8833", "0.3403", "0.8833", "0.354", "0.511", "0.354" ], [ "0.1184", "0.3499", "0.4895", "0.3486", "0.4895", "0.3623", "0.1184", "0.3635" ], [ "0.513", "0.357", "0.885", "0.357", "0.885", "0.3706", "0.513", "0.3706" ], [ "0.12006", "0.3652", "0.4895", "0.3652", "0.4895", "0.379", "0.12006", "0.379" ], [ "0.513", "0.3735", "0.8833", "0.3735", "0.8833", "0.3872", "0.513", "0.3872" ], [ "0.12006", "0.3818", "0.4878", "0.3818", "0.4878", "0.3955", "0.12006", "0.3955" ], [ "0.511", "0.3884", "0.8833", "0.3896", "0.8833", "0.4038", "0.511", "0.4026" ], [ "0.1184", "0.398", "0.4895", "0.398", "0.4895", "0.411", "0.1184", "0.411" ], [ "0.513", "0.405", "0.885", "0.405", "0.885", "0.4192", "0.513", "0.4192" ], [ "0.1184", "0.4133", "0.4878", "0.4133", "0.4878", "0.4275", "0.1184", "0.4275" ], [ "0.513", "0.4216", "0.885", "0.4216", "0.885", "0.4353", "0.513", "0.4353" ], [ "0.11755", "0.43", "0.4878", "0.43", "0.4878", "0.4436", "0.11755", "0.4436" ], [ "0.513", "0.437", "0.885", "0.437", "0.885", "0.452", "0.513", "0.452" ], [ "0.11755", "0.4448", "0.4895", "0.4453", "0.4895", "0.4602", "0.11755", "0.459" ], [ "0.5146", "0.4536", "0.8833", "0.4536", "0.8833", "0.4673", "0.5146", "0.4673" ], [ "0.11755", "0.4614", "0.4895", "0.462", "0.4895", "0.4768", "0.11755", "0.4756" ], [ "0.513", "0.4697", "0.8833", "0.4697", "0.8833", "0.4827", "0.513", "0.4827" ], [ "0.12177", "0.4797", "0.4878", "0.4797", "0.4878", "0.4905", "0.12177", "0.4905" ], [ "0.513", "0.4863", "0.882", "0.4863", "0.882", "0.4995", "0.513", "0.4995" ], [ "0.1184", "0.4946", "0.4878", "0.4946", "0.4878", "0.5083", "0.1184", "0.5083" ], [ "0.513", "0.5005", "0.8833", "0.502", "0.8833", "0.5166", "0.513", "0.5156" ], [ "0.11755", "0.511", "0.486", "0.511", "0.486", "0.525", "0.11755", "0.525" ], [ "0.513", "0.5186", "0.882", "0.5186", "0.882", "0.532", "0.513", "0.532" ], [ "0.12006", "0.527", "0.4878", "0.527", "0.4878", "0.5405", "0.12006", "0.5405" ], [ "0.513", "0.5337", "0.8833", "0.5337", "0.8833", "0.5474", "0.513", "0.5474" ], [ "0.1184", "0.542", "0.2854", "0.542", "0.2854", "0.5557", "0.1184", "0.5557" ], [ "0.511", "0.549", "0.882", "0.549", "0.882", "0.5625", "0.511", "0.5625" ], [ "0.5146", "0.5654", "0.6943", "0.5654", "0.6943", "0.579", "0.5146", "0.579" ], [ "0.1368", "0.5723", "0.4895", "0.5723", "0.4895", "0.586", "0.1368", "0.586" ], [ "0.531", "0.5806", "0.8833", "0.583", "0.8833", "0.5967", "0.531", "0.594" ], [ "0.11755", "0.589", "0.4878", "0.5884", "0.4878", "0.6025", "0.11755", "0.604" ], [ "0.1184", "0.604", "0.4878", "0.604", "0.4878", "0.618", "0.1184", "0.618" ], [ "0.513", "0.5986", "0.882", "0.5986", "0.882", "0.6133", "0.513", "0.6133" ], [ "0.1184", "0.619", "0.4878", "0.619", "0.4878", "0.633", "0.1184", "0.633" ], [ "0.513", "0.615", "0.8833", "0.615", "0.8833", "0.629", "0.513", "0.629" ], [ "0.513", "0.6313", "0.8833", "0.6313", "0.8833", "0.6445", "0.513", "0.6445" ], [ "0.12006", "0.637", "0.4895", "0.637", "0.4895", "0.651", "0.12006", "0.651" ], [ "0.513", "0.6455", "0.885", "0.6465", "0.885", "0.661", "0.513", "0.66" ], [ "0.11755", "0.652", "0.4878", "0.653", "0.4878", "0.6675", "0.11755", "0.6665" ], [ "0.511", "0.662", "0.885", "0.663", "0.885", "0.6777", "0.511", "0.677" ], [ "0.12006", "0.6694", "0.486", "0.6694", "0.486", "0.683", "0.12006", "0.683" ], [ "0.511", "0.6777", "0.885", "0.6787", "0.885", "0.6934", "0.511", "0.6924" ], [ "0.12006", "0.686", "0.4895", "0.686", "0.4895", "0.6997", "0.12006", "0.6997" ], [ "0.1184", "0.7007", "0.4878", "0.7017", "0.4878", "0.7163", "0.1184", "0.7153" ], [ "0.5146", "0.6953", "0.8833", "0.6953", "0.8833", "0.7104", "0.5146", "0.7104" ], [ "0.5146", "0.7114", "0.8833", "0.7114", "0.8833", "0.7256", "0.5146", "0.7256" ], [ "0.12177", "0.7197", "0.4878", "0.7197", "0.4878", "0.73", "0.12177", "0.73" ], [ "0.513", "0.728", "0.8833", "0.728", "0.8833", "0.741", "0.513", "0.741" ], [ "0.12006", "0.734", "0.4878", "0.734", "0.4878", "0.7476", "0.12006", "0.7476" ], [ "0.511", "0.742", "0.636", "0.7437", "0.636", "0.7583", "0.511", "0.7573" ], [ "0.12006", "0.752", "0.486", "0.752", "0.486", "0.7627", "0.12006", "0.7627" ], [ "0.531", "0.7603", "0.885", "0.7603", "0.885", "0.774", "0.531", "0.774" ], [ "0.12006", "0.7666", "0.4895", "0.7666", "0.4895", "0.78", "0.12006", "0.78" ], [ "0.1184", "0.781", "0.486", "0.781", "0.486", "0.7944", "0.1184", "0.7944" ], [ "0.513", "0.777", "0.6753", "0.777", "0.6753", "0.7905", "0.513", "0.7905" ], [ "0.11755", "0.799", "0.4878", "0.799", "0.4878", "0.8125", "0.11755", "0.8125" ], [ "0.513", "0.804", "0.6543", "0.804", "0.6543", "0.8184", "0.513", "0.8184" ], [ "0.11755", "0.814", "0.486", "0.814", "0.486", "0.8276", "0.11755", "0.8276" ], [ "0.1184", "0.8306", "0.4895", "0.83", "0.4895", "0.844", "0.1184", "0.8457" ], [ "0.513", "0.826", "0.882", "0.826", "0.882", "0.8403", "0.513", "0.8403" ], [ "0.11755", "0.8467", "0.492", "0.8467", "0.492", "0.861", "0.11755", "0.861" ], [ "0.513", "0.843", "0.882", "0.843", "0.882", "0.8555", "0.513", "0.8555" ], [ "0.511", "0.857", "0.885", "0.858", "0.885", "0.8726", "0.511", "0.8716" ], [ "0.1184", "0.8633", "0.4902", "0.8633", "0.4902", "0.877", "0.1184", "0.877" ], [ "0.11755", "0.8774", "0.4902", "0.879", "0.4902", "0.8936", "0.11755", "0.8926" ], [ "0.532", "0.882", "0.8506", "0.883", "0.8506", "0.898", "0.532", "0.8965" ], [ "0.1184", "0.8945", "0.4895", "0.8955", "0.4895", "0.91", "0.1184", "0.909" ], [ "0.513", "0.898", "0.6274", "0.898", "0.6274", "0.91", "0.513", "0.91" ] ]

[ "MacroP", "Macro R", "Macro F1", "79.2", "87.0", "82.6", "Numeric value + IsBolded + Table caption", "Tablecaption", "71.1", "77.7", "74.0", "Numeric value + Row label+ Table caption", "55.5", "71.4", "61.4", "67.2", "55.4", "49.8", "Numeric value + Column label + Table caption", " Numeric value + IsBolded + Row label + Column label + Table caption", "36.6", "60.9", "43.0", "Table 1: Table extraction results of our table parser on 50 tables from 10 NLP papers in PDF format.", "mEval 2018 Task 7) with a dataset of 350 anno-", "the standard transformer encoder (Ashish et al.,", "tated abstracts. Ammar et al. (2017, 2018); Luan", "2017; Devlin et al., 2019) to train two NLI models.", "et al. (2017); Augenstein and Sogaard (2017) ex-", "3.", "Dataset Construction", "on scientific literature. Luan et al. (2018) extend", "ploit these datasets to test neural models for IE", "We create two datasets for testing our approach", "those datasets by adding more relation types and", "for task, dataset, metric, and score (TDMS) iden-", "cross-sentence relations using coreference links.", "tification. Both datasets are taken from a collec-", "The authors also develop a framework called Sci-", "tion of NLP papers in PDF format and both re-", "entificInformation Extractorfortheextractionof", "quire similar pre-processing. First, we parse the", "six types of scientific entities (Task, Method, Met-", "PDFs using GROBID (Lopez, 2009) to extract the", "ric, Material, Other-ScientificTerm and Generic)", "title, abstract, and for each section, the section ti-", "and seven relation types (Compare, Part-of, Con-", "tle and its corresponding content. Then we apply", "junction, Evaluate-for, Feature-of, Used-for, and", "an improved table parser we developed, built on", "Hyponym-of). They reach 64.2 Fi on entity recog-", "GROBID's output, to extract all tables containing", "nition and 39.2 Fi on relation extraction. Differ-", "numeric cells from the paper. Each extracted ta-", "ently from (Luan et al., 2018), (1) we concentrate", "ble contains the table caption and a list of numeric", "on the identification of entities from a taxonomy", "cells. For each numeric cell, we detect whether it", "thatarenecessaryforthereconstruction of leader-", "has a bold typeface, and associate it to its corre-", "boards (i.e., task, dataset, metric); (2) we anal-", "sponding row and column headers. For instance,", "yse the entire paper, not only the abstract (the rea-", "for the sample paper shown in Figure 1, after pro-", "son being that the score information is rarely con-", "cessing the table shown, we extract the bolded", "tained in the abstract).", "number \"85.60\" and find its corresponding column", "headers “[Test, NER}\".", "Our method for TDMS identification resembles", "We evaluated our table parser on a set of 10 pa-", "some approaches used for textual entailment (Da-", "gan et al., 2006) or natural language inference", "pers from different venues (e.g., EMNLP, Compu-", "(NLI) (Bowman et al., 2015). We follow the exam-", "tational Linguistics journal). In total, these papers", "contain 50 tables with 1,063 numeric content cells.", "ple of White et al. (2017) and Poliak et al. (2018)", "Table1 shows the results for extracting different", "who reframe different NLP tasks, including ex-", "table elements. Our table parser achieves a macro", "traction tasks, as NLI problems. Eichler et al.", "Fi score of 82.6 for identifying table captions, and", "(2017) and Obamuyide and Vlachos (2018) have", "both used NLI approaches for relation extraction.", "74.0 macro F1 for extracting tuples of <Numeric", "value, Bolded Info, Table caption>. In general, it", "Ourworkdiffersintheinformationextractedand", "obtains higher recall than precision in all evalua-", "consequently in what context and hypothesis infor-", "tion dimensions.", "mationwe model.Currently,oneofthebestper-", "In the remainder of this section we describe our", "forming NLI models (e.g., on the SNLI dataset)", "for three way classification is (Liu et al., 2019)", "two datasets in detail.", "The authors apply deep neural networks and make", "3.1NLP-TDMS", "use of BERT (Devlin et al., 2019), a novel lan-", "guage representation model. They reach an accu-", "The content of the NLP-progress Github reposi-", "racy of 91.1%. Kim et al. (2019) exploit densely-", "tory4 provides us with expert annotations of vari-", "ous leaderboards for a few hundred papers in the", "connected co-attentive recurrent neural network,", "and reach 90% accuracy. In our scenario, we gen-", "thttps://github.com/sebastianruder/", "erate pseudo premises and hypotheses, then apply", "NLP-progress" ]

{ "arxiv_id": "1906.09317", "language": "en", "page_idx": 2 }

[ [ "0.3594", "0.0748", "0.3938", "0.0748", "0.3938", "0.0873", "0.3594", "0.0873" ], [ "0.4148", "0.07245", "0.4534", "0.076", "0.451", "0.0903", "0.4114", "0.0873" ], [ "0.638", "0.076", "0.6987", "0.076", "0.6987", "0.0873", "0.638", "0.0873" ], [ "0.712", "0.0748", "0.846", "0.0748", "0.846", "0.0885", "0.712", "0.0885" ], [ "0.1545", "0.0903", "0.2083", "0.0903", "0.2083", "0.1051", "0.1545", "0.1051" ], [ "0.3586", "0.0903", "0.3938", "0.0903", "0.3938", "0.1057", "0.3586", "0.1057" ], [ "0.4165", "0.09143", "0.451", "0.09143", "0.451", "0.10394", "0.4165", "0.10394" ], [ "0.55", "0.09143", "0.592", "0.09143", "0.592", "0.1051", "0.55", "0.1051" ], [ "0.6465", "0.0903", "0.6885", "0.0903", "0.6885", "0.1057", "0.6465", "0.1057" ], [ "0.76", "0.09143", "0.7983", "0.09143", "0.7983", "0.10394", "0.76", "0.10394" ], [ "0.3552", "0.1057", "0.3972", "0.1057", "0.3972", "0.1217", "0.3552", "0.1217" ], [ "0.413", "0.1057", "0.455", "0.1057", "0.455", "0.1217", "0.413", "0.1217" ], [ "0.7573", "0.1057", "0.8003", "0.1057", "0.8003", "0.1217", "0.7573", "0.1217" ], [ "0.6455", "0.1069", "0.687", "0.1069", "0.687", "0.1223", "0.6455", "0.1223" ], [ "0.1545", "0.1069", "0.2761", "0.1069", "0.2761", "0.1217", "0.1545", "0.1217" ], [ "0.553", "0.1081", "0.6187", "0.1081", "0.6187", "0.1194", "0.553", "0.1194" ], [ "0.1528", "0.1217", "0.3342", "0.12354", "0.3342", "0.1383", "0.1528", "0.136" ], [ "0.4207", "0.1194", "0.451", "0.1223", "0.4475", "0.1395", "0.4182", "0.136" ], [ "0.65", "0.1223", "0.686", "0.1223", "0.686", "0.1383", "0.65", "0.1383" ], [ "0.7583", "0.1223", "0.8003", "0.1223", "0.8003", "0.1383", "0.7583", "0.1383" ], [ "0.55", "0.12354", "0.6187", "0.12354", "0.6187", "0.1383", "0.55", "0.1383" ], [ "0.3728", "0.1277", "0.3787", "0.1277", "0.3787", "0.133", "0.3728", "0.133" ], [ "0.4148", "0.1395", "0.451", "0.1395", "0.451", "0.155", "0.4148", "0.155" ], [ "0.546", "0.1395", "0.596", "0.1395", "0.596", "0.155", "0.546", "0.155" ], [ "0.644", "0.1395", "0.687", "0.1395", "0.687", "0.155", "0.644", "0.155" ], [ "0.7573", "0.1395", "0.8003", "0.1395", "0.8003", "0.155", "0.7573", "0.155" ], [ "0.3594", "0.1401", "0.392", "0.1401", "0.392", "0.1526", "0.3594", "0.1526" ], [ "0.1545", "0.1401", "0.34", "0.1401", "0.34", "0.1538", "0.1545", "0.1538" ], [ "0.1721", "0.1561", "0.3333", "0.1561", "0.3333", "0.1692", "0.1721", "0.1692" ], [ "0.3586", "0.155", "0.392", "0.155", "0.392", "0.1704", "0.3586", "0.1704" ], [ "0.4207", "0.1561", "0.4458", "0.1561", "0.4458", "0.1692", "0.4207", "0.1692" ], [ "0.1704", "0.1704", "0.2778", "0.1729", "0.2778", "0.187", "0.1704", "0.1853" ], [ "0.3628", "0.1716", "0.388", "0.1716", "0.388", "0.1859", "0.3628", "0.1859" ], [ "0.4224", "0.1716", "0.4458", "0.1716", "0.4458", "0.1853", "0.4224", "0.1853" ], [ "0.5146", "0.1746", "0.586", "0.1746", "0.586", "0.1859", "0.5146", "0.1859" ], [ "0.5845", "0.1746", "0.885", "0.1746", "0.885", "0.1882", "0.5845", "0.1882" ], [ "0.1704", "0.187", "0.297", "0.1882", "0.297", "0.2031", "0.1704", "0.2019" ], [ "0.3628", "0.187", "0.388", "0.187", "0.388", "0.2019", "0.3628", "0.2019" ], [ "0.4207", "0.187", "0.4458", "0.187", "0.4458", "0.2019", "0.4207", "0.2019" ], [ "0.51", "0.1859", "0.5938", "0.1882", "0.5938", "0.2031", "0.51", "0.2007" ], [ "0.1721", "0.2037", "0.294", "0.2037", "0.294", "0.2173", "0.1721", "0.2173" ], [ "0.3645", "0.2031", "0.388", "0.2031", "0.388", "0.2173", "0.3645", "0.2173" ], [ "0.4207", "0.2037", "0.4475", "0.2037", "0.4475", "0.2173", "0.4207", "0.2173" ], [ "0.5146", "0.2299", "0.5425", "0.2299", "0.5425", "0.2411", "0.5146", "0.2411" ], [ "0.5356", "0.228", "0.6406", "0.228", "0.6406", "0.2423", "0.5356", "0.2423" ], [ "0.11755", "0.2363", "0.4878", "0.2351", "0.4878", "0.2494", "0.11755", "0.2505" ], [ "0.1184", "0.2517", "0.3652", "0.2517", "0.3652", "0.2654", "0.1184", "0.2654" ], [ "0.5146", "0.2505", "0.885", "0.2505", "0.885", "0.2642", "0.5146", "0.2642" ], [ "0.511", "0.2654", "0.882", "0.2666", "0.882", "0.2808", "0.511", "0.2798" ], [ "0.513", "0.282", "0.882", "0.2808", "0.882", "0.2957", "0.513", "0.2964" ], [ "0.11755", "0.2915", "0.486", "0.2915", "0.486", "0.3047", "0.11755", "0.3047" ], [ "0.5146", "0.3005", "0.8833", "0.3005", "0.8833", "0.3113", "0.5146", "0.3113" ], [ "0.1184", "0.3088", "0.4878", "0.3088", "0.4878", "0.3225", "0.1184", "0.3225" ], [ "0.513", "0.3154", "0.885", "0.3154", "0.885", "0.3289", "0.513", "0.3289" ], [ "0.12006", "0.3267", "0.4895", "0.3267", "0.4895", "0.3374", "0.12006", "0.3374" ], [ "0.511", "0.3308", "0.885", "0.33", "0.885", "0.3445", "0.511", "0.3457" ], [ "0.1184", "0.339", "0.4878", "0.339", "0.4878", "0.354", "0.1184", "0.354" ], [ "0.511", "0.3467", "0.8833", "0.3467", "0.8833", "0.3599", "0.511", "0.3599" ], [ "0.12006", "0.357", "0.4895", "0.357", "0.4895", "0.3706", "0.12006", "0.3706" ], [ "0.51", "0.3623", "0.8564", "0.3635", "0.8564", "0.3777", "0.51", "0.3772" ], [ "0.1184", "0.3735", "0.4878", "0.3735", "0.4878", "0.3872", "0.1184", "0.3872" ], [ "0.11755", "0.3896", "0.4878", "0.3884", "0.4878", "0.4026", "0.11755", "0.4038" ], [ "0.5156", "0.3943", "0.529", "0.3943", "0.529", "0.4026", "0.5156", "0.4026" ], [ "0.5356", "0.3955", "0.5415", "0.3955", "0.5415", "0.401", "0.5356", "0.401" ], [ "0.538", "0.3901", "0.827", "0.3914", "0.827", "0.4062", "0.538", "0.405" ], [ "0.11584", "0.405", "0.4895", "0.4026", "0.4895", "0.4192", "0.11584", "0.4216" ], [ "0.11755", "0.4216", "0.4878", "0.4216", "0.4878", "0.4353", "0.11755", "0.4353" ], [ "0.513", "0.4163", "0.708", "0.4175", "0.708", "0.4316", "0.513", "0.4312" ], [ "0.1184", "0.4382", "0.4902", "0.4382", "0.4902", "0.452", "0.1184", "0.452" ], [ "0.511", "0.4395", "0.881", "0.4395", "0.881", "0.4531", "0.511", "0.4531" ], [ "0.1184", "0.4531", "0.4902", "0.4531", "0.4902", "0.4673", "0.1184", "0.4673" ], [ "0.5146", "0.456", "0.885", "0.456", "0.885", "0.4697", "0.5146", "0.4697" ], [ "0.1184", "0.4685", "0.4902", "0.4697", "0.4902", "0.484", "0.1184", "0.4827" ], [ "0.513", "0.4714", "0.8833", "0.4714", "0.8833", "0.4863", "0.513", "0.4863" ], [ "0.1184", "0.4863", "0.398", "0.4863", "0.398", "0.4995", "0.1184", "0.4995" ], [ "0.511", "0.4868", "0.8833", "0.4868", "0.8833", "0.502", "0.511", "0.502" ], [ "0.1377", "0.502", "0.486", "0.502", "0.486", "0.5156", "0.1377", "0.5156" ], [ "0.511", "0.5044", "0.8833", "0.5044", "0.8833", "0.517", "0.511", "0.517" ], [ "0.11755", "0.5186", "0.486", "0.5186", "0.486", "0.532", "0.11755", "0.532" ], [ "0.513", "0.517", "0.607", "0.5195", "0.607", "0.5337", "0.513", "0.532" ], [ "0.12006", "0.535", "0.4895", "0.535", "0.4895", "0.549", "0.12006", "0.549" ], [ "0.532", "0.536", "0.8833", "0.536", "0.8833", "0.5503", "0.532", "0.5503" ], [ "0.12006", "0.5503", "0.4878", "0.5503", "0.4878", "0.564", "0.12006", "0.564" ], [ "0.513", "0.5527", "0.885", "0.5527", "0.885", "0.5664", "0.513", "0.5664" ], [ "0.1184", "0.5664", "0.4895", "0.5664", "0.4895", "0.58", "0.1184", "0.58" ], [ "0.511", "0.568", "0.882", "0.568", "0.882", "0.5806", "0.511", "0.5806" ], [ "0.12177", "0.5845", "0.4878", "0.5845", "0.4878", "0.594", "0.12177", "0.594" ], [ "0.513", "0.5845", "0.882", "0.5845", "0.882", "0.5986", "0.513", "0.5986" ], [ "0.1184", "0.5986", "0.4895", "0.5986", "0.4895", "0.6123", "0.1184", "0.6123" ], [ "0.513", "0.601", "0.8833", "0.601", "0.8833", "0.6143", "0.513", "0.6143" ], [ "0.1184", "0.615", "0.4902", "0.615", "0.4902", "0.629", "0.1184", "0.629" ], [ "0.513", "0.6177", "0.839", "0.6177", "0.839", "0.6274", "0.513", "0.6274" ], [ "0.12006", "0.6313", "0.4895", "0.6313", "0.4895", "0.6445", "0.12006", "0.6445" ], [ "0.11755", "0.6465", "0.486", "0.6465", "0.486", "0.66", "0.11755", "0.66" ], [ "0.511", "0.641", "0.8833", "0.6426", "0.8833", "0.6567", "0.511", "0.6562" ], [ "0.1184", "0.663", "0.3997", "0.663", "0.3997", "0.677", "0.1184", "0.677" ], [ "0.5146", "0.659", "0.885", "0.659", "0.885", "0.673", "0.5146", "0.673" ], [ "0.1377", "0.6787", "0.4895", "0.6787", "0.4895", "0.6934", "0.1377", "0.6934" ], [ "0.511", "0.6733", "0.881", "0.6733", "0.881", "0.687", "0.511", "0.687" ], [ "0.1184", "0.6953", "0.4895", "0.6953", "0.4895", "0.709", "0.1184", "0.709" ], [ "0.513", "0.6914", "0.8833", "0.6914", "0.8833", "0.705", "0.513", "0.705" ], [ "0.1184", "0.712", "0.4902", "0.7104", "0.4902", "0.7236", "0.1184", "0.7256" ], [ "0.513", "0.707", "0.885", "0.707", "0.885", "0.72", "0.513", "0.72" ], [ "0.1184", "0.728", "0.4895", "0.728", "0.4895", "0.741", "0.1184", "0.741" ], [ "0.513", "0.7236", "0.8833", "0.7236", "0.8833", "0.737", "0.513", "0.737" ], [ "0.11755", "0.7437", "0.4902", "0.742", "0.4902", "0.7573", "0.11755", "0.7583" ], [ "0.513", "0.7383", "0.8833", "0.7383", "0.8833", "0.753", "0.513", "0.753" ], [ "0.11755", "0.759", "0.4895", "0.7583", "0.4895", "0.7725", "0.11755", "0.774" ], [ "0.511", "0.7544", "0.8833", "0.755", "0.8833", "0.7695", "0.511", "0.7686" ], [ "0.12006", "0.7754", "0.4895", "0.7754", "0.4895", "0.789", "0.12006", "0.789" ], [ "0.513", "0.7715", "0.885", "0.7715", "0.885", "0.785", "0.513", "0.785" ], [ "0.1184", "0.792", "0.4902", "0.792", "0.4902", "0.8057", "0.1184", "0.8057" ], [ "0.511", "0.7876", "0.885", "0.786", "0.885", "0.8003", "0.511", "0.802" ], [ "0.11755", "0.807", "0.4895", "0.808", "0.4895", "0.8223", "0.11755", "0.822" ], [ "0.531", "0.81", "0.885", "0.8125", "0.885", "0.8267", "0.531", "0.8247" ], [ "0.11755", "0.8237", "0.392", "0.8223", "0.392", "0.8374", "0.11755", "0.8384" ], [ "0.513", "0.8237", "0.7715", "0.8247", "0.7715", "0.836", "0.513", "0.835" ], [ "0.534", "0.835", "0.8833", "0.8374", "0.8833", "0.851", "0.534", "0.8486" ], [ "0.1377", "0.8467", "0.4902", "0.847", "0.4902", "0.8623", "0.1377", "0.861" ], [ "0.513", "0.847", "0.882", "0.8486", "0.882", "0.8623", "0.513", "0.861" ], [ "0.11755", "0.86", "0.4878", "0.86", "0.4878", "0.8735", "0.11755", "0.8735" ], [ "0.511", "0.86", "0.8833", "0.86", "0.8833", "0.8735", "0.511", "0.8735" ], [ "0.1184", "0.8716", "0.4902", "0.8716", "0.4902", "0.8853", "0.1184", "0.8853" ], [ "0.513", "0.8726", "0.8833", "0.8726", "0.8833", "0.8857", "0.513", "0.8857" ], [ "0.11755", "0.883", "0.4902", "0.8843", "0.4902", "0.898", "0.11755", "0.8965" ], [ "0.511", "0.8843", "0.8867", "0.8843", "0.8867", "0.898", "0.511", "0.898" ], [ "0.12006", "0.898", "0.2761", "0.898", "0.2761", "0.9077", "0.12006", "0.9077" ], [ "0.513", "0.898", "0.586", "0.898", "0.586", "0.909", "0.513", "0.909" ] ]

[ "Full", "Exp", "#Papers", "#Extracted tables", "Papers", "332", "332", "ACL", "1958", "4537", "1269", "1269", "3488", "1167", "Extracted tables", "EMNLP", "\"Unknown\"annotations", "90", "730", "1559", "NAACL", "-", "606", "Total", "3855", "9584", "848", "Leaderboard annotations", "Distinct leaderboards", "168", "77", "Distinct tasks", "35", "18", "Table 3:", "Statistics of papers and extracted tables in", "Distinct datasets", "99", "44", "ARC-PDN.", "Distinct metrics", "72", "30", "3.2", "ARC-PDN", "Table 2: Statistics of leaderboard annotations in NLP.", "TDMS (Full) and NLP-TDMS (Exp).", "To test our model in a more realistic scenario, we", "create a second dataset ARC-PDN.6 We select pa-", "pers (in PDF format) published in ACL, EMNLP,", "NLP domain. The repository is organized follow-", "andNAACLbetween2010to2015fromthemost", "ing a “\"language-domain/task-dataset-leaderboard\"", "recent version of the ACL Anthology Reference", "structure.After crawlingthis informationtogether", "Corpus (ARC) (Bird et al., 2008). Table 3 shows", "with the corresponding papers (in PDF format),", "statistics about papers and extracted tables in this", "we clean the dataset manually. This includes: ()", "dataset after the PDF parsing described above.", "normalizing task name, dataset name, and evalu-", "ation metrics across leaderboards created by dif-", "4", "", "Method for TDMS Identification", "ferent experts, e.g., using “F1\" to represent \"F-", "score\" and \"Fscore\"; (2) for each leaderboard ta-", "4.1 ProblemDefinition", "ble, only keeping the best result from the same pa-", "We represent each leaderboard as a <Task,", "per; (3) splitting a leaderboard table into several", "Dataset, Metric> triple (TDM triple). Given an", "leaderboard tables if its column headers represent", "experimental scientific paper D, we want to iden-", "datasets instead of evaluation metrics.", "tify relevant TDM triples from a taxonomy and", "The resulting dataset NLP-TDMS (Full) con-", "extract the best numeric score for each predicted", "tains 332 papers with 848 leaderboard annotations.", "TDM triple.", "Each leaderboard annotation is a tuple containing", "However, scientific papers are often long doc-", "task, dataset, metric, and score (as shown in Fig-", "uments and only some parts of the document are", "ure 1). In total, we have 168 distinct leaderboards", "useful to predict TDM triples and the associated", "(i.e.,<Task,Dataset.Metric>triples) and only", "scores. Hence, we define a document representa-", "around half of them (77) are associated with at", "tion, called DocTAET and a table score represen-", "least five papers. We treat these manually curated", "tationcalled SC (score context),asfollows:", "TDM triples as an NLP knowledge taxonomy and", "we aim to explore how well we can associate a pa-", "DocTAET.For each scientific paper, its Doc-", "per to the corresponding TDM triples.", "TAET representation contains the following four", "We further create NLP-TDMS (Exp) by remov-", "parts: Title, Abstract, ExpSetup, and Tablelnfo. Ti-", "ing those leaderboards that are associated with", "tle and Abstract often help in predicting Task. Ex-", "fewer than five papers. If all leaderboard anno-", "pSetup contains all sentences which are likely to", "tations of a paper belong to these removed leader-", "describe the experimental setup, which can help to", "boards, we tag this paper as \"Unknown\". Table 2", "predict Dataset and Metric. We use a few heuris-", "compares statistics of NLP-TDMS (Full) and NLP", "tics to extract such sentences.7 Finally, table cap-", "TDMS (Exp). All experiments in this paper (ex-", "tions and column headers are important in predict-", "cept experiments in the zero-shot setup in Section", "ing Dataset and Metric. We collect them in the", "7) are on NLP-TDMS (Exp) and going forward we", "‘PDN comes from the anthology's directory prefixes for", "will refer to that only as NLP-TDMS.", "ACL,EMNLP,and NAACL,respectively.", "7A sentence is included in ExpSetup if it: (1) contains any", "5In this paper, we focus on tagging papers with different", "of the following cue words/phrases: [experiment on, exper-", "leaderboards (i.e., TDM triples). For each leaderboard table,", "iment in, evaluation(s), evaluate, evaluated, dataset(s), cor-", "an ideal situation would be to extract all results reported in", "pus, corpora); and (2) belongs to a section whose title con-", "the same paper and associate them to different methods, we", "tains any of the following words: [experiment(s), evaluation,", "leavethisforfuturework.", "dataset(s)." ]

{ "arxiv_id": "1906.09317", "language": "en", "page_idx": 3 }

[ [ "0.12006", "0.0766", "0.4937", "0.0766", "0.4937", "0.0903", "0.12006", "0.0903" ], [ "0.513", "0.0766", "0.8833", "0.0766", "0.8833", "0.0903", "0.513", "0.0903" ], [ "0.11755", "0.09143", "0.4324", "0.0932", "0.4324", "0.1069", "0.11755", "0.1051" ], [ "0.513", "0.0932", "0.875", "0.0932", "0.875", "0.1069", "0.513", "0.1069" ], [ "0.12006", "0.11816", "0.4895", "0.11816", "0.4895", "0.133", "0.12006", "0.133" ], [ "0.513", "0.11755", "0.882", "0.11755", "0.882", "0.1312", "0.513", "0.1312" ], [ "0.1184", "0.1342", "0.4878", "0.1342", "0.4878", "0.1478", "0.1184", "0.1478" ], [ "0.511", "0.133", "0.882", "0.133", "0.882", "0.1467", "0.511", "0.1467" ], [ "0.12006", "0.1509", "0.4895", "0.1509", "0.4895", "0.1644", "0.12006", "0.1644" ], [ "0.513", "0.1509", "0.8833", "0.1509", "0.8833", "0.1644", "0.513", "0.1644" ], [ "0.12006", "0.1663", "0.4902", "0.1663", "0.4902", "0.1812", "0.12006", "0.1812" ], [ "0.513", "0.1663", "0.8833", "0.1663", "0.8833", "0.1812", "0.513", "0.1812" ], [ "0.1184", "0.1829", "0.4937", "0.1829", "0.4937", "0.1965", "0.1184", "0.1965" ], [ "0.51", "0.1829", "0.882", "0.1829", "0.882", "0.1965", "0.51", "0.1965" ], [ "0.11755", "0.1996", "0.4895", "0.1996", "0.4895", "0.2131", "0.11755", "0.2131" ], [ "0.513", "0.199", "0.882", "0.199", "0.882", "0.212", "0.513", "0.212" ], [ "0.1184", "0.2156", "0.4902", "0.2156", "0.4902", "0.2286", "0.1184", "0.2286" ], [ "0.513", "0.2156", "0.885", "0.2156", "0.885", "0.2286", "0.513", "0.2286" ], [ "0.11755", "0.2299", "0.3745", "0.231", "0.3745", "0.2452", "0.11755", "0.2446" ], [ "0.513", "0.231", "0.81", "0.231", "0.81", "0.2446", "0.513", "0.2446" ], [ "0.11755", "0.2583", "0.3015", "0.2588", "0.3015", "0.2737", "0.11755", "0.2725" ], [ "0.511", "0.2542", "0.8833", "0.2559", "0.8833", "0.2708", "0.511", "0.2683" ], [ "0.513", "0.2725", "0.8833", "0.2725", "0.8833", "0.2861", "0.513", "0.2861" ], [ "0.1184", "0.2808", "0.4895", "0.2808", "0.4895", "0.2957", "0.1184", "0.2957" ], [ "0.513", "0.288", "0.8833", "0.288", "0.8833", "0.3018", "0.513", "0.3018" ], [ "0.12006", "0.2976", "0.4902", "0.2976", "0.4902", "0.3113", "0.12006", "0.3113" ], [ "0.513", "0.304", "0.885", "0.304", "0.885", "0.3184", "0.513", "0.3184" ], [ "0.1184", "0.3135", "0.4895", "0.3135", "0.4895", "0.3267", "0.1184", "0.3267" ], [ "0.511", "0.3206", "0.8833", "0.3196", "0.8833", "0.3342", "0.511", "0.335" ], [ "0.11584", "0.3289", "0.4902", "0.328", "0.4902", "0.3425", "0.11584", "0.3433" ], [ "0.5146", "0.3362", "0.8833", "0.3362", "0.8833", "0.3499", "0.5146", "0.3499" ], [ "0.12006", "0.3457", "0.4475", "0.3457", "0.4475", "0.3594", "0.12006", "0.3594" ], [ "0.5146", "0.3528", "0.882", "0.3528", "0.882", "0.3665", "0.5146", "0.3665" ], [ "0.1184", "0.3694", "0.4265", "0.3706", "0.4265", "0.3855", "0.1184", "0.3843" ], [ "0.511", "0.3677", "0.882", "0.3677", "0.882", "0.3813", "0.511", "0.3813" ], [ "0.11584", "0.3901", "0.4895", "0.3901", "0.4895", "0.4038", "0.11584", "0.4038" ], [ "0.513", "0.3855", "0.8833", "0.3855", "0.8833", "0.3984", "0.513", "0.3984" ], [ "0.511", "0.3997", "0.8833", "0.401", "0.8833", "0.415", "0.511", "0.4146" ], [ "0.1184", "0.408", "0.4895", "0.408", "0.4895", "0.4216", "0.1184", "0.4216" ], [ "0.11584", "0.4229", "0.4895", "0.4216", "0.4895", "0.4358", "0.11584", "0.437" ], [ "0.513", "0.4175", "0.882", "0.4175", "0.882", "0.4312", "0.513", "0.4312" ], [ "0.511", "0.4329", "0.882", "0.4329", "0.882", "0.4478", "0.511", "0.4478" ], [ "0.1184", "0.4395", "0.4895", "0.4395", "0.4895", "0.4531", "0.1184", "0.4531" ], [ "0.11755", "0.4548", "0.4878", "0.4548", "0.4878", "0.4697", "0.11755", "0.4697" ], [ "0.513", "0.4495", "0.8833", "0.4495", "0.8833", "0.4631", "0.513", "0.4631" ], [ "0.5146", "0.4656", "0.882", "0.4656", "0.882", "0.4785", "0.5146", "0.4785" ], [ "0.1184", "0.4727", "0.4895", "0.4727", "0.4895", "0.4863", "0.1184", "0.4863" ], [ "0.51", "0.4785", "0.885", "0.4797", "0.885", "0.4976", "0.51", "0.4963" ], [ "0.1184", "0.488", "0.4902", "0.488", "0.4902", "0.502", "0.1184", "0.502" ], [ "0.511", "0.4963", "0.885", "0.4954", "0.885", "0.5103", "0.511", "0.511" ], [ "0.1184", "0.5044", "0.4895", "0.5044", "0.4895", "0.517", "0.1184", "0.517" ], [ "0.513", "0.513", "0.885", "0.513", "0.885", "0.527", "0.513", "0.527" ], [ "0.12006", "0.521", "0.4902", "0.521", "0.4902", "0.5337", "0.12006", "0.5337" ], [ "0.513", "0.53", "0.882", "0.53", "0.882", "0.5435", "0.513", "0.5435" ], [ "0.12006", "0.536", "0.4895", "0.536", "0.4895", "0.5503", "0.12006", "0.5503" ], [ "0.513", "0.5444", "0.8833", "0.546", "0.8833", "0.56", "0.513", "0.5586" ], [ "0.12006", "0.5527", "0.4878", "0.5527", "0.4878", "0.5664", "0.12006", "0.5664" ], [ "0.5146", "0.561", "0.8833", "0.561", "0.8833", "0.576", "0.5146", "0.576" ], [ "0.12006", "0.5684", "0.4895", "0.5684", "0.4895", "0.582", "0.12006", "0.582" ], [ "0.1184", "0.582", "0.4878", "0.5845", "0.4878", "0.5986", "0.1184", "0.5967" ], [ "0.513", "0.5767", "0.8833", "0.5776", "0.8833", "0.593", "0.513", "0.5913" ], [ "0.513", "0.594", "0.8833", "0.594", "0.8833", "0.6094", "0.513", "0.6094" ], [ "0.1184", "0.601", "0.486", "0.601", "0.486", "0.611", "0.1184", "0.611" ], [ "0.5146", "0.61", "0.6094", "0.61", "0.6094", "0.6245", "0.5146", "0.6245" ], [ "0.12006", "0.616", "0.4895", "0.616", "0.4895", "0.6313", "0.12006", "0.6313" ], [ "0.12006", "0.633", "0.4895", "0.633", "0.4895", "0.6465", "0.12006", "0.6465" ], [ "0.513", "0.6353", "0.6465", "0.6353", "0.6465", "0.6484", "0.513", "0.6484" ], [ "0.1184", "0.6484", "0.4902", "0.648", "0.4902", "0.662", "0.1184", "0.663" ], [ "0.5146", "0.654", "0.882", "0.654", "0.882", "0.6685", "0.5146", "0.6685" ], [ "0.12006", "0.665", "0.4902", "0.665", "0.4902", "0.6787", "0.12006", "0.6787" ], [ "0.513", "0.6704", "0.8867", "0.6704", "0.8867", "0.685", "0.513", "0.685" ], [ "0.11755", "0.681", "0.4895", "0.6797", "0.4895", "0.695", "0.11755", "0.6953" ], [ "0.513", "0.687", "0.8833", "0.687", "0.8833", "0.7007", "0.513", "0.7007" ], [ "0.11755", "0.6978", "0.4895", "0.6978", "0.4895", "0.7114", "0.11755", "0.7114" ], [ "0.5146", "0.705", "0.882", "0.705", "0.882", "0.7153", "0.5146", "0.7153" ], [ "0.1184", "0.7134", "0.4434", "0.7134", "0.4434", "0.727", "0.1184", "0.727" ], [ "0.513", "0.7197", "0.8833", "0.7197", "0.8833", "0.733", "0.513", "0.733" ], [ "0.1368", "0.7285", "0.4895", "0.7285", "0.4895", "0.742", "0.1368", "0.742" ], [ "0.513", "0.7354", "0.882", "0.7354", "0.882", "0.749", "0.513", "0.749" ], [ "0.1184", "0.745", "0.4902", "0.745", "0.4902", "0.759", "0.1184", "0.759" ], [ "0.511", "0.7505", "0.8833", "0.75", "0.8833", "0.764", "0.511", "0.7656" ], [ "0.1184", "0.7627", "0.4895", "0.7627", "0.4895", "0.7754", "0.1184", "0.7754" ], [ "0.5146", "0.767", "0.882", "0.767", "0.882", "0.781", "0.5146", "0.781" ], [ "0.1184", "0.777", "0.4895", "0.778", "0.4895", "0.792", "0.1184", "0.7915" ], [ "0.511", "0.783", "0.8833", "0.782", "0.8833", "0.7964", "0.511", "0.7974" ], [ "0.11755", "0.792", "0.4895", "0.7935", "0.4895", "0.808", "0.11755", "0.807" ], [ "0.513", "0.799", "0.8833", "0.7974", "0.8833", "0.8125", "0.513", "0.8135" ], [ "0.11755", "0.81", "0.4895", "0.81", "0.4895", "0.8237", "0.11755", "0.8237" ], [ "0.513", "0.8154", "0.8833", "0.8154", "0.8833", "0.829", "0.513", "0.829" ], [ "0.1394", "0.833", "0.4902", "0.8345", "0.4902", "0.8486", "0.1394", "0.847" ], [ "0.511", "0.8306", "0.8833", "0.83", "0.8833", "0.844", "0.511", "0.8457" ], [ "0.1184", "0.8467", "0.4902", "0.8467", "0.4902", "0.86", "0.1184", "0.86" ], [ "0.511", "0.847", "0.8833", "0.847", "0.8833", "0.8623", "0.511", "0.8623" ], [ "0.1184", "0.8594", "0.4902", "0.8594", "0.4902", "0.8726", "0.1184", "0.8726" ], [ "0.12006", "0.8726", "0.4492", "0.8726", "0.4492", "0.883", "0.12006", "0.883" ], [ "0.5264", "0.868", "0.885", "0.8706", "0.885", "0.8853", "0.5264", "0.883" ], [ "0.1377", "0.882", "0.4902", "0.883", "0.4902", "0.898", "0.1377", "0.8965" ], [ "0.509", "0.8843", "0.885", "0.8843", "0.885", "0.898", "0.509", "0.898" ], [ "0.11755", "0.8945", "0.3787", "0.8955", "0.3787", "0.909", "0.11755", "0.9077" ], [ "0.513", "0.898", "0.58", "0.898", "0.58", "0.909", "0.513", "0.909" ] ]

[ "Tablelnfo part. Figure 2 (upper right) illustrates", "extraction. In the following we describe how we", "the DocTAET extraction for a given paper.", "generate training instances for these two models.", "SC. For each table in a scientific paper, we fo-", "DocTAET-TDM model.Illustrated in Figure 3", "cus on boldfaced numeric scores because they are", "(upper left), this model predicts whether a TDM", "more likely to be the best scores for the corre-", "triple can be inferred from a DocTAET. For a set", "sponding TDM triples.8 For a specific boldfaced", "of n TDM triples ({ti, t2, .., tnj) from a taxon-", "numeric score in a table, its context (SC) contains ", " omy, if a paper d; (DocTAET) is annotated with ti", "its corresponding column headers and the table", "and t2, we then generate two positive training in-", "caption. Figure 2 (lower right) shows the extracted", "stances (d; → ti and d; → t2) and n - 2 negative", "SC for the scores 85.60 and 61.71.", "training instances (d; tj, 2 <j ≤ n).", "4.2TDMS-IE System", "SC-DM model.Illustrated in Figure 3 (lower", "left), this model predicts whether a score con-", "We develop a system called TDMS-IE to associate", "text SC indicates a DM pair. To form train-", "TDM triples to a given experimental scientific pa-", "ing instances, we start with the list of DM pairs", "per: Our system also extracts the best numeric", "((pi,P2,...Pm)from a taxonomyandapaper dj,", "score for each predicted TDM triple.Figure3", "which is annotated with a TDM triple t (contain-", "shows the system architecture for TDMS-IE.", "ing pi) and a numeric score s. We first try to ex-", "4.2.1 TDMS-IE Classification Models", "tract the score contexts (SC) for all bolded numeric", "To predict correct TDM triples and associate the", "scores. If di's annotated score s is equal to one", "of the bolded scores sk (typically there should not", "appropriate scores, we adopt a natural language", "inference approach (NLI) (Poliak et al., 2018)", "be more than one), we generate a positive train-", "ing instance (SCsk=1 =→ pi). Negative instances", "and learn a binary classifier for pairs of docu-", "ment contexts and TDM label hypotheses. Specif-", "can be generated for this context by choosing other", "DMs not associated with the context, i.e., m - 1", "ically, we split the problem into two tasks: (1)", "negative training instances (SCsk=1 ≠ Pj, 1<", "given a document representation DocTAET, we", "j ≤ m). For example, an SC with “ROUGE", "would like to predict whether a specific TDM", "for anonymized CNN/Daily Mail\" might form a", "triple can be inferred (e.g., give a document we", "positive instance with DM <CNN / Daily Mail", "infer <Summarization, Gigaword, ROUGE-2>);", "ROUGE-L>, and then a negative instance with", "(2) we predict whether a <Dataset, Metric> tu-", "DM <Penn Treebank, LAS>. Additional nega-", "ple (DM) can be inferred given a score context", "Sc.9 This setup has two advantages: first, it nat-", "tive training instances come from bolded scores sk", "which do not match s (e.g., SCs Pj, 1 < k,", "urally captures the inter-relations between differ-", "1≤j≤m).", "ent labels by encoding the three types of labels", "(i.e., task, dataset, metric) into the same hypoth-", "4.2.2Inference", "esis. Second, similar to approaches for NLI, it", "During the inference stage (see Figure 3 (right),", "forces the model to focus on learning the similar-", "for a given scientific paper in PDF format, our", "ity patterns between DocTAET and various TDM", "system first uses the PDF parser and table extrac-", "triples. For instance, the model will capture the", "tor(describedinSection3)togeneratethedocu-", "similarities between ROUGE-2 and \"Rg-2\".", "ment representation DocTAET. We also extract all", "Recently, a multi-head self-attention encoder", "boldfaced scores and their contexts from each ta-", "(Ashish et al., 2017) has been shown to perform", "ble. Next, we apply the DocTAET-TDM model to", "well in various NLP tasks, including NLI (Devlin", "predict TDM triples among all TDM triple candi-", "et al., 2019). We apply the standard transformer", "dates for the paperlo. Then, to extract scores for", "encoder (Devlin et al., 2019) to train our models,", "the predicted TDM triples, we apply the SC-DM", "one for TDM triple prediction, and one for score", "model to every extracted score context (SC) and", "8 We randomly choose 10 papers from NLP-TDMS (Full)", "predicted DM pair (taken from the predicted TDM", "and compare their TDMS tuple annotations with the results", "triples). This step tells us how likely it is that a", "reported in the original tables. We found that 78% (18/23) of", "the annotated tuples containboldfaced numeric scores.", "IOThe TDM triple candidates could be the valid TDM", "We look for the relation SC-DM, rather then SC-TDM,", " triples from the training set, or a set of TDM triples from a", "because rarely the task is mentioned in SC.", "taxonomy." ]

{ "arxiv_id": "1906.09317", "language": "en", "page_idx": 4 }

[ [ "0.08496", "0.065", "0.1299", "0.06757", "0.1283", "0.0821", "0.08417", "0.0802" ], [ "0.10297", "0.0808", "0.4844", "0.0808", "0.4844", "0.0947", "0.10297", "0.0947" ], [ "0.562", "0.0903", "0.5767", "0.09344", "0.574", "0.101", "0.5596", "0.09784" ], [ "0.58", "0.09155", "0.6167", "0.09155", "0.6167", "0.0985", "0.58", "0.0985" ], [ "0.581", "0.1016", "0.6445", "0.1016", "0.6445", "0.113", "0.581", "0.113" ], [ "0.0866", "0.09534", "0.486", "0.09534", "0.486", "0.1092", "0.0866", "0.1092" ], [ "0.08496", "0.10986", "0.486", "0.10986", "0.486", "0.1237", "0.08496", "0.1237" ], [ "0.8037", "0.1048", "0.847", "0.1048", "0.847", "0.11237", "0.8037", "0.11237" ], [ "0.5825", "0.11426", "0.652", "0.11426", "0.652", "0.1256", "0.5825", "0.1256" ], [ "0.08496", "0.1237", "0.486", "0.1237", "0.486", "0.137", "0.08496", "0.137" ], [ "0.0866", "0.1395", "0.4844", "0.1395", "0.4844", "0.1497", "0.0866", "0.1497" ], [ "0.7534", "0.1445", "0.7607", "0.1445", "0.7607", "0.1477", "0.7534", "0.1477" ], [ "0.7686", "0.1414", "0.8", "0.1445", "0.7983", "0.1515", "0.767", "0.1483" ], [ "0.08496", "0.1515", "0.486", "0.1515", "0.486", "0.1654", "0.08496", "0.1654" ], [ "0.7646", "0.1515", "0.8", "0.1515", "0.8", "0.159", "0.7646", "0.159" ], [ "0.08496", "0.1654", "0.4878", "0.1654", "0.4878", "0.1787", "0.08496", "0.1787" ], [ "0.8", "0.1604", "0.8735", "0.1622", "0.8735", "0.1736", "0.8", "0.1718" ], [ "0.08496", "0.178", "0.486", "0.1799", "0.486", "0.1932", "0.08496", "0.1913" ], [ "0.08496", "0.1932", "0.486", "0.1932", "0.486", "0.207", "0.08496", "0.207" ], [ "0.08496", "0.207", "0.486", "0.207", "0.486", "0.2203", "0.08496", "0.2203" ], [ "0.761", "0.2153", "0.8325", "0.2153", "0.8325", "0.2229", "0.761", "0.2229" ], [ "0.835", "0.2166", "0.8545", "0.2166", "0.8545", "0.2203", "0.835", "0.2203" ], [ "0.0866", "0.2229", "0.486", "0.2229", "0.486", "0.2329", "0.0866", "0.2329" ], [ "0.5522", "0.2235", "0.638", "0.2235", "0.638", "0.2311", "0.5522", "0.2311" ], [ "0.0866", "0.2349", "0.486", "0.2349", "0.486", "0.2488", "0.0866", "0.2488" ], [ "0.574", "0.2343", "0.607", "0.2343", "0.607", "0.2412", "0.574", "0.2412" ], [ "0.08496", "0.2488", "0.4836", "0.2488", "0.4836", "0.262", "0.08496", "0.262" ], [ "0.0866", "0.2632", "0.4844", "0.2632", "0.4844", "0.2766", "0.0866", "0.2766" ], [ "0.8164", "0.2695", "0.841", "0.2695", "0.841", "0.281", "0.8164", "0.281" ], [ "0.08496", "0.276", "0.4844", "0.276", "0.4844", "0.289", "0.08496", "0.289" ], [ "0.814", "0.2798", "0.8545", "0.2798", "0.8545", "0.2842", "0.814", "0.2842" ], [ "0.08496", "0.2903", "0.4844", "0.2903", "0.4844", "0.3037", "0.08496", "0.3037" ], [ "0.7466", "0.2942", "0.8667", "0.2942", "0.8667", "0.305", "0.7466", "0.305" ], [ "0.0866", "0.3062", "0.4844", "0.3062", "0.4844", "0.3164", "0.0866", "0.3164" ], [ "0.751", "0.3062", "0.864", "0.3062", "0.864", "0.3132", "0.751", "0.3132" ], [ "0.08496", "0.3176", "0.4844", "0.3176", "0.4844", "0.3308", "0.08496", "0.3308" ], [ "0.7915", "0.3164", "0.822", "0.3164", "0.822", "0.324", "0.7915", "0.324" ], [ "0.08417", "0.332", "0.4395", "0.3308", "0.4395", "0.3452", "0.08417", "0.3467" ], [ "0.10455", "0.3467", "0.4836", "0.3467", "0.4836", "0.3567", "0.10455", "0.3567" ], [ "0.5664", "0.3516", "0.8325", "0.3516", "0.8325", "0.3655", "0.5664", "0.3655" ], [ "0.0866", "0.3599", "0.486", "0.3599", "0.486", "0.3738", "0.0866", "0.3738" ], [ "0.08496", "0.3726", "0.486", "0.3726", "0.486", "0.3865", "0.08496", "0.3865" ], [ "0.08496", "0.3894", "0.246", "0.3894", "0.246", "0.3997", "0.08496", "0.3997" ], [ "0.513", "0.3901", "0.9126", "0.3901", "0.9126", "0.404", "0.513", "0.404" ], [ "0.1054", "0.4016", "0.486", "0.4016", "0.486", "0.4153", "0.1054", "0.4153" ], [ "0.511", "0.4048", "0.914", "0.4048", "0.914", "0.4185", "0.511", "0.4185" ], [ "0.08496", "0.4153", "0.4844", "0.4153", "0.4844", "0.4287", "0.08496", "0.4287" ], [ "0.513", "0.4192", "0.9126", "0.4192", "0.9126", "0.433", "0.513", "0.433" ], [ "0.08496", "0.43", "0.4885", "0.43", "0.4885", "0.443", "0.08496", "0.443" ], [ "0.513", "0.433", "0.914", "0.433", "0.914", "0.4463", "0.513", "0.4463" ], [ "0.08496", "0.443", "0.2336", "0.443", "0.2336", "0.457", "0.08496", "0.457" ], [ "0.513", "0.4463", "0.914", "0.4463", "0.914", "0.4602", "0.513", "0.4602" ], [ "0.1013", "0.457", "0.4844", "0.457", "0.4844", "0.4702", "0.1013", "0.4702" ], [ "0.513", "0.4602", "0.9126", "0.4602", "0.9126", "0.4734", "0.513", "0.4734" ], [ "0.08417", "0.4717", "0.486", "0.4702", "0.486", "0.4841", "0.08417", "0.4849" ], [ "0.511", "0.4734", "0.9126", "0.4749", "0.9126", "0.488", "0.511", "0.4873" ], [ "0.08496", "0.4849", "0.486", "0.4849", "0.486", "0.4988", "0.08496", "0.4988" ], [ "0.511", "0.488", "0.914", "0.488", "0.914", "0.502", "0.511", "0.502" ], [ "0.08417", "0.4988", "0.4844", "0.4988", "0.4844", "0.512", "0.08417", "0.512" ], [ "0.511", "0.502", "0.914", "0.5024", "0.914", "0.5166", "0.511", "0.515" ], [ "0.10455", "0.5146", "0.4836", "0.5146", "0.4836", "0.5244", "0.10455", "0.5244" ], [ "0.511", "0.5166", "0.7305", "0.518", "0.7305", "0.529", "0.511", "0.5264" ], [ "0.08496", "0.5264", "0.486", "0.5264", "0.486", "0.5405", "0.08496", "0.5405" ], [ "0.5146", "0.535", "0.615", "0.535", "0.615", "0.5454", "0.5146", "0.5454" ], [ "0.6177", "0.5386", "0.63", "0.5386", "0.63", "0.5425", "0.6177", "0.5425" ], [ "0.6274", "0.534", "0.9126", "0.534", "0.9126", "0.5474", "0.6274", "0.5474" ], [ "0.08496", "0.5405", "0.4878", "0.5405", "0.4878", "0.5537", "0.08496", "0.5537" ], [ "0.513", "0.5474", "0.9126", "0.5474", "0.9126", "0.561", "0.513", "0.561" ], [ "0.0866", "0.556", "0.2361", "0.556", "0.2361", "0.5664", "0.0866", "0.5664" ], [ "0.10297", "0.5684", "0.4878", "0.5684", "0.4878", "0.582", "0.10297", "0.582" ], [ "0.5146", "0.563", "0.9126", "0.563", "0.9126", "0.5737", "0.5146", "0.5737" ], [ "0.513", "0.5757", "0.914", "0.5757", "0.914", "0.589", "0.513", "0.589" ], [ "0.08417", "0.582", "0.486", "0.582", "0.486", "0.595", "0.08417", "0.595" ], [ "0.511", "0.5884", "0.9126", "0.589", "0.9126", "0.604", "0.511", "0.603" ], [ "0.08496", "0.5967", "0.486", "0.5967", "0.486", "0.61", "0.08496", "0.61" ], [ "0.08496", "0.61", "0.429", "0.61", "0.429", "0.6235", "0.08496", "0.6235" ], [ "0.513", "0.604", "0.914", "0.604", "0.914", "0.617", "0.513", "0.617" ], [ "0.5107", "0.6157", "0.915", "0.613", "0.915", "0.632", "0.5107", "0.634" ], [ "0.513", "0.632", "0.812", "0.632", "0.812", "0.646", "0.513", "0.646" ], [ "0.08496", "0.64", "0.1111", "0.64", "0.1111", "0.6514", "0.08496", "0.6514" ], [ "0.10864", "0.643", "0.1136", "0.643", "0.1136", "0.6484", "0.10864", "0.6484" ], [ "0.1095", "0.6396", "0.2402", "0.6396", "0.2402", "0.654", "0.1095", "0.654" ], [ "0.511", "0.651", "0.796", "0.6514", "0.796", "0.6665", "0.511", "0.6655" ], [ "0.08496", "0.66", "0.486", "0.66", "0.486", "0.674", "0.08496", "0.674" ], [ "0.0866", "0.675", "0.4878", "0.675", "0.4878", "0.688", "0.0866", "0.688" ], [ "0.511", "0.6704", "0.9126", "0.6694", "0.9126", "0.6846", "0.511", "0.685" ], [ "0.08496", "0.688", "0.4885", "0.688", "0.4885", "0.702", "0.08496", "0.702" ], [ "0.513", "0.685", "0.9126", "0.685", "0.9126", "0.6987", "0.513", "0.6987" ], [ "0.0866", "0.702", "0.486", "0.702", "0.486", "0.7153", "0.0866", "0.7153" ], [ "0.511", "0.6997", "0.9126", "0.6997", "0.9126", "0.7134", "0.511", "0.7134" ], [ "0.0866", "0.7163", "0.486", "0.7163", "0.486", "0.73", "0.0866", "0.73" ], [ "0.513", "0.7134", "0.914", "0.7134", "0.914", "0.7266", "0.513", "0.7266" ], [ "0.0866", "0.731", "0.4878", "0.731", "0.4878", "0.741", "0.0866", "0.741" ], [ "0.513", "0.728", "0.9126", "0.728", "0.9126", "0.741", "0.513", "0.741" ], [ "0.08496", "0.7437", "0.486", "0.7437", "0.486", "0.757", "0.08496", "0.757" ], [ "0.513", "0.741", "0.9126", "0.741", "0.9126", "0.755", "0.513", "0.755" ], [ "0.08496", "0.7583", "0.2418", "0.7583", "0.2418", "0.7715", "0.08496", "0.7715" ], [ "0.513", "0.755", "0.914", "0.755", "0.914", "0.768", "0.513", "0.768" ], [ "0.1013", "0.773", "0.4878", "0.773", "0.4878", "0.7827", "0.1013", "0.7827" ], [ "0.5146", "0.771", "0.9126", "0.771", "0.9126", "0.781", "0.5146", "0.781" ], [ "0.08496", "0.785", "0.4878", "0.785", "0.4878", "0.799", "0.08496", "0.799" ], [ "0.513", "0.7827", "0.914", "0.7827", "0.914", "0.797", "0.513", "0.797" ], [ "0.08496", "0.8", "0.4878", "0.8", "0.4878", "0.813", "0.08496", "0.813" ], [ "0.511", "0.797", "0.915", "0.797", "0.915", "0.811", "0.511", "0.811" ], [ "0.08496", "0.813", "0.2361", "0.813", "0.2361", "0.827", "0.08496", "0.827" ], [ "0.5146", "0.8125", "0.5825", "0.8125", "0.5825", "0.8237", "0.5146", "0.8237" ], [ "0.511", "0.83", "0.7295", "0.8306", "0.7295", "0.845", "0.511", "0.8447" ], [ "0.0866", "0.839", "0.3464", "0.839", "0.3464", "0.853", "0.0866", "0.853" ], [ "0.0866", "0.857", "0.4844", "0.857", "0.4844", "0.872", "0.0866", "0.872" ], [ "0.513", "0.851", "0.9126", "0.851", "0.9126", "0.8643", "0.513", "0.8643" ], [ "0.511", "0.8643", "0.8896", "0.8643", "0.8896", "0.8774", "0.511", "0.8774" ], [ "0.08417", "0.872", "0.4878", "0.872", "0.4878", "0.886", "0.08417", "0.886" ], [ "0.884", "0.867", "0.9087", "0.867", "0.9087", "0.875", "0.884", "0.875" ] ]

[ "2018]", "However, existing attention mechanisms only focus on the", "A", "ancho", "Positive data", "visual feature of the image regions while ignore the relative", "position information in the images. In general, if an object", "ipletloss", "Negative data", "region is closer to the center, it may express the main", "semantics of the imagewith higherprobability,while the", "F", "alText", "marginal ones may not be that important. Just as shown in", "Attentio", "Figure I(a), the main semantic part corresponding to the word", "WordFeaturee", "\"men\"locates at the center of the picture, while the peddling", "objects lie on the brink. From this observation, an intuitive", "idea is to simply pay more attention to the regions closer to", "GRU食GRU食", "", "the center.However.not all regions near the center are that", "Position Focuse", "important, as shown in Figure 1(b) which exhibits that a", "Attent", "woman (the most important object) lies in the lower left part.", "Furthermore, simply assigning attention to the region based", "Word", "on the fixed position (the center for example) cause a bad", "mbeddin", "extendibility.From above observations and considerations,", "The man is in the pool and", "we design a position feature for the region to integrate", "throwingasmallboyinto", "position information and propose an attention mechanism to", "theair", "generate the valuable position feature for each region.", "Inthispaper.anovel positionfocused attentionnetworkis", "Figure 2: The flowchart of proposed network", "developed to study the fine-grained interplay between the", "image regions and the words. Our contributions can be", "summarizedasfollows:", "the number of image regions. Since our attention mechanism", "I) A novel position feature is designed for image regions,", "is focused on the image regions, especially the objects in", "by which we integrate the position information of regions to", "image. Therefore, we detect the objects in image utilizing the", "investigate the correspondences between objects in image", "Faster R-CNN model [Ren et al., 2017]. In order to get a", "and words in sentence", "better feature representation, we feed the detected object into", "2) We propose a position focused attention mechanism to", "the ResNet-101 [He et al., 2016] pre-trained on Visual", "generate the valuable position feature for the image region,", "Genomes [Krishna et al., 2017] by Anderson et al. [Anderson", "and the position feature together with the visual feature form", "et al., 2018] to extract the visual feature. Finally, the input", "a more reliable and complete expression of the image region.", "image is represented by n D-dimensional feature vectors,", "3)Besidestwopublic datasets.wemakethefirstattempt", "andDis2048inourexperiment.", "to evaluate the application value on a practical news dataset", "Text feature.", "", "On the subject of corresponding image", "and our method achieves the state-of-art performance on all", "description, the basic item is the word in the sentence. Each", "of thesethree datasets.", "The remainder of this paper is organized as follows: In", "word isrepresented withone-hotvector,whichindicatesthe", "index in the vocabulary. Then the one-hot representation is", "section 2, we elaborate the details of each process in our", "embedded into d-dimensional vector by a linear mapping", "system.Experiments are shown in section 3. Finally,", "conclusions and future works are given in section 4.", "layer, xt = Wxwt, where Wt is a one-hot of word in a", "sentence with T words (wi,w2,..,wr), w e RdxN is the", "embedding matrix, N is the vocabulary size.", "2", "", " Our Approach", "2.2PositionInformation Integration", "In this section, we will elaborate the details of our proposed", "framework. Figure 2 shows the flowchart of this paper, we", "The relative position of the object in the whole image is an", "first extract the features of the region and the position, the", "important and useful clue, which is helpful to infer the", "visual feature together with the generated position feature", "significance of the object region, just as shown in Figure l.", "form the final region's representation, and the alignments", "Motivated by this observation, we fuse the position", "between the region and the word are studied by a visual-", "information into the learning procedure to capture more", "textual attention [Lee et al., 2018]. The network is trained by", "reliable and credible fine-grained interplay between the", "the triplet ranking loss.", "image and the text elements. In this subsection, we present", "Next.we describe the inputrepresentation in subsection", "ourposition attention mechanism.We first introduce the", "2.1. In subsection 2.2, the position information integration is", "initial positional representation, and then elaborate the block", "presented, subsection 2.3 presents the image-sentence", "embedding. Finally, our position focused attention is", "relevance calculation.", "presented.", "Initial Position Representation", "2.1 Input Feature Representation", "Image feature.In this paper, an image I is represented by", "Given an image I = [vi, V2,..., vn], in order to reveal the", "relative position for a region Vi in the whole image I,", "a set of features (vi, V2, .., Vn], and vi E R', where n is", "we" ]

{ "arxiv_id": "1907.09748", "language": "en", "page_idx": 1 }

[ [ "0.513", "0.06757", "0.9126", "0.06757", "0.9126", "0.0808", "0.513", "0.0808" ], [ "0.513", "0.0821", "0.911", "0.0821", "0.911", "0.09534", "0.513", "0.09534" ], [ "0.513", "0.09534", "0.911", "0.09534", "0.911", "0.1092", "0.513", "0.1092" ], [ "0.513", "0.10986", "0.9126", "0.10986", "0.9126", "0.1237", "0.513", "0.1237" ], [ "0.5146", "0.125", "0.6167", "0.125", "0.6167", "0.1364", "0.5146", "0.1364" ], [ "0.5146", "0.1395", "0.634", "0.1395", "0.634", "0.1497", "0.5146", "0.1497" ], [ "0.6436", "0.1364", "0.911", "0.137", "0.911", "0.1509", "0.6436", "0.1497" ], [ "0.513", "0.1509", "0.9126", "0.1509", "0.9126", "0.1654", "0.513", "0.1654" ], [ "0.513", "0.1654", "0.9126", "0.1654", "0.9126", "0.1787", "0.513", "0.1787" ], [ "0.5146", "0.1812", "0.91", "0.1812", "0.91", "0.1913", "0.5146", "0.1913" ], [ "0.511", "0.1913", "0.9126", "0.1925", "0.9126", "0.207", "0.511", "0.2058" ], [ "0.09314", "0.1976", "0.452", "0.1976", "0.452", "0.2084", "0.09314", "0.2084" ], [ "0.09235", "0.2102", "0.4534", "0.2102", "0.4534", "0.2235", "0.09235", "0.2235" ], [ "0.513", "0.207", "0.9126", "0.207", "0.9126", "0.2203", "0.513", "0.2203" ], [ "0.0907", "0.2247", "0.4534", "0.2247", "0.4534", "0.238", "0.0907", "0.238" ], [ "0.5146", "0.2197", "0.9126", "0.2203", "0.9126", "0.2343", "0.5146", "0.2329" ], [ "0.0907", "0.2393", "0.2296", "0.2393", "0.2296", "0.2524", "0.0907", "0.2524" ], [ "0.513", "0.2343", "0.9126", "0.2343", "0.9126", "0.2474", "0.513", "0.2474" ], [ "0.511", "0.2488", "0.9087", "0.2488", "0.9087", "0.262", "0.511", "0.262" ], [ "0.513", "0.262", "0.9126", "0.262", "0.9126", "0.276", "0.513", "0.276" ], [ "0.0866", "0.2715", "0.4844", "0.2715", "0.4844", "0.2854", "0.0866", "0.2854" ], [ "0.513", "0.276", "0.9126", "0.276", "0.9126", "0.2903", "0.513", "0.2903" ], [ "0.0866", "0.2874", "0.486", "0.2874", "0.486", "0.2974", "0.0866", "0.2974" ], [ "0.513", "0.2903", "0.911", "0.2903", "0.911", "0.3037", "0.513", "0.3037" ], [ "0.08417", "0.2998", "0.4656", "0.2998", "0.4656", "0.3132", "0.08417", "0.3132" ], [ "0.5146", "0.3062", "0.911", "0.3062", "0.911", "0.3164", "0.5146", "0.3164" ], [ "0.5146", "0.318", "0.9126", "0.318", "0.9126", "0.332", "0.5146", "0.332" ], [ "0.08496", "0.3289", "0.2255", "0.3289", "0.2255", "0.3435", "0.08496", "0.3435" ], [ "0.511", "0.332", "0.9126", "0.332", "0.9126", "0.3452", "0.511", "0.3452" ], [ "0.08496", "0.351", "0.4878", "0.351", "0.4878", "0.3643", "0.08496", "0.3643" ], [ "0.5146", "0.348", "0.911", "0.348", "0.911", "0.358", "0.5146", "0.358" ], [ "0.08496", "0.3643", "0.4878", "0.3643", "0.4878", "0.3774", "0.08496", "0.3774" ], [ "0.513", "0.3599", "0.9126", "0.3599", "0.9126", "0.3738", "0.513", "0.3738" ], [ "0.08496", "0.3774", "0.486", "0.3774", "0.486", "0.3914", "0.08496", "0.3914" ], [ "0.511", "0.3738", "0.9126", "0.3738", "0.9126", "0.387", "0.511", "0.387" ], [ "0.5107", "0.3865", "0.914", "0.387", "0.914", "0.4016", "0.5107", "0.401" ], [ "0.0866", "0.3926", "0.4844", "0.3926", "0.4844", "0.406", "0.0866", "0.406" ], [ "0.08417", "0.404", "0.4878", "0.4048", "0.4878", "0.4192", "0.08417", "0.4185" ], [ "0.5107", "0.4016", "0.681", "0.4016", "0.681", "0.4153", "0.5107", "0.4153" ], [ "0.0866", "0.4216", "0.486", "0.4216", "0.486", "0.432", "0.0866", "0.432" ], [ "0.513", "0.4268", "0.732", "0.4268", "0.732", "0.44", "0.513", "0.44" ], [ "0.0866", "0.4343", "0.486", "0.4343", "0.486", "0.4475", "0.0866", "0.4475" ], [ "0.08826", "0.4495", "0.486", "0.4495", "0.486", "0.4602", "0.08826", "0.4602" ], [ "0.513", "0.449", "0.808", "0.449", "0.808", "0.4622", "0.513", "0.4622" ], [ "0.08496", "0.4622", "0.4878", "0.4622", "0.4878", "0.476", "0.08496", "0.476" ], [ "0.511", "0.4622", "0.9126", "0.461", "0.9126", "0.476", "0.511", "0.4766" ], [ "0.08496", "0.476", "0.4885", "0.476", "0.4885", "0.4893", "0.08496", "0.4893" ], [ "0.513", "0.4766", "0.914", "0.4766", "0.914", "0.4905", "0.513", "0.4905" ], [ "0.08496", "0.4893", "0.1993", "0.4893", "0.1993", "0.5024", "0.08496", "0.5024" ], [ "0.5146", "0.4924", "0.9126", "0.4924", "0.9126", "0.5024", "0.5146", "0.5024" ], [ "0.5146", "0.506", "0.9126", "0.506", "0.9126", "0.5166", "0.5146", "0.5166" ], [ "0.08496", "0.5146", "0.259", "0.5146", "0.259", "0.529", "0.08496", "0.529" ], [ "0.513", "0.518", "0.9126", "0.518", "0.9126", "0.532", "0.513", "0.532" ], [ "0.08496", "0.5327", "0.4878", "0.5327", "0.4878", "0.547", "0.08496", "0.547" ], [ "0.511", "0.5327", "0.914", "0.532", "0.914", "0.5454", "0.511", "0.547" ], [ "0.0866", "0.5474", "0.4878", "0.5474", "0.4878", "0.561", "0.0866", "0.561" ], [ "0.5156", "0.5474", "0.9126", "0.5474", "0.9126", "0.558", "0.5156", "0.558" ], [ "0.08826", "0.5625", "0.4836", "0.5625", "0.4836", "0.5728", "0.08826", "0.5728" ], [ "0.513", "0.5625", "0.911", "0.5625", "0.911", "0.5728", "0.513", "0.5728" ], [ "0.08496", "0.5737", "0.4844", "0.5737", "0.4844", "0.587", "0.08496", "0.587" ], [ "0.511", "0.5728", "0.9126", "0.5737", "0.9126", "0.5884", "0.511", "0.587" ], [ "0.08496", "0.589", "0.4885", "0.589", "0.4885", "0.603", "0.08496", "0.603" ], [ "0.5146", "0.5884", "0.9126", "0.5884", "0.9126", "0.5986", "0.5146", "0.5986" ], [ "0.08496", "0.603", "0.4844", "0.603", "0.4844", "0.616", "0.08496", "0.616" ], [ "0.513", "0.6016", "0.9126", "0.6016", "0.9126", "0.6157", "0.513", "0.6157" ], [ "0.0866", "0.617", "0.486", "0.617", "0.486", "0.631", "0.0866", "0.631" ], [ "0.511", "0.616", "0.914", "0.616", "0.914", "0.63", "0.511", "0.63" ], [ "0.08496", "0.63", "0.2957", "0.632", "0.2957", "0.6465", "0.08496", "0.6445" ], [ "0.513", "0.631", "0.568", "0.631", "0.568", "0.6416", "0.513", "0.6416" ], [ "0.0866", "0.6484", "0.2035", "0.6484", "0.2035", "0.662", "0.0866", "0.662" ], [ "0.1953", "0.6484", "0.486", "0.6484", "0.486", "0.662", "0.1953", "0.662" ], [ "0.528", "0.643", "0.9126", "0.643", "0.9126", "0.657", "0.528", "0.657" ], [ "0.511", "0.657", "0.9126", "0.6577", "0.9126", "0.672", "0.511", "0.6704" ], [ "0.08826", "0.664", "0.4844", "0.664", "0.4844", "0.675", "0.08826", "0.675" ], [ "0.5107", "0.6704", "0.9126", "0.6704", "0.9126", "0.6846", "0.5107", "0.6846" ], [ "0.08496", "0.677", "0.4844", "0.677", "0.4844", "0.69", "0.08496", "0.69" ], [ "0.0866", "0.69", "0.486", "0.69", "0.486", "0.705", "0.0866", "0.705" ], [ "0.511", "0.685", "0.6216", "0.685", "0.6216", "0.6987", "0.511", "0.6987" ], [ "0.08496", "0.7026", "0.4844", "0.7026", "0.4844", "0.7163", "0.08496", "0.7163" ], [ "0.527", "0.6987", "0.9126", "0.6997", "0.9126", "0.7134", "0.527", "0.712" ], [ "0.0866", "0.7173", "0.4878", "0.7173", "0.4878", "0.731", "0.0866", "0.731" ], [ "0.511", "0.7134", "0.9126", "0.7134", "0.9126", "0.7266", "0.511", "0.7266" ], [ "0.08417", "0.731", "0.4878", "0.732", "0.4878", "0.7456", "0.08417", "0.744" ], [ "0.511", "0.726", "0.911", "0.726", "0.911", "0.7393", "0.511", "0.7393" ], [ "0.511", "0.7393", "0.9126", "0.7407", "0.9126", "0.755", "0.511", "0.754" ], [ "0.0866", "0.7476", "0.4844", "0.7476", "0.4844", "0.7583", "0.0866", "0.7583" ], [ "0.0866", "0.7603", "0.486", "0.7603", "0.486", "0.7734", "0.0866", "0.7734" ], [ "0.5107", "0.755", "0.9126", "0.7524", "0.9126", "0.7666", "0.5107", "0.768" ], [ "0.08496", "0.7734", "0.486", "0.7734", "0.486", "0.787", "0.08496", "0.787" ], [ "0.513", "0.768", "0.911", "0.768", "0.911", "0.782", "0.513", "0.782" ], [ "0.08496", "0.787", "0.4844", "0.7886", "0.4844", "0.8", "0.08496", "0.799" ], [ "0.513", "0.7827", "0.9126", "0.7827", "0.9126", "0.797", "0.513", "0.797" ], [ "0.08417", "0.8", "0.486", "0.802", "0.486", "0.815", "0.08417", "0.813" ], [ "0.513", "0.797", "0.914", "0.797", "0.914", "0.81", "0.513", "0.81" ], [ "0.08496", "0.815", "0.4885", "0.815", "0.4885", "0.829", "0.08496", "0.829" ], [ "0.511", "0.811", "0.8486", "0.811", "0.8486", "0.8247", "0.511", "0.8247" ], [ "0.511", "0.836", "0.7075", "0.836", "0.7075", "0.8496", "0.511", "0.8496" ], [ "0.09235", "0.853", "0.4739", "0.854", "0.4739", "0.867", "0.09235", "0.866" ], [ "0.513", "0.854", "0.911", "0.854", "0.911", "0.867", "0.513", "0.867" ], [ "0.09235", "0.8687", "0.4714", "0.8687", "0.4714", "0.879", "0.09235", "0.879" ], [ "0.513", "0.8687", "0.914", "0.8687", "0.914", "0.882", "0.513", "0.882" ] ]

[ "title, and the model can automatically output several related", "candidate images for this news, which can remarkably", "alleviate the effort of editors and speed up the news publish.", "In this application scene, we only focus on the task of the", "image retrieval.", "Training details.", "Allofourexperimentsareconductedon", "a workstation with NVIDIA Tesla GPU. Adam optimization", "algorithm is used to train the overall network. The mini-batch", "size is 128.The image region is extracted bythe FasterR-", "CNN model [Ren et al., 2017], and we retain 36 detected", "Figure4:AnexemplarimageofTencent-Newswithtitle“莫", "迪希望在安保及经济方面强化与日本战略关系”（Modi", "regions for the image representation. Each image is split into", "hopes to strengthen strategic relationship with Japan in", "16x16blocks (K=16),and we set L as 15.The dimension", "security and economy)", "of joint embedding is fixed as 1024. The block index is first", "embedded into 200-dimensional space,and the original 2048", "dimensional visual vector together with 200-dimensional", "common ranking objective for image-text matching. The", "position feature is mapped into the 1024-dimensional space", "training triplets are sampled within a mini-batch by the", "by a linear projection layer. On the subject of word, the one-", "sampling strategy in [Lee et al., 2018; Faghri et al.,2018]", "hot vector is first embedded into 3o0-dimensional dense", "representation, then the dense representation is fed into the", "3Experiments", "bi-GRU whose hidden dimension is set as 1024 as well. For", "In order to demonstrate the effectiveness of our proposed", "Flickr3OkandMS-COCOdataset,thetrainingdetailsarethe", "methods,we conduct our Position Focused Attention", "same as reference [Lee et al., 2018]. On the Tencent-News", "Network (PFAN) on two public datasets: Flickr30K and MS-", "dataset, the parameter settings except the embedding size are", "the same as the Flickr30K, we set the embedding size as 512", "COCO, and a practical Chinese news dataset: Tencent-", "News.We systematically make comparisons with several", "to get better performance.", "latest start-of-the-artmethodsandthoroughlyinvestigatethe", "3.2PerformanceEvaluation", "performance of the proposed PFAN. As for the performance", "measurecriterionforsentenceretrievalorimageretrieval,we", "Comparison with the Competing Methods", "apply the commonly used recall on top H (R@H), which is", "Table 1 shows the performances of all methods on Flickr30k", "defined as the percentage of correct items in the top H", "dataset, where PFAN t-i means only employing the loss of", "retrieved results.", "attendingtexttoimagetotrainthenetworkandPFANt-i+i-t", "fusesthemodelsfromPFANt-iandPFANi-t.Asshownin", "3.1 Implement Details", "Table 1, our PFAN t-i and PFAN i-t both achieve satisfied", "Dataset..We evaluate our PFAN on the widely used and", "performance.For example, when the depth varies fromI to", "authoritative dataset Flickr30K, MS-COCO, the data splits", "10.therecallofPFANt-icanreach66.0.89.6and94.3", "forthesetwodatasetsfollowthework[Karpathyetal..2o15]", "respectively on the sentenceretrieval given animage.The", "and [Lee et al., 2018]. Besides the public datasets, a practical", "fused model PFAN t-i+i-t makes our performance step", "news dataset, Tencent-News, is also collected to evaluate the", "further and outperforms the current state-of-art method", "value of the proposed method in the practical application.", "SCAN (i-t+t-i fused), the R@1 on sentence retrieval task can", "This dataset can be used for training image-text models to", "even reach 70.0, which validates the superior of our proposed", "further support Chinese corpus.", "method.", "Tencent-News.", "For a piece of news, the title and one", "Table 2 exhibits the performances on the MS-COCO", "dataset,from which we canalsoobserve that PFAN ismore", "perfect matching image in this news make up a basic data", "outstanding than other methods and the similar conclusions", "item, and the title is regarded as the description of this image.", "By this way, we collect 143,317 training pairs, and 1,000", "can be obtained.", "pairs for validating and there are 141,736 different images", "Table3shows theperformances onTencent-News dataset.", "130,230 different titles in total. In the test procedure, we", "In order to evaluate the retrieved images by a given news title,", "manually label 510 news and several corresponding images", "we use the Mean Average Precision (MAP) [Qian et al., 2017:", "Wang et al., 2018J and the Accuracy (A) to evaluate the", "(≥5)for performance evaluation.There are 510 titles and", "2,794 labeled images in total. Each title has 5.5 candidate", "performance.It is clearfromTable3thatourPFANis still", "images, 2.3 irrelevant images and 3.2 relevant images on", "more outstanding than the state-of-art method SCAN, PFAN", "average.Figure 4 shows an exemplar image of this dataset.", "can outperform SCAN by six points on average under both", "In this practical dataset, we focus on the news auto-image", "MAP and Accuracy. Tables 1-3 show the effectiveness and", "recommendation task, i.e. the news editor inputs the news", "the practical applied value of the proposed method.", "3.3Result Visualization", "IThe Tencent News data download link and our code can be found at:", "In this subsection, we will visualize some results to see the", "https:/github.com/HaoYang0123/Position-Focused-Attention-Network/", "ability of the proposed network on relationship capture aspect" ]

{ "arxiv_id": "1907.09748", "language": "en", "page_idx": 3 }

[ [ "0.429", "0.0663", "0.5767", "0.0663", "0.5767", "0.0808", "0.429", "0.0808" ], [ "0.6567", "0.0663", "0.8057", "0.0663", "0.8057", "0.0802", "0.6567", "0.0802" ], [ "0.246", "0.07385", "0.298", "0.07385", "0.298", "0.0852", "0.246", "0.0852" ], [ "0.4126", "0.07764", "0.4453", "0.0802", "0.4436", "0.09216", "0.411", "0.0903" ], [ "0.4885", "0.0802", "0.524", "0.0802", "0.524", "0.09155", "0.4885", "0.09155" ], [ "0.5605", "0.0802", "0.603", "0.0802", "0.603", "0.09155", "0.5605", "0.09155" ], [ "0.6313", "0.0802", "0.6685", "0.0802", "0.6685", "0.09155", "0.6313", "0.09155" ], [ "0.7075", "0.07764", "0.7427", "0.0802", "0.7417", "0.09216", "0.706", "0.0903" ], [ "0.785", "0.0802", "0.8286", "0.0802", "0.8286", "0.09155", "0.785", "0.09155" ], [ "0.1765", "0.09155", "0.3652", "0.09155", "0.3652", "0.1048", "0.1765", "0.1048" ], [ "0.4927", "0.09155", "0.523", "0.09155", "0.523", "0.1042", "0.4927", "0.1042" ], [ "0.5664", "0.09155", "0.5947", "0.09155", "0.5947", "0.1042", "0.5664", "0.1042" ], [ "0.6357", "0.09155", "0.664", "0.09155", "0.664", "0.1042", "0.6357", "0.1042" ], [ "0.71", "0.09155", "0.7417", "0.09155", "0.7417", "0.1042", "0.71", "0.1042" ], [ "0.7935", "0.09155", "0.822", "0.09155", "0.822", "0.1042", "0.7935", "0.1042" ], [ "0.165", "0.1029", "0.3757", "0.1029", "0.3757", "0.11615", "0.165", "0.11615" ], [ "0.415", "0.09216", "0.4436", "0.09216", "0.4436", "0.1042", "0.415", "0.1042" ], [ "0.4143", "0.1029", "0.4436", "0.1029", "0.4436", "0.11554", "0.4143", "0.11554" ], [ "0.492", "0.1029", "0.523", "0.1029", "0.523", "0.11554", "0.492", "0.11554" ], [ "0.634", "0.1029", "0.665", "0.1029", "0.665", "0.11554", "0.634", "0.11554" ], [ "0.71", "0.1029", "0.7417", "0.1029", "0.7417", "0.11554", "0.71", "0.11554" ], [ "0.564", "0.1111", "0.5986", "0.11426", "0.596", "0.1288", "0.5605", "0.1256" ], [ "0.79", "0.11237", "0.8228", "0.11237", "0.8228", "0.1276", "0.79", "0.1276" ], [ "0.4143", "0.11426", "0.4453", "0.11426", "0.4453", "0.127", "0.4143", "0.127" ], [ "0.2059", "0.11426", "0.357", "0.11426", "0.357", "0.1276", "0.2059", "0.1276" ], [ "0.492", "0.11426", "0.523", "0.11426", "0.523", "0.127", "0.492", "0.127" ], [ "0.6357", "0.11426", "0.665", "0.11426", "0.665", "0.127", "0.6357", "0.127" ], [ "0.71", "0.11426", "0.7417", "0.11426", "0.7417", "0.127", "0.71", "0.127" ], [ "0.1896", "0.1225", "0.3586", "0.125", "0.3586", "0.1395", "0.1896", "0.137" ], [ "0.415", "0.125", "0.4436", "0.125", "0.4436", "0.137", "0.415", "0.137" ], [ "0.5645", "0.1256", "0.596", "0.1256", "0.596", "0.1383", "0.5645", "0.1383" ], [ "0.6357", "0.1256", "0.6665", "0.1256", "0.6665", "0.1383", "0.6357", "0.1383" ], [ "0.7915", "0.125", "0.822", "0.125", "0.822", "0.137", "0.7915", "0.137" ], [ "0.4143", "0.1338", "0.4478", "0.137", "0.4453", "0.1515", "0.411", "0.1483" ], [ "0.564", "0.1338", "0.5986", "0.137", "0.596", "0.1515", "0.5605", "0.1483" ], [ "0.634", "0.1351", "0.664", "0.137", "0.6626", "0.1497", "0.6323", "0.1477" ], [ "0.4927", "0.1364", "0.523", "0.1364", "0.523", "0.1483", "0.4927", "0.1483" ], [ "0.1936", "0.1364", "0.3472", "0.1364", "0.3472", "0.1509", "0.1936", "0.1509" ], [ "0.71", "0.137", "0.7417", "0.137", "0.7417", "0.1497", "0.71", "0.1497" ], [ "0.7915", "0.1364", "0.8228", "0.1364", "0.8228", "0.1483", "0.7915", "0.1483" ], [ "0.1936", "0.1477", "0.3472", "0.1477", "0.3472", "0.1622", "0.1936", "0.1622" ], [ "0.415", "0.1477", "0.4436", "0.1477", "0.4436", "0.1604", "0.415", "0.1604" ], [ "0.492", "0.1477", "0.523", "0.1477", "0.523", "0.1604", "0.492", "0.1604" ], [ "0.564", "0.1465", "0.597", "0.1465", "0.597", "0.1622", "0.564", "0.1622" ], [ "0.6357", "0.1477", "0.664", "0.1477", "0.664", "0.1604", "0.6357", "0.1604" ], [ "0.71", "0.1477", "0.74", "0.1477", "0.74", "0.1604", "0.71", "0.1604" ], [ "0.7915", "0.1483", "0.822", "0.1483", "0.822", "0.1604", "0.7915", "0.1604" ], [ "0.1936", "0.1578", "0.343", "0.1578", "0.343", "0.1718", "0.1936", "0.1718" ], [ "0.4143", "0.159", "0.4436", "0.159", "0.4436", "0.1718", "0.4143", "0.1718" ], [ "0.492", "0.159", "0.523", "0.159", "0.523", "0.1718", "0.492", "0.1718" ], [ "0.5664", "0.159", "0.596", "0.159", "0.596", "0.1718", "0.5664", "0.1718" ], [ "0.6357", "0.159", "0.6665", "0.159", "0.6665", "0.1718", "0.6357", "0.1718" ], [ "0.71", "0.159", "0.74", "0.159", "0.74", "0.1718", "0.71", "0.1718" ], [ "0.79", "0.159", "0.822", "0.159", "0.822", "0.1718", "0.79", "0.1718" ], [ "0.4143", "0.1718", "0.4453", "0.1718", "0.4453", "0.1843", "0.4143", "0.1843" ], [ "0.492", "0.1718", "0.523", "0.1718", "0.523", "0.1843", "0.492", "0.1843" ], [ "0.5645", "0.1718", "0.596", "0.1718", "0.596", "0.1843", "0.5645", "0.1843" ], [ "0.6323", "0.1704", "0.6685", "0.1704", "0.6685", "0.1863", "0.6323", "0.1863" ], [ "0.71", "0.1718", "0.7417", "0.1718", "0.7417", "0.1843", "0.71", "0.1843" ], [ "0.79", "0.1718", "0.822", "0.1718", "0.822", "0.1843", "0.79", "0.1843" ], [ "0.5645", "0.1812", "0.597", "0.1812", "0.597", "0.1963", "0.5645", "0.1963" ], [ "0.2427", "0.1724", "0.3015", "0.1724", "0.3015", "0.1843", "0.2427", "0.1843" ], [ "0.4143", "0.1818", "0.4453", "0.1818", "0.4453", "0.1976", "0.4143", "0.1976" ], [ "0.634", "0.1818", "0.6665", "0.1818", "0.6665", "0.1976", "0.634", "0.1976" ], [ "0.709", "0.1818", "0.7417", "0.1818", "0.7417", "0.1976", "0.709", "0.1976" ], [ "0.79", "0.1818", "0.8228", "0.1818", "0.8228", "0.1976", "0.79", "0.1976" ], [ "0.492", "0.1831", "0.523", "0.1831", "0.523", "0.1957", "0.492", "0.1957" ], [ "0.2427", "0.1831", "0.3015", "0.1831", "0.3015", "0.1945", "0.2427", "0.1945" ], [ "0.2336", "0.1945", "0.3096", "0.1945", "0.3096", "0.2058", "0.2336", "0.2058" ], [ "0.415", "0.1945", "0.4436", "0.1945", "0.4436", "0.207", "0.415", "0.207" ], [ "0.4927", "0.1945", "0.5215", "0.1945", "0.5215", "0.207", "0.4927", "0.207" ], [ "0.5664", "0.1945", "0.5947", "0.1945", "0.5947", "0.207", "0.5664", "0.207" ], [ "0.6357", "0.1945", "0.664", "0.1945", "0.664", "0.207", "0.6357", "0.207" ], [ "0.7114", "0.1945", "0.74", "0.1945", "0.74", "0.207", "0.7114", "0.207" ], [ "0.7915", "0.1945", "0.8203", "0.1945", "0.8203", "0.207", "0.7915", "0.207" ], [ "0.2115", "0.209", "0.7876", "0.209", "0.7876", "0.2235", "0.2115", "0.2235" ], [ "0.4307", "0.2374", "0.5767", "0.2374", "0.5767", "0.252", "0.4307", "0.252" ], [ "0.66", "0.2374", "0.8057", "0.2374", "0.8057", "0.2507", "0.66", "0.2507" ], [ "0.2443", "0.2443", "0.298", "0.2443", "0.298", "0.2556", "0.2443", "0.2556" ], [ "0.5605", "0.2488", "0.603", "0.2488", "0.603", "0.2632", "0.5605", "0.2632" ], [ "0.709", "0.25", "0.7427", "0.25", "0.7427", "0.262", "0.709", "0.262" ], [ "0.4902", "0.2507", "0.524", "0.2507", "0.524", "0.262", "0.4902", "0.262" ], [ "0.4143", "0.2507", "0.4453", "0.2507", "0.4453", "0.262", "0.4143", "0.262" ], [ "0.6323", "0.2507", "0.665", "0.2507", "0.665", "0.262", "0.6323", "0.262" ], [ "0.7876", "0.2507", "0.8267", "0.2507", "0.8267", "0.262", "0.7876", "0.262" ], [ "0.1772", "0.2615", "0.3635", "0.2615", "0.3635", "0.2747", "0.1772", "0.2747" ], [ "0.415", "0.262", "0.4436", "0.262", "0.4436", "0.2747", "0.415", "0.2747" ], [ "0.492", "0.262", "0.5215", "0.262", "0.5215", "0.2747", "0.492", "0.2747" ], [ "0.5664", "0.262", "0.5947", "0.262", "0.5947", "0.2747", "0.5664", "0.2747" ], [ "0.6357", "0.262", "0.664", "0.262", "0.664", "0.2747", "0.6357", "0.2747" ], [ "0.71", "0.262", "0.74", "0.262", "0.74", "0.2747", "0.71", "0.2747" ], [ "0.7915", "0.262", "0.822", "0.262", "0.822", "0.2747", "0.7915", "0.2747" ], [ "0.165", "0.2727", "0.3757", "0.2727", "0.3757", "0.286", "0.165", "0.286" ], [ "0.415", "0.2727", "0.4436", "0.2727", "0.4436", "0.2854", "0.415", "0.2854" ], [ "0.4927", "0.2727", "0.523", "0.2727", "0.523", "0.2854", "0.4927", "0.2854" ], [ "0.6357", "0.2727", "0.664", "0.2727", "0.664", "0.2854", "0.6357", "0.2854" ], [ "0.71", "0.2727", "0.74", "0.2727", "0.74", "0.2854", "0.71", "0.2854" ], [ "0.492", "0.2827", "0.524", "0.2827", "0.524", "0.2986", "0.492", "0.2986" ], [ "0.564", "0.2827", "0.597", "0.2827", "0.597", "0.2986", "0.564", "0.2986" ], [ "0.709", "0.2827", "0.7417", "0.2827", "0.7417", "0.2986", "0.709", "0.2986" ], [ "0.415", "0.2842", "0.4453", "0.2842", "0.4453", "0.2966", "0.415", "0.2966" ], [ "0.2059", "0.2842", "0.357", "0.2842", "0.357", "0.2974", "0.2059", "0.2974" ], [ "0.6357", "0.2842", "0.665", "0.2842", "0.665", "0.2966", "0.6357", "0.2966" ], [ "0.7915", "0.2842", "0.822", "0.2842", "0.822", "0.2966", "0.7915", "0.2966" ], [ "0.1912", "0.2954", "0.3594", "0.2954", "0.3594", "0.3093", "0.1912", "0.3093" ], [ "0.564", "0.2922", "0.5986", "0.2954", "0.596", "0.31", "0.5605", "0.307" ], [ "0.634", "0.2954", "0.665", "0.2954", "0.665", "0.308", "0.634", "0.308" ], [ "0.7915", "0.2954", "0.822", "0.2954", "0.822", "0.308", "0.7915", "0.308" ], [ "0.6323", "0.3062", "0.6665", "0.3062", "0.6665", "0.3213", "0.6323", "0.3213" ], [ "0.4143", "0.2966", "0.4453", "0.2966", "0.4453", "0.3093", "0.4143", "0.3093" ], [ "0.1936", "0.307", "0.3472", "0.307", "0.3472", "0.3208", "0.1936", "0.3208" ], [ "0.4927", "0.307", "0.523", "0.307", "0.523", "0.3193", "0.4927", "0.3193" ], [ "0.5645", "0.307", "0.596", "0.307", "0.596", "0.3193", "0.5645", "0.3193" ], [ "0.71", "0.307", "0.7417", "0.307", "0.7417", "0.3193", "0.71", "0.3193" ], [ "0.7915", "0.307", "0.8228", "0.307", "0.8228", "0.3193", "0.7915", "0.3193" ], [ "0.1936", "0.318", "0.3472", "0.318", "0.3472", "0.332", "0.1936", "0.332" ], [ "0.4143", "0.308", "0.4453", "0.308", "0.4453", "0.3193", "0.4143", "0.3193" ], [ "0.4143", "0.318", "0.4453", "0.318", "0.4453", "0.3308", "0.4143", "0.3308" ], [ "0.492", "0.318", "0.523", "0.318", "0.523", "0.3308", "0.492", "0.3308" ], [ "0.5645", "0.318", "0.596", "0.318", "0.596", "0.3308", "0.5645", "0.3308" ], [ "0.634", "0.318", "0.665", "0.318", "0.665", "0.3308", "0.634", "0.3308" ], [ "0.71", "0.318", "0.74", "0.318", "0.74", "0.3308", "0.71", "0.3308" ], [ "0.79", "0.318", "0.8228", "0.318", "0.8228", "0.3308", "0.79", "0.3308" ], [ "0.1936", "0.3276", "0.3447", "0.33", "0.3447", "0.3447", "0.1936", "0.3423" ], [ "0.6357", "0.3276", "0.665", "0.33", "0.664", "0.3435", "0.634", "0.3416" ], [ "0.7915", "0.3276", "0.8228", "0.33", "0.822", "0.3435", "0.79", "0.3416" ], [ "0.492", "0.33", "0.523", "0.33", "0.523", "0.3423", "0.492", "0.3423" ], [ "0.4143", "0.33", "0.4436", "0.33", "0.4436", "0.3423", "0.4143", "0.3423" ], [ "0.5645", "0.33", "0.596", "0.33", "0.596", "0.3423", "0.5645", "0.3423" ], [ "0.71", "0.3308", "0.7417", "0.3308", "0.7417", "0.3435", "0.71", "0.3435" ], [ "0.415", "0.3403", "0.4436", "0.3423", "0.4429", "0.356", "0.4143", "0.3542" ], [ "0.4927", "0.3423", "0.523", "0.3423", "0.523", "0.3547", "0.4927", "0.3547" ], [ "0.5645", "0.3423", "0.596", "0.3423", "0.596", "0.3547", "0.5645", "0.3547" ], [ "0.709", "0.3416", "0.7417", "0.3416", "0.7417", "0.3567", "0.709", "0.3567" ], [ "0.7915", "0.3423", "0.822", "0.3423", "0.822", "0.3547", "0.7915", "0.3547" ], [ "0.634", "0.3435", "0.6665", "0.3435", "0.6665", "0.356", "0.634", "0.356" ], [ "0.2427", "0.3435", "0.3015", "0.3435", "0.3015", "0.3547", "0.2427", "0.3547" ], [ "0.2402", "0.3516", "0.3022", "0.3542", "0.3022", "0.3687", "0.2402", "0.3662" ], [ "0.4143", "0.3528", "0.4453", "0.3528", "0.4453", "0.3694", "0.4143", "0.3694" ], [ "0.492", "0.3528", "0.523", "0.3528", "0.523", "0.3687", "0.492", "0.3687" ], [ "0.5645", "0.3528", "0.597", "0.3528", "0.597", "0.3687", "0.5645", "0.3687" ], [ "0.634", "0.3528", "0.6685", "0.3528", "0.6685", "0.3687", "0.634", "0.3687" ], [ "0.709", "0.3528", "0.7417", "0.3528", "0.7417", "0.3687", "0.709", "0.3687" ], [ "0.79", "0.351", "0.8228", "0.3542", "0.8203", "0.3706", "0.7876", "0.3674" ], [ "0.634", "0.363", "0.6665", "0.3655", "0.664", "0.3806", "0.6323", "0.379" ], [ "0.79", "0.363", "0.822", "0.363", "0.822", "0.379", "0.79", "0.379" ], [ "0.492", "0.3643", "0.524", "0.3643", "0.524", "0.3801", "0.492", "0.3801" ], [ "0.4143", "0.3643", "0.4453", "0.3643", "0.4453", "0.3801", "0.4143", "0.3801" ], [ "0.5645", "0.3643", "0.596", "0.3643", "0.596", "0.3801", "0.5645", "0.3801" ], [ "0.709", "0.3643", "0.7417", "0.3643", "0.7417", "0.3801", "0.709", "0.3801" ], [ "0.2091", "0.3806", "0.79", "0.3806", "0.79", "0.3958", "0.2091", "0.3958" ], [ "0.232", "0.3655", "0.3105", "0.3655", "0.3105", "0.377", "0.232", "0.377" ], [ "0.1422", "0.4153", "0.2035", "0.4153", "0.2035", "0.4268", "0.1422", "0.4268" ], [ "0.2075", "0.4153", "0.2695", "0.4153", "0.2695", "0.4268", "0.2075", "0.4268" ], [ "0.273", "0.4153", "0.3325", "0.4153", "0.3325", "0.4268", "0.273", "0.4268" ], [ "0.3381", "0.4153", "0.3743", "0.4153", "0.3743", "0.4268", "0.3381", "0.4268" ], [ "0.3896", "0.4153", "0.4346", "0.4153", "0.4346", "0.4268", "0.3896", "0.4268" ], [ "0.4355", "0.4153", "0.478", "0.4153", "0.478", "0.4268", "0.4355", "0.4268" ], [ "0.5947", "0.4224", "0.6167", "0.4216", "0.6167", "0.4255", "0.5947", "0.4268" ], [ "0.701", "0.4216", "0.7812", "0.4216", "0.7812", "0.4255", "0.701", "0.4255" ], [ "0.2255", "0.428", "0.2532", "0.428", "0.2532", "0.4426", "0.2255", "0.4426" ], [ "0.29", "0.428", "0.3162", "0.428", "0.3162", "0.4414", "0.29", "0.4414" ], [ "0.3472", "0.428", "0.3774", "0.428", "0.3774", "0.44", "0.3472", "0.44" ], [ "0.3938", "0.428", "0.429", "0.428", "0.429", "0.44", "0.3938", "0.44" ], [ "0.4412", "0.4268", "0.4763", "0.4287", "0.4756", "0.4414", "0.4395", "0.4395" ], [ "0.1602", "0.4287", "0.1854", "0.4287", "0.1854", "0.4414", "0.1602", "0.4414" ], [ "0.0866", "0.4287", "0.1365", "0.4287", "0.1365", "0.44", "0.0866", "0.44" ], [ "0.786", "0.4312", "0.7935", "0.4312", "0.7935", "0.4363", "0.786", "0.4363" ], [ "0.0907", "0.4414", "0.1365", "0.4414", "0.1365", "0.4526", "0.0907", "0.4526" ], [ "0.1602", "0.4414", "0.1871", "0.4414", "0.1871", "0.4526", "0.1602", "0.4526" ], [ "0.2255", "0.4395", "0.2532", "0.4414", "0.2507", "0.4539", "0.2239", "0.4507" ], [ "0.2893", "0.4414", "0.3179", "0.4414", "0.3179", "0.4526", "0.2893", "0.4526" ], [ "0.3489", "0.4414", "0.3774", "0.4414", "0.3774", "0.4526", "0.3489", "0.4526" ], [ "0.3962", "0.4414", "0.4272", "0.4414", "0.4272", "0.4526", "0.3962", "0.4526" ], [ "0.4395", "0.4414", "0.4756", "0.4414", "0.4756", "0.4526", "0.4395", "0.4526" ], [ "0.8228", "0.4414", "0.841", "0.4414", "0.841", "0.4495", "0.8228", "0.4495" ], [ "0.1438", "0.457", "0.433", "0.457", "0.433", "0.4702", "0.1438", "0.4702" ], [ "0.8228", "0.4685", "0.841", "0.4685", "0.841", "0.4766", "0.8228", "0.4766" ], [ "0.0866", "0.5005", "0.3245", "0.5005", "0.3245", "0.5146", "0.0866", "0.5146" ], [ "0.822", "0.4956", "0.841", "0.4956", "0.841", "0.504", "0.822", "0.504" ], [ "0.08496", "0.5146", "0.4902", "0.5146", "0.4902", "0.528", "0.08496", "0.528" ], [ "0.8228", "0.523", "0.841", "0.523", "0.841", "0.531", "0.8228", "0.531" ], [ "0.0866", "0.53", "0.4885", "0.53", "0.4885", "0.5405", "0.0866", "0.5405" ], [ "0.08496", "0.5425", "0.4885", "0.5425", "0.4885", "0.556", "0.08496", "0.556" ], [ "0.8228", "0.55", "0.841", "0.55", "0.841", "0.558", "0.8228", "0.558" ], [ "0.08496", "0.5566", "0.4902", "0.5566", "0.4902", "0.571", "0.08496", "0.571" ], [ "0.08496", "0.571", "0.4885", "0.571", "0.4885", "0.584", "0.08496", "0.584" ], [ "0.0866", "0.584", "0.4878", "0.584", "0.4878", "0.5977", "0.0866", "0.5977" ], [ "0.5337", "0.579", "0.892", "0.58", "0.892", "0.5947", "0.5337", "0.5933" ], [ "0.08417", "0.5933", "0.492", "0.5967", "0.492", "0.6143", "0.08417", "0.6113" ], [ "0.08496", "0.6123", "0.4902", "0.6123", "0.4902", "0.6255", "0.08496", "0.6255" ], [ "0.08496", "0.6255", "0.4902", "0.6255", "0.4902", "0.6396", "0.08496", "0.6396" ], [ "0.509", "0.6235", "0.914", "0.625", "0.914", "0.638", "0.509", "0.637" ], [ "0.0866", "0.6416", "0.1708", "0.6416", "0.1708", "0.653", "0.0866", "0.653" ], [ "0.5107", "0.638", "0.914", "0.638", "0.914", "0.6514", "0.5107", "0.6514" ], [ "0.3652", "0.657", "0.4192", "0.657", "0.4192", "0.6685", "0.3652", "0.6685" ], [ "0.5107", "0.6514", "0.914", "0.6514", "0.914", "0.6655", "0.5107", "0.6655" ], [ "0.134", "0.662", "0.2075", "0.662", "0.2075", "0.677", "0.134", "0.677" ], [ "0.2296", "0.66", "0.2378", "0.66", "0.2378", "0.6655", "0.2296", "0.6655" ], [ "0.2532", "0.6655", "0.3464", "0.662", "0.3464", "0.677", "0.2532", "0.6797" ], [ "0.4592", "0.6636", "0.4885", "0.6636", "0.4885", "0.6753", "0.4592", "0.6753" ], [ "0.509", "0.6665", "0.914", "0.6665", "0.914", "0.6797", "0.509", "0.6797" ], [ "0.3816", "0.6704", "0.4043", "0.6704", "0.4043", "0.6816", "0.3816", "0.6816" ], [ "0.2091", "0.6694", "0.2566", "0.6724", "0.255", "0.6846", "0.2075", "0.681" ], [ "0.0866", "0.6816", "0.492", "0.6816", "0.492", "0.696", "0.0866", "0.696" ], [ "0.5107", "0.6797", "0.914", "0.6797", "0.914", "0.6934", "0.5107", "0.6934" ], [ "0.08496", "0.6963", "0.4902", "0.6963", "0.4902", "0.7104", "0.08496", "0.7104" ], [ "0.509", "0.6934", "0.9126", "0.6934", "0.9126", "0.707", "0.509", "0.707" ], [ "0.08496", "0.7104", "0.492", "0.7104", "0.492", "0.7236", "0.08496", "0.7236" ], [ "0.509", "0.7085", "0.914", "0.7085", "0.914", "0.7217", "0.509", "0.7217" ], [ "0.08496", "0.7246", "0.492", "0.7246", "0.492", "0.738", "0.08496", "0.738" ], [ "0.5063", "0.72", "0.9126", "0.72", "0.9126", "0.7344", "0.5063", "0.7344" ], [ "0.509", "0.735", "0.841", "0.735", "0.841", "0.7485", "0.509", "0.7485" ], [ "0.834", "0.7363", "0.9126", "0.7363", "0.9126", "0.7476", "0.834", "0.7476" ], [ "0.1013", "0.752", "0.4902", "0.752", "0.4902", "0.765", "0.1013", "0.765" ], [ "0.08496", "0.738", "0.2303", "0.7373", "0.2303", "0.7505", "0.08496", "0.752" ], [ "0.509", "0.752", "0.5923", "0.752", "0.5923", "0.763", "0.509", "0.763" ], [ "0.0866", "0.7676", "0.4885", "0.7676", "0.4885", "0.778", "0.0866", "0.778" ], [ "0.511", "0.771", "0.675", "0.771", "0.675", "0.784", "0.511", "0.784" ], [ "0.08496", "0.78", "0.4902", "0.78", "0.4902", "0.7935", "0.08496", "0.7935" ], [ "0.5073", "0.784", "0.91", "0.784", "0.91", "0.7974", "0.5073", "0.7974" ], [ "0.08496", "0.7935", "0.4885", "0.7935", "0.4885", "0.8066", "0.08496", "0.8066" ], [ "0.509", "0.799", "0.9126", "0.799", "0.9126", "0.8125", "0.509", "0.8125" ], [ "0.0866", "0.8096", "0.4902", "0.8096", "0.4902", "0.8193", "0.0866", "0.8193" ], [ "0.5107", "0.8145", "0.9126", "0.8145", "0.9126", "0.8247", "0.5107", "0.8247" ], [ "0.08496", "0.8213", "0.4182", "0.8213", "0.4182", "0.8354", "0.08496", "0.8354" ], [ "0.5073", "0.8257", "0.914", "0.827", "0.914", "0.8403", "0.5073", "0.839" ], [ "0.10297", "0.8354", "0.4902", "0.8354", "0.4902", "0.8486", "0.10297", "0.8486" ], [ "0.5073", "0.8403", "0.706", "0.8403", "0.706", "0.854", "0.5073", "0.854" ], [ "0.08496", "0.8496", "0.4885", "0.8496", "0.4885", "0.863", "0.08496", "0.863" ], [ "0.527", "0.855", "0.9126", "0.855", "0.9126", "0.8687", "0.527", "0.8687" ], [ "0.08496", "0.863", "0.4885", "0.863", "0.4885", "0.877", "0.08496", "0.877" ], [ "0.509", "0.8687", "0.914", "0.8687", "0.914", "0.882", "0.509", "0.882" ] ]

[ "Image-to-Text Retrieval", "Text-to-Image Retrieval", "methods", "R@1", "R@5", "R@10", "R@1", "R@5", "R@10", "SM-LSTM[Huang et al., 2017]", "71.9", "81.5", "30.2", "60.4", "72.3", "2WayNet [Eisenschtat et al., 2017]", "42.5", "49.8", "67.5", "36.0", "55.6", "89.0", "79.1", "55.0", "DAN [Nam et al., 2017]", "81.8", "39.4", "69.2", "VSE++ [Faghri et al., 2018]", "52.9", "87.2", "39.6", "79.5", "55.6", "89.5", "39.1", "81.9", "DPC [Zheng et al., 2018]", "69.2", "80.9", "SCO [Huang et al., 2018]", "55.5", "82.0", "89.3", "41.1", "70.5", "80.1", "SCAN [Lee et al., 2018]", "67.4", "90.3", "95.8", "48.6", "77.7", "85.2", "66.0", "89.6", "94.3", "49.6", "77.0", "84.2", "93.8", "PFAN t-i", "67.6", "45.7", "74.7", "83.6", "90.0", "PFANi-t", "PFAN t-i+i-t", "70.0", "91.8", "95.0", "50.4", "78.7", "86.1", "Table 1: Comparisons of cross-modal retrieval on Flickr30K dataset with the competing methods", "Image-to-Text Retrieval", "Text-to-Image Retrieval", "methods", "R@10", "R@5", "R@5", "R@1", "R@1", "R@10", "SM-LSTM[Huang et al., 2017]", "53.2", "83.1", "91.5", "40.7", "75.8", "87.4", "2WayNet [Eisenschtat et al., 20i7]", "55.8", "75.2", "39.7", "66.3", "81.8", "89.0", "69.2", "55.0", "DAN [Nam et al., 2017]", "39.4", "79.1", "VSE++ [Faghri et al., 2018]", "95.7", "52.0", "92.0", "47.1", "64.6", "DPC [Zheng et al., 2018]", "89.8", "95.5", "79.9", "90.0", "SCo [Huang et al., 2018]", "65.6", "69.9", "92.9", "97.5", "56.7", "87.5", "94.8", "SCAN [Lee et al., 2018]", "58.8", "94.8", "94.8", "72.7", "98.4", "88.4", "75.8", "95.9", "99.0", "89.1", "95.1", "61.0", "PFANt-i", "PFANi-t", "70.7", "94.1", "97.8", "53.0", "84.5", "92.6", "61.6", "95.2", "96.3", "76.5", "99.0", "89.6", "Table 2: Comparisons of cross-modal retrieval on MS-COCO dataset with the competing methods", "PFAN t-i+i-t", "MAP@1", "MAP@3", "MAP@3", "A@1", "A@2", "A@3", "2", "1011213141", "70.6", "75.7", "67.2", "69.1", "73.6", "67.2", "SCAN", "", "PFAN", "76.0", "79.0", "82.0", "76.0", "76.3", "79.7", "0.70", "Table 3: Performances on Tencent-News dataset", "0.68", "Position Embedding Visualization", "0.66", "In this work, we employ an embedding layer to project the", "0.64", "block index into a dense representation.Intuitively, the", "learned block index embedding should preserve the locality,", "0.62", "i.e. the neighbor position embeddings should be close to each", "other. In order to see if the learned position embedding", "preserve the locality, we compute the similarity matrix SM =-", "Figure 5: The visualization of position embedding similarity", "R16x16 of the block position embeddings, and the component", "SM(i,j) is defined as the average similarity between the", "block embedding in i-th row and j-th column and its adjacent", "We guess that this is because the regions in the corners are", "embeddings:", "relatively similar, just as shown in the lower right region in", "[|bm-d]]", "Figure 1(a) and (b). Therefore, the position features in corner", "SM(i,j) =", "1.", "Zdet(bm) exp", "(12)", "frequently meet similar visual content and fit the similar", "2g2", "1c(bm)l", "the SM is calculated based on Gaussian kernel function, where", "visual content in each iteration. The final learned position", "the m = ix16 + j, L(bm) is the collection of the position", "embedding should be similar. However, the content close to", "embeddings that are directly adjacent to the bm, E R is a", "the center in different images are violently changed, the", "scalar, we simply set it as the average distance between all the", "corresponding block position needs to fit various visual", "contents, therefore, there will be obvious", "differences", "Figure 5 shows the visualization result of matrix SM, we", "position embeddings.", "accordingly.", "can find that the closer block position embeddings share", "Attention Visualization", "higher similarity in most cases. For example, the four blocks", "We design a position attention mechanism to adaptively", "marked as (a)-(d) are very similar to each other, and many", "determine the importance of the block position according to", "analogysituationscanbefoundinFigure5,whichmeansthe", "the attention weights instead of the fixed position.More", "learned position embeddings preserve the locality.", "details can be found in section 2. In this subsection, we", "We can also observe that the similarities of some position", "visualize the attention results.", "embeddings close to the center are with lower value (c), (d)),", "An exemplary visualization result is shown in Figure 6,", "while many marginal blocks are highly similar to each other.", "where the green box indicates the image region, the word with" ]

{ "arxiv_id": "1907.09748", "language": "en", "page_idx": 4 }

[ [ "0.2837", "0.08435", "0.7197", "0.08435", "0.7197", "0.0986", "0.2837", "0.0986" ], [ "0.3376", "0.10394", "0.665", "0.10394", "0.665", "0.11816", "0.3376", "0.11816" ], [ "0.2644", "0.1538", "0.413", "0.152", "0.413", "0.1692", "0.2644", "0.1716" ], [ "0.6094", "0.1538", "0.727", "0.1538", "0.727", "0.1687", "0.6094", "0.1687" ], [ "0.2183", "0.1704", "0.4517", "0.1704", "0.4517", "0.1853", "0.2183", "0.1853" ], [ "0.586", "0.1716", "0.75", "0.1716", "0.75", "0.1853", "0.586", "0.1853" ], [ "0.2747", "0.1859", "0.3972", "0.1882", "0.3972", "0.2031", "0.2747", "0.2007" ], [ "0.5557", "0.1882", "0.781", "0.1882", "0.781", "0.2019", "0.5557", "0.2019" ], [ "0.266", "0.2654", "0.3435", "0.2654", "0.3435", "0.2798", "0.266", "0.2798" ], [ "0.513", "0.2666", "0.885", "0.2666", "0.885", "0.2798", "0.513", "0.2798" ], [ "0.5146", "0.2832", "0.885", "0.2832", "0.885", "0.2964", "0.5146", "0.2964" ], [ "0.147", "0.2944", "0.4592", "0.2944", "0.4592", "0.308", "0.147", "0.308" ], [ "0.5146", "0.2986", "0.8833", "0.2986", "0.8833", "0.3123", "0.5146", "0.3123" ], [ "0.02184", "0.31", "0.05878", "0.31", "0.05878", "0.4358", "0.02184", "0.4358" ], [ "0.1453", "0.3088", "0.461", "0.308", "0.461", "0.3225", "0.1453", "0.3237" ], [ "0.511", "0.3135", "0.8833", "0.3142", "0.8833", "0.3289", "0.511", "0.328" ], [ "0.147", "0.3237", "0.461", "0.3237", "0.461", "0.3374", "0.147", "0.3374" ], [ "0.5146", "0.33", "0.8867", "0.33", "0.8867", "0.3445", "0.5146", "0.3445" ], [ "0.147", "0.3384", "0.461", "0.3384", "0.461", "0.3516", "0.147", "0.3516" ], [ "0.513", "0.3445", "0.8833", "0.3445", "0.8833", "0.3594", "0.513", "0.3594" ], [ "0.1453", "0.3516", "0.461", "0.3508", "0.461", "0.3652", "0.1453", "0.3665" ], [ "0.1486", "0.3665", "0.4592", "0.3665", "0.4592", "0.3801", "0.1486", "0.3801" ], [ "0.513", "0.3623", "0.885", "0.3623", "0.885", "0.376", "0.513", "0.376" ], [ "0.147", "0.3813", "0.461", "0.3813", "0.461", "0.3943", "0.147", "0.3943" ], [ "0.5146", "0.3801", "0.8867", "0.3801", "0.8867", "0.3938", "0.5146", "0.3938" ], [ "0.1486", "0.3967", "0.4592", "0.3967", "0.4592", "0.4067", "0.1486", "0.4067" ], [ "0.513", "0.3967", "0.613", "0.3967", "0.613", "0.408", "0.513", "0.408" ], [ "0.147", "0.4092", "0.4626", "0.4092", "0.4626", "0.4229", "0.147", "0.4229" ], [ "0.532", "0.411", "0.8833", "0.411", "0.8833", "0.4246", "0.532", "0.4246" ], [ "0.147", "0.4229", "0.4592", "0.4229", "0.4592", "0.437", "0.147", "0.437" ], [ "0.03864", "0.4316", "0.0487", "0.4316", "0.0487", "0.437", "0.03864", "0.437" ], [ "0.513", "0.4275", "0.8833", "0.4275", "0.8833", "0.4412", "0.513", "0.4412" ], [ "0.02184", "0.437", "0.06464", "0.437", "0.06464", "0.5186", "0.02184", "0.5186" ], [ "0.1486", "0.4395", "0.461", "0.4395", "0.461", "0.4495", "0.1486", "0.4495" ], [ "0.5146", "0.4448", "0.885", "0.4448", "0.885", "0.4578", "0.5146", "0.4578" ], [ "0.147", "0.452", "0.461", "0.452", "0.461", "0.4656", "0.147", "0.4656" ], [ "0.147", "0.4656", "0.461", "0.4656", "0.461", "0.4785", "0.147", "0.4785" ], [ "0.513", "0.4602", "0.8833", "0.4602", "0.8833", "0.4739", "0.513", "0.4739" ], [ "0.147", "0.4797", "0.461", "0.4797", "0.461", "0.4934", "0.147", "0.4934" ], [ "0.513", "0.4756", "0.885", "0.4744", "0.885", "0.4893", "0.513", "0.4905" ], [ "0.1486", "0.4954", "0.4592", "0.4954", "0.4592", "0.506", "0.1486", "0.506" ], [ "0.511", "0.4922", "0.881", "0.4922", "0.881", "0.506", "0.511", "0.506" ], [ "0.147", "0.5083", "0.461", "0.5083", "0.461", "0.5215", "0.147", "0.5215" ], [ "0.5146", "0.509", "0.8833", "0.509", "0.8833", "0.5225", "0.5146", "0.5225" ], [ "0.02016", "0.5215", "0.06046", "0.5215", "0.06046", "0.73", "0.02016", "0.73" ], [ "0.147", "0.5225", "0.4626", "0.5225", "0.4626", "0.536", "0.147", "0.536" ], [ "0.5146", "0.524", "0.8867", "0.524", "0.8867", "0.5376", "0.5146", "0.5376" ], [ "0.1486", "0.536", "0.4592", "0.536", "0.4592", "0.55", "0.1486", "0.55" ], [ "0.513", "0.539", "0.8867", "0.5405", "0.8867", "0.5547", "0.513", "0.554" ], [ "0.1453", "0.5503", "0.4626", "0.55", "0.4626", "0.564", "0.1453", "0.5654" ], [ "0.513", "0.5547", "0.8833", "0.5557", "0.8833", "0.571", "0.513", "0.5693" ], [ "0.147", "0.5654", "0.461", "0.5654", "0.461", "0.579", "0.147", "0.579" ], [ "0.5146", "0.5723", "0.885", "0.5723", "0.885", "0.586", "0.5146", "0.586" ], [ "0.147", "0.579", "0.4592", "0.579", "0.4592", "0.593", "0.147", "0.593" ], [ "0.5146", "0.5884", "0.885", "0.5884", "0.885", "0.6025", "0.5146", "0.6025" ], [ "0.1486", "0.5957", "0.4592", "0.5957", "0.4592", "0.6055", "0.1486", "0.6055" ], [ "0.1486", "0.608", "0.461", "0.608", "0.461", "0.6216", "0.1486", "0.6216" ], [ "0.513", "0.605", "0.885", "0.605", "0.885", "0.618", "0.513", "0.618" ], [ "0.147", "0.6206", "0.4592", "0.6206", "0.4592", "0.6343", "0.147", "0.6343" ], [ "0.513", "0.619", "0.8867", "0.6206", "0.8867", "0.6353", "0.513", "0.6343" ], [ "0.147", "0.6357", "0.461", "0.6357", "0.461", "0.6494", "0.147", "0.6494" ], [ "0.513", "0.637", "0.885", "0.637", "0.885", "0.651", "0.513", "0.651" ], [ "0.1486", "0.651", "0.4592", "0.651", "0.4592", "0.6646", "0.1486", "0.6646" ], [ "0.513", "0.652", "0.588", "0.654", "0.588", "0.6675", "0.513", "0.665" ], [ "0.147", "0.663", "0.4592", "0.663", "0.4592", "0.677", "0.147", "0.677" ], [ "0.531", "0.6694", "0.8833", "0.6685", "0.8833", "0.683", "0.531", "0.684" ], [ "0.147", "0.6797", "0.4575", "0.6787", "0.4575", "0.6904", "0.147", "0.6914" ], [ "0.513", "0.686", "0.8833", "0.686", "0.8833", "0.6997", "0.513", "0.6997" ], [ "0.147", "0.6934", "0.461", "0.6934", "0.461", "0.707", "0.147", "0.707" ], [ "0.513", "0.7017", "0.8833", "0.7017", "0.8833", "0.7163", "0.513", "0.7163" ], [ "0.1486", "0.709", "0.4575", "0.709", "0.4575", "0.7197", "0.1486", "0.7197" ], [ "0.147", "0.72", "0.4592", "0.72", "0.4592", "0.734", "0.147", "0.734" ], [ "0.513", "0.7188", "0.885", "0.7188", "0.885", "0.7324", "0.513", "0.7324" ], [ "0.147", "0.737", "0.4592", "0.737", "0.4592", "0.7476", "0.147", "0.7476" ], [ "0.513", "0.734", "0.8833", "0.734", "0.8833", "0.7476", "0.513", "0.7476" ], [ "0.1453", "0.75", "0.372", "0.75", "0.372", "0.763", "0.1453", "0.763" ], [ "0.5146", "0.7505", "0.8833", "0.7505", "0.8833", "0.764", "0.5146", "0.764" ], [ "0.5146", "0.7666", "0.8833", "0.7666", "0.8833", "0.78", "0.5146", "0.78" ], [ "0.1301", "0.78", "0.2612", "0.78", "0.2612", "0.7935", "0.1301", "0.7935" ], [ "0.511", "0.782", "0.8833", "0.781", "0.8833", "0.796", "0.511", "0.7964" ], [ "0.513", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.513", "0.8125" ], [ "0.12427", "0.783", "0.1318", "0.783", "0.1318", "0.789", "0.12427", "0.789" ], [ "0.12006", "0.807", "0.4902", "0.807", "0.4902", "0.821", "0.12006", "0.821" ], [ "0.511", "0.8135", "0.885", "0.814", "0.885", "0.829", "0.511", "0.8276" ], [ "0.1184", "0.8237", "0.4902", "0.8237", "0.4902", "0.8374", "0.1184", "0.8374" ], [ "0.513", "0.8306", "0.8833", "0.8306", "0.8833", "0.8457", "0.513", "0.8457" ], [ "0.1184", "0.839", "0.4902", "0.839", "0.4902", "0.8525", "0.1184", "0.8525" ], [ "0.511", "0.8467", "0.8833", "0.8457", "0.8833", "0.86", "0.511", "0.861" ], [ "0.12006", "0.8555", "0.4895", "0.8555", "0.4895", "0.869", "0.12006", "0.869" ], [ "0.5146", "0.8633", "0.8833", "0.8633", "0.8833", "0.877", "0.5146", "0.877" ], [ "0.1184", "0.8716", "0.4895", "0.8716", "0.4895", "0.8857", "0.1184", "0.8857" ], [ "0.513", "0.879", "0.881", "0.879", "0.881", "0.8926", "0.513", "0.8926" ], [ "0.1436", "0.8945", "0.424", "0.8936", "0.424", "0.9077", "0.1436", "0.909" ], [ "0.513", "0.8955", "0.8833", "0.8955", "0.8833", "0.91", "0.513", "0.91" ] ]

[ "Plan ahead: Self-Supervised Text Planning", "for Paragraph Completion Task", "Dongyeop Kang*", "Eduard Hovy", "dongyeopk@berkeley.edu", "hovy@cs.cmu.edu", "UC Berkeley", "Carnegie Mellon University", "Abstract", "mainly because per-word predictions from the au-", "toregressive models can not capture the long-term", "Despite the recent success of contextualized", "flow of text, while humans often make structural", "", "language models on various NLP tasks, lan-", "decisions on what and how to say about before they", "guage model itself cannot capture textual co", "speak (Byrne, 1979; McKeown, 1985; Hovy, 1990;", "herence of a long, multi-sentence document", "Swan, 2002; Kang, 2020). Guiding the surface-", "(e.g., a paragraph).Humans often make", "structural decisions on what and how to say", "level realization with such high-level decisions and", "about before making utterances. Guiding sur-", "coherently structuring output text is called a plan-", "facerealization with suchhigh-level decisions", "ning process.", "and structuring text in a coherent way is es-", "Where can the model learn such high-level de-", "sentially called a planning process.Where", "", "cisions related to long-term coherence? A written", "[T']", "can the model learn such high-level coher-", "paragraph itself can be a pot of golden resources,", "ence? A paragraph itself contains various", "forms of inductive coherence signals called", "containing various forms of inductive coherence", "self-supervision in this work, such as sentence", "signals. Different types of coherence signals in", "orders,topicalkeywords,rhetorical structures.", " a paragraph have been studied and used in many", "and so on. Motivated by that, this work pro-", "different ways: a sequence of words or sentences", "", "poses a new paragraph completion task PAR-", "(Devlin et al., 2019; Radford et al., 2019), a dis-", "CoM; predicting masked sentences in a para", "course structure of a text (Appelt, 1982; Hovy,", "graph. However, the task suffers from predict-", "1991; Kang et al., 2017, 2019), an order of sen-", "ing and selecting appropriate topical content", "tences (Chambers and Jurafsky, 2008; Barzilay and", "with respect to the given context. To address", "Lapata, 2008), topic introduction, co-reference, a", "that,wepropose a self-supervised text plan-", "ner ssPlanner that predicts what to say first", "sequence of events (Tomkins, 1978; Schank and", "(content prediction), then guides the pretrained", "Abelson, 2013), and more. In this work, we pri-", "language model (surface realization) using the", "marily focus on the effect of topical content in text", "predicted content. SSPlanner outperforms the", "planning.", "baseline generation models on the paragraph", "Despite the recent advances of contextualized", "completiontaskinbothautomaticandhuman", "language models (Devlin et al., 2019; Radford et al..", "evaluation. We also find that a combination of", "2019), the lack of appropriate tasks makes it diffi-", "nounand verbtypesof keywords isthemost", "effective for content selection. As more num-", "cult to evaluate generation models' long-term co-", "ber of contentkeywords are provided,overall", "herence. Prior tasks fall into classification or rank-", "generation quality also increases.", "ing problems, such as narrative close task (Cham-", "bers and Jurafsky,2008; Mostafazadeh et al., 2016),", "Introduction", "sentence ordering (Barzilay and Lapata, 2008),", "and next sentence prediction (Devlin et al., 2019).", "", "One may think textual coherence can be achieved", "Some recent works focused on designing genera-", "from a gigantic language model trained on massive", "tion tasks: story generation (Fan et al., 2019), text", "data. This might be true in simple cases, such as", "infilling (Huang et al., 2019; Fedus et al., 2018;", "generating short replies (Kannan et al., 2016), but", "Hua and Wang, 2019), or paragraph bridging (Kang", "not in a long, multi-sentence generation. This is", "et al., 2019). However, most of them suffer from", "*This work was done while DK was at CMU.", "predicting appropriate topical content given limited" ]

{ "arxiv_id": "2010.05141", "language": "en", "page_idx": 0 }

[ [ "0.11755", "0.076", "0.4895", "0.076", "0.4895", "0.08905", "0.11755", "0.08905" ], [ "0.513", "0.0766", "0.8833", "0.0766", "0.8833", "0.0903", "0.513", "0.0903" ], [ "0.1184", "0.0932", "0.3208", "0.0932", "0.3208", "0.1069", "0.1184", "0.1069" ], [ "0.513", "0.0932", "0.839", "0.0932", "0.839", "0.1069", "0.513", "0.1069" ], [ "0.1368", "0.10925", "0.4895", "0.10925", "0.4895", "0.1223", "0.1368", "0.1223" ], [ "0.529", "0.1069", "0.885", "0.1081", "0.885", "0.1223", "0.529", "0.1217" ], [ "0.1184", "0.1247", "0.4902", "0.1247", "0.4902", "0.1383", "0.1184", "0.1383" ], [ "0.51", "0.12354", "0.8867", "0.1259", "0.8867", "0.1395", "0.51", "0.1372" ], [ "0.1184", "0.1414", "0.492", "0.1414", "0.492", "0.155", "0.1184", "0.155" ], [ "0.513", "0.1426", "0.8867", "0.1426", "0.8867", "0.1561", "0.513", "0.1561" ], [ "0.1184", "0.1567", "0.4902", "0.1567", "0.4902", "0.1704", "0.1184", "0.1704" ], [ "0.511", "0.1561", "0.885", "0.158", "0.885", "0.1729", "0.511", "0.1704" ], [ "0.1184", "0.1733", "0.4902", "0.1733", "0.4902", "0.187", "0.1184", "0.187" ], [ "0.513", "0.1733", "0.885", "0.1733", "0.885", "0.187", "0.513", "0.187" ], [ "0.1184", "0.1882", "0.492", "0.1882", "0.492", "0.2019", "0.1184", "0.2019" ], [ "0.513", "0.1907", "0.681", "0.1907", "0.681", "0.2037", "0.513", "0.2037" ], [ "0.1184", "0.2048", "0.4895", "0.2048", "0.4895", "0.2197", "0.1184", "0.2197" ], [ "0.529", "0.206", "0.8833", "0.206", "0.8833", "0.2203", "0.529", "0.2203" ], [ "0.12006", "0.2214", "0.4902", "0.2214", "0.4902", "0.2351", "0.12006", "0.2351" ], [ "0.511", "0.2227", "0.8833", "0.2227", "0.8833", "0.2363", "0.511", "0.2363" ], [ "0.1184", "0.2382", "0.4895", "0.2382", "0.4895", "0.2517", "0.1184", "0.2517" ], [ "0.511", "0.2382", "0.8867", "0.2382", "0.8867", "0.2517", "0.511", "0.2517" ], [ "0.12006", "0.2546", "0.1587", "0.2546", "0.1587", "0.2666", "0.12006", "0.2666" ], [ "0.5146", "0.2559", "0.8833", "0.2559", "0.8833", "0.2666", "0.5146", "0.2666" ], [ "0.1377", "0.2695", "0.4895", "0.2695", "0.4895", "0.284", "0.1377", "0.284" ], [ "0.511", "0.2695", "0.8833", "0.2695", "0.8833", "0.2832", "0.511", "0.2832" ], [ "0.1184", "0.2852", "0.4902", "0.2852", "0.4902", "0.2998", "0.1184", "0.2998" ], [ "0.511", "0.2852", "0.8833", "0.2852", "0.8833", "0.2998", "0.511", "0.2998" ], [ "0.11755", "0.3005", "0.4878", "0.3005", "0.4878", "0.3142", "0.11755", "0.3142" ], [ "0.513", "0.3027", "0.6733", "0.3027", "0.6733", "0.3164", "0.513", "0.3164" ], [ "0.12006", "0.3184", "0.4902", "0.3184", "0.4902", "0.332", "0.12006", "0.332" ], [ "0.531", "0.3184", "0.885", "0.3184", "0.885", "0.3333", "0.531", "0.3333" ], [ "0.12006", "0.335", "0.4878", "0.335", "0.4878", "0.3457", "0.12006", "0.3457" ], [ "0.513", "0.335", "0.8833", "0.335", "0.8833", "0.3486", "0.513", "0.3486" ], [ "0.11755", "0.3486", "0.4878", "0.3486", "0.4878", "0.3623", "0.11755", "0.3623" ], [ "0.513", "0.3508", "0.885", "0.3508", "0.885", "0.3652", "0.513", "0.3652" ], [ "0.12006", "0.3665", "0.4895", "0.3665", "0.4895", "0.3801", "0.12006", "0.3801" ], [ "0.513", "0.3677", "0.885", "0.3677", "0.885", "0.3813", "0.513", "0.3813" ], [ "0.12006", "0.3818", "0.4902", "0.3818", "0.4902", "0.3955", "0.12006", "0.3955" ], [ "0.5146", "0.3855", "0.882", "0.3855", "0.882", "0.3955", "0.5146", "0.3955" ], [ "0.12006", "0.3984", "0.4902", "0.3984", "0.4902", "0.4133", "0.12006", "0.4133" ], [ "0.513", "0.3984", "0.8833", "0.3984", "0.8833", "0.412", "0.513", "0.412" ], [ "0.1184", "0.415", "0.4895", "0.415", "0.4895", "0.4287", "0.1184", "0.4287" ], [ "0.513", "0.4163", "0.885", "0.4163", "0.885", "0.43", "0.513", "0.43" ], [ "0.12006", "0.4312", "0.4895", "0.4312", "0.4895", "0.4448", "0.12006", "0.4448" ], [ "0.513", "0.4316", "0.885", "0.4316", "0.885", "0.4453", "0.513", "0.4453" ], [ "0.12006", "0.4478", "0.492", "0.4478", "0.492", "0.4614", "0.12006", "0.4614" ], [ "0.513", "0.449", "0.6514", "0.449", "0.6514", "0.462", "0.513", "0.462" ], [ "0.1184", "0.4631", "0.4902", "0.4631", "0.4902", "0.4768", "0.1184", "0.4768" ], [ "0.1184", "0.4785", "0.4895", "0.4785", "0.4895", "0.4922", "0.1184", "0.4922" ], [ "0.5557", "0.4785", "0.6035", "0.4785", "0.6035", "0.4905", "0.5557", "0.4905" ], [ "0.6406", "0.478", "0.882", "0.478", "0.882", "0.4912", "0.6406", "0.4912" ], [ "0.561", "0.4905", "0.5977", "0.4905", "0.5977", "0.502", "0.561", "0.502" ], [ "0.6484", "0.4905", "0.684", "0.4905", "0.684", "0.502", "0.6484", "0.502" ], [ "0.6963", "0.4893", "0.888", "0.4893", "0.888", "0.503", "0.6963", "0.503" ], [ "0.12006", "0.4954", "0.4902", "0.4954", "0.4902", "0.509", "0.12006", "0.509" ], [ "0.5146", "0.5073", "0.623", "0.5073", "0.623", "0.5186", "0.5146", "0.5186" ], [ "0.663", "0.5083", "0.6753", "0.5083", "0.6753", "0.5186", "0.663", "0.5186" ], [ "0.723", "0.5083", "0.7363", "0.5083", "0.7363", "0.5195", "0.723", "0.5195" ], [ "0.785", "0.511", "0.7983", "0.5073", "0.804", "0.5166", "0.791", "0.521" ], [ "0.8506", "0.511", "0.858", "0.511", "0.858", "0.517", "0.8506", "0.517" ], [ "0.5146", "0.5186", "0.607", "0.5186", "0.607", "0.53", "0.5146", "0.53" ], [ "0.1184", "0.511", "0.4895", "0.511", "0.4895", "0.525", "0.1184", "0.525" ], [ "0.723", "0.521", "0.735", "0.521", "0.735", "0.53", "0.723", "0.53" ], [ "0.8506", "0.5225", "0.858", "0.5225", "0.858", "0.528", "0.8506", "0.528" ], [ "0.665", "0.524", "0.6724", "0.524", "0.6724", "0.5293", "0.665", "0.5293" ], [ "0.1184", "0.527", "0.4902", "0.527", "0.4902", "0.5405", "0.1184", "0.5405" ], [ "0.5146", "0.53", "0.6206", "0.53", "0.6206", "0.5405", "0.5146", "0.5405" ], [ "0.79", "0.532", "0.7983", "0.532", "0.7983", "0.5415", "0.79", "0.5415" ], [ "0.8506", "0.533", "0.8564", "0.533", "0.8564", "0.539", "0.8506", "0.539" ], [ "0.1184", "0.5435", "0.4902", "0.5435", "0.4902", "0.557", "0.1184", "0.557" ], [ "0.513", "0.539", "0.638", "0.539", "0.638", "0.5527", "0.513", "0.5527" ], [ "0.5146", "0.552", "0.614", "0.552", "0.614", "0.5635", "0.5146", "0.5635" ], [ "0.1184", "0.5586", "0.4575", "0.5586", "0.4575", "0.5723", "0.1184", "0.5723" ], [ "0.7886", "0.554", "0.8003", "0.554", "0.8003", "0.5635", "0.7886", "0.5635" ], [ "0.513", "0.5625", "0.6113", "0.564", "0.6113", "0.576", "0.513", "0.5737" ], [ "0.7256", "0.5664", "0.733", "0.5664", "0.733", "0.571", "0.7256", "0.571" ], [ "0.791", "0.5664", "0.7983", "0.5664", "0.7983", "0.572", "0.791", "0.572" ], [ "0.12177", "0.5884", "0.1503", "0.5884", "0.1503", "0.5967", "0.12177", "0.5967" ], [ "0.1436", "0.586", "0.272", "0.586", "0.272", "0.601", "0.1436", "0.601" ], [ "0.513", "0.5903", "0.885", "0.5903", "0.885", "0.604", "0.513", "0.604" ], [ "0.511", "0.604", "0.7886", "0.605", "0.7886", "0.619", "0.511", "0.618" ], [ "0.11755", "0.611", "0.4878", "0.611", "0.4878", "0.626", "0.11755", "0.626" ], [ "0.12006", "0.629", "0.4902", "0.629", "0.4902", "0.6426", "0.12006", "0.6426" ], [ "0.532", "0.637", "0.8833", "0.637", "0.8833", "0.652", "0.532", "0.652" ], [ "0.1184", "0.6445", "0.4902", "0.6445", "0.4902", "0.6577", "0.1184", "0.6577" ], [ "0.513", "0.652", "0.8867", "0.653", "0.8867", "0.6675", "0.513", "0.6665" ], [ "0.11584", "0.66", "0.3628", "0.659", "0.3628", "0.6733", "0.11584", "0.675" ], [ "0.513", "0.6694", "0.885", "0.6694", "0.885", "0.683", "0.513", "0.683" ], [ "0.2737", "0.6895", "0.444", "0.6895", "0.444", "0.703", "0.2737", "0.703" ], [ "0.513", "0.685", "0.8833", "0.686", "0.8833", "0.6997", "0.513", "0.6987" ], [ "0.1184", "0.6953", "0.1453", "0.6953", "0.1453", "0.707", "0.1184", "0.707" ], [ "0.1957", "0.6953", "0.2334", "0.6953", "0.2334", "0.707", "0.1957", "0.707" ], [ "0.4375", "0.6914", "0.4895", "0.6914", "0.4895", "0.7017", "0.4375", "0.7017" ], [ "0.272", "0.6997", "0.4978", "0.6997", "0.4978", "0.7134", "0.272", "0.7134" ], [ "0.511", "0.6997", "0.885", "0.7007", "0.885", "0.7153", "0.511", "0.7144" ], [ "0.1394", "0.7163", "0.2107", "0.7163", "0.2107", "0.728", "0.1394", "0.728" ], [ "0.351", "0.7153", "0.3628", "0.7236", "0.351", "0.733", "0.339", "0.7246" ], [ "0.513", "0.7173", "0.8833", "0.7173", "0.8833", "0.731", "0.513", "0.731" ], [ "0.3938", "0.7236", "0.4055", "0.7173", "0.4114", "0.7227", "0.3997", "0.7285" ], [ "0.126", "0.728", "0.2284", "0.728", "0.2284", "0.7383", "0.126", "0.7383" ], [ "0.351", "0.728", "0.3586", "0.7324", "0.3525", "0.737", "0.345", "0.733" ], [ "0.3938", "0.7285", "0.4055", "0.7256", "0.4106", "0.733", "0.398", "0.7363" ], [ "0.513", "0.734", "0.881", "0.734", "0.881", "0.7446", "0.513", "0.7446" ], [ "0.1276", "0.7383", "0.2124", "0.7383", "0.2124", "0.749", "0.1276", "0.749" ], [ "0.2957", "0.72", "0.3015", "0.72", "0.3015", "0.728", "0.2957", "0.728" ], [ "0.345", "0.7407", "0.357", "0.7407", "0.357", "0.75", "0.345", "0.75" ], [ "0.3997", "0.7407", "0.4072", "0.7407", "0.4072", "0.745", "0.3997", "0.745" ], [ "0.1368", "0.7476", "0.2737", "0.7466", "0.2737", "0.7583", "0.1368", "0.759" ], [ "0.293", "0.753", "0.3047", "0.756", "0.2988", "0.7666", "0.287", "0.763" ], [ "0.3435", "0.755", "0.357", "0.755", "0.357", "0.764", "0.3435", "0.764" ], [ "0.3972", "0.755", "0.4114", "0.755", "0.4114", "0.764", "0.3972", "0.764" ], [ "0.513", "0.75", "0.885", "0.75", "0.885", "0.763", "0.513", "0.763" ], [ "0.1276", "0.756", "0.1411", "0.756", "0.1411", "0.7627", "0.1276", "0.7627" ], [ "0.1377", "0.7573", "0.287", "0.7573", "0.287", "0.771", "0.1377", "0.771" ], [ "0.4592", "0.7573", "0.4685", "0.7573", "0.4685", "0.7627", "0.4592", "0.7627" ], [ "0.513", "0.764", "0.8867", "0.7656", "0.8867", "0.78", "0.513", "0.7793" ], [ "0.11755", "0.7837", "0.4895", "0.7837", "0.4895", "0.7974", "0.11755", "0.7974" ], [ "0.5146", "0.783", "0.6943", "0.783", "0.6943", "0.7935", "0.5146", "0.7935" ], [ "0.11755", "0.799", "0.471", "0.8", "0.471", "0.8135", "0.11755", "0.8125" ], [ "0.529", "0.7974", "0.885", "0.7964", "0.885", "0.811", "0.529", "0.8125" ], [ "0.513", "0.814", "0.8833", "0.814", "0.8833", "0.8276", "0.513", "0.8276" ], [ "0.1377", "0.8306", "0.4937", "0.8306", "0.4937", "0.844", "0.1377", "0.844" ], [ "0.513", "0.8306", "0.8833", "0.8306", "0.8833", "0.844", "0.513", "0.844" ], [ "0.1184", "0.847", "0.4937", "0.847", "0.4937", "0.861", "0.1184", "0.861" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.1184", "0.8633", "0.4937", "0.8633", "0.4937", "0.877", "0.1184", "0.877" ], [ "0.51", "0.8633", "0.8867", "0.8633", "0.8867", "0.877", "0.51", "0.877" ], [ "0.11755", "0.8774", "0.4902", "0.879", "0.4902", "0.8936", "0.11755", "0.8926" ], [ "0.51", "0.879", "0.8867", "0.8774", "0.8867", "0.8926", "0.51", "0.8936" ], [ "0.1184", "0.8955", "0.4902", "0.8955", "0.4902", "0.909", "0.1184", "0.909" ], [ "0.513", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.513", "0.909" ] ]

[ "context,duetothe limitedusageof self-supervision", "is explicitly provided as a data form, the planner", "signals from the paragraph.", "mostly orders and structures, not prediction.", "This work proposes a new open-ended paragraph", "C T: In abstractive summarization, all context", "completion task; PARCoM; predicting the masked", "information is entirely given in the source docu-", "sentences in a paragraph. Unlike the prior works,", "ment, as a superset of target summaries to predict.", "our task uses two effective ways of self-supervision", "Thus, generation only pays attention to abstracting", "learnt from a written paragraph itself: (1) we aug-", "the context into a shorter form instead of content", "ment more training instances via permutation mask-", "prediction or ordering", "ing and (2) resolve the context sparsity problem by", "C ~ T: Paraphrasing is transforming surface", "providing a set of ground-truth content keywords", "patterns of text while preserving its semantics. Fu", "and predicting them directly from context at testing", " et al. (2019) used variational autoencoders for sur-", "time.", "facerealization witha latentbag of words model", "For the task, we propose a self-supervised text", "for differentiable content planning, where content", "planner (SSPlanner) that explicitly predicts con-", "to generate itself is given in context, not requiring", "tent keywords (content prediction) from context", "any content planning.", "and guides the pretrained language model (surface-", "C Il T: Story generation (Fan et al., 2019), text", "realization)using thepredicted content.The distri-", "infilling (Fedus et al., 2018; Huang et al., 2019)", "bution of predicted keywords is then combined with", "paragraph bridging (Kang et al., 2019), and our pro-", "the distribution of words in the language model", "posed PARCoM are very challenging tasks where", "using copy mechanism (See et al., 2017). The pre-", "contextandtargethavenooverlap(open-ended)", "dicted content keywords are an approximation of", "but they should be coherently connected. Table 2", "topical intents by the generator, providing a hint to", "categorize various generation models applied on", "guide the surface realizer to bridge the coherency", "the open-ended tasks (CILT), based on its self-", "gap between the given context and text to gener-", "supervision types:", "ate. Overall, SsPlanner combines two advantages;", "micro-level language fluency from the pre-trained", "Models", "Bidirect.Permutation Content Content", "(H)", "Flow", "Masking Guidance Prediction", "language model (bottom-up) and macro-level con-", "Keskar et al. (2019)", "V", "V", "V", "x", "Fan et al. (2019)", "tent choice controlled by the macro-level planning", "x", "X", "X", "(top-down). Our experiment shows that SsPlanner", "Huang et al. (2019)", "X", ")", "achieves significant improvements over the base-", "Hua and Wang (2019)", "Kang et al. (2019)", "lines in both automatic and human evaluation.", "x", "SSPlanner (ours)", "", "", "2", "Related Work", "Table 2:Comparison of generation models in ClLT", "tasks by different self-supervision types.", "We first categorize a wide range of long-term co-", "herent generation tasks (Table 1), based on their", "Keskar et al. (2019) conditioned language mod-", "inclusion relationship (C-T) between the given con-", "els with topical words to control the target text.", "text (C) and target to predict (T).", "Fan et al. (2019) developed a surface realizer on", "Content Content Content", "anonymized entities using semantic role labeling", "C-T", "Tasks", "Surface", "Selection Planning Ordering Realization", "Hua and Wang (2019) used pre-extracted topics", "Data-to-text", "+", "to guide a generator to produce stylized argumen-", "", "SSummarization", "+", "A", "tation text However. they are given the topical", "Paraphrasing", "X", "x", "X", "StoryGen,Text Infilling.", "J", "", "", "content as input (content guidance), while our", "L.", "Bridging, PARCOM (ours)", "", "SsPlanner directly predicts plan words from con-", "Table 1: Comparison of generation tasks by different", "text (contentprediction)", "inclusion relationships between Context and Target.", "Fedus et al. (2018): Huang et al. (2019) devel-", "oped various methods for text infilling task. Very", "C C T: Data-to-text produces text from struc-", "similar to our task, Kang et al. (2019) developed", "tured data (e.g., table). Moryossef et al. (2019);", "language models informed by discourse relations", "Puduppully et al. (2019); Shen et al. (2019); Mi-", " on the bridging task; given the first and last sen-", "culicich et al. (2019) combine content planning", "tences, predicting the intermediate sentences (bidi-", "with surface realization. However, since content", "rectional flow). However, they did not explicitly" ]

{ "arxiv_id": "2010.05141", "language": "en", "page_idx": 1 }

[ [ "0.1335", "0.0748", "0.4902", "0.076", "0.4902", "0.0903", "0.1335", "0.08905" ], [ "0.513", "0.076", "0.882", "0.076", "0.882", "0.0903", "0.513", "0.0903" ], [ "0.1335", "0.09265", "0.492", "0.09143", "0.492", "0.1057", "0.1335", "0.1069" ], [ "0.51", "0.09143", "0.881", "0.09143", "0.881", "0.1051", "0.51", "0.1051" ], [ "0.1335", "0.1081", "0.4937", "0.1081", "0.4937", "0.1217", "0.1335", "0.1217" ], [ "0.513", "0.10925", "0.8833", "0.10925", "0.8833", "0.1223", "0.513", "0.1223" ], [ "0.1335", "0.1247", "0.4937", "0.1247", "0.4937", "0.1395", "0.1335", "0.1395" ], [ "0.513", "0.1247", "0.6543", "0.1247", "0.6543", "0.1383", "0.513", "0.1383" ], [ "0.1351", "0.1426", "0.315", "0.1426", "0.315", "0.1526", "0.1351", "0.1526" ], [ "0.531", "0.1437", "0.8833", "0.1437", "0.8833", "0.158", "0.531", "0.158" ], [ "0.1184", "0.1567", "0.4902", "0.1567", "0.4902", "0.1704", "0.1184", "0.1704" ], [ "0.511", "0.1592", "0.881", "0.158", "0.881", "0.1729", "0.511", "0.1733" ], [ "0.1335", "0.1729", "0.4902", "0.1729", "0.4902", "0.1859", "0.1335", "0.1859" ], [ "0.511", "0.1758", "0.8833", "0.1746", "0.8833", "0.1895", "0.511", "0.1907" ], [ "0.1318", "0.187", "0.4902", "0.1882", "0.4902", "0.2031", "0.1318", "0.2019" ], [ "0.513", "0.1936", "0.8833", "0.1936", "0.8833", "0.2079", "0.513", "0.2079" ], [ "0.1335", "0.2048", "0.4902", "0.2048", "0.4902", "0.2185", "0.1335", "0.2185" ], [ "0.513", "0.209", "0.885", "0.209", "0.885", "0.2239", "0.513", "0.2239" ], [ "0.1301", "0.2203", "0.4895", "0.2197", "0.4895", "0.234", "0.1301", "0.2351" ], [ "0.513", "0.2245", "0.8867", "0.2245", "0.8867", "0.2393", "0.513", "0.2393" ], [ "0.1301", "0.2363", "0.4902", "0.2369", "0.4902", "0.2517", "0.1301", "0.2505" ], [ "0.513", "0.2405", "0.8833", "0.2393", "0.8833", "0.2542", "0.513", "0.2546" ], [ "0.1301", "0.253", "0.4878", "0.253", "0.4878", "0.2673", "0.1301", "0.2673" ], [ "0.511", "0.2559", "0.885", "0.2546", "0.885", "0.2695", "0.511", "0.2708" ], [ "0.1318", "0.2695", "0.4895", "0.2695", "0.4895", "0.284", "0.1318", "0.284" ], [ "0.513", "0.2715", "0.885", "0.2715", "0.885", "0.2861", "0.513", "0.2861" ], [ "0.1335", "0.2861", "0.4575", "0.2861", "0.4575", "0.2998", "0.1335", "0.2998" ], [ "0.513", "0.288", "0.8867", "0.288", "0.8867", "0.3027", "0.513", "0.3027" ], [ "0.1351", "0.3005", "0.4895", "0.3018", "0.4895", "0.3164", "0.1351", "0.3154" ], [ "0.511", "0.3027", "0.848", "0.3027", "0.848", "0.3176", "0.511", "0.3176" ], [ "0.12006", "0.3184", "0.4895", "0.3184", "0.4895", "0.332", "0.12006", "0.332" ], [ "0.531", "0.3237", "0.8867", "0.3237", "0.8867", "0.3374", "0.531", "0.3374" ], [ "0.1184", "0.3333", "0.486", "0.3333", "0.486", "0.3467", "0.1184", "0.3467" ], [ "0.5146", "0.3403", "0.8867", "0.3403", "0.8867", "0.354", "0.5146", "0.354" ], [ "0.12177", "0.3516", "0.4878", "0.3516", "0.4878", "0.3623", "0.12177", "0.3623" ], [ "0.513", "0.3557", "0.885", "0.3557", "0.885", "0.3694", "0.513", "0.3694" ], [ "0.12006", "0.3677", "0.486", "0.3677", "0.486", "0.3777", "0.12006", "0.3777" ], [ "0.513", "0.3718", "0.8867", "0.3718", "0.8867", "0.3855", "0.513", "0.3855" ], [ "0.12006", "0.383", "0.4895", "0.383", "0.4895", "0.3967", "0.12006", "0.3967" ], [ "0.513", "0.3884", "0.8867", "0.3884", "0.8867", "0.402", "0.513", "0.402" ], [ "0.12006", "0.3997", "0.4895", "0.3997", "0.4895", "0.4133", "0.12006", "0.4133" ], [ "0.513", "0.4038", "0.8867", "0.4026", "0.8867", "0.4175", "0.513", "0.4187" ], [ "0.11755", "0.415", "0.4878", "0.415", "0.4878", "0.4287", "0.11755", "0.4287" ], [ "0.513", "0.4204", "0.885", "0.4192", "0.885", "0.434", "0.513", "0.4353" ], [ "0.1184", "0.4312", "0.4895", "0.4312", "0.4895", "0.4448", "0.1184", "0.4448" ], [ "0.513", "0.4358", "0.8833", "0.4358", "0.8833", "0.4507", "0.513", "0.4507" ], [ "0.12006", "0.4478", "0.4895", "0.4478", "0.4895", "0.4614", "0.12006", "0.4614" ], [ "0.511", "0.452", "0.8833", "0.4507", "0.8833", "0.4656", "0.511", "0.466" ], [ "0.1184", "0.4631", "0.4895", "0.4631", "0.4895", "0.4768", "0.1184", "0.4768" ], [ "0.511", "0.4673", "0.8833", "0.466", "0.8833", "0.484", "0.511", "0.485" ], [ "0.1184", "0.4797", "0.4902", "0.4797", "0.4902", "0.4934", "0.1184", "0.4934" ], [ "0.513", "0.484", "0.882", "0.485", "0.882", "0.4995", "0.513", "0.4988" ], [ "0.12006", "0.4963", "0.2225", "0.4963", "0.2225", "0.5083", "0.12006", "0.5083" ], [ "0.513", "0.5005", "0.8833", "0.502", "0.8833", "0.5166", "0.513", "0.5156" ], [ "0.11584", "0.5215", "0.409", "0.524", "0.409", "0.5415", "0.11584", "0.539" ], [ "0.5146", "0.5186", "0.8833", "0.5186", "0.8833", "0.532", "0.5146", "0.532" ], [ "0.5146", "0.536", "0.5977", "0.536", "0.5977", "0.5474", "0.5146", "0.5474" ], [ "0.1503", "0.5435", "0.2568", "0.5435", "0.2568", "0.5547", "0.1503", "0.5547" ], [ "0.532", "0.554", "0.8833", "0.554", "0.8833", "0.568", "0.532", "0.568" ], [ "0.1184", "0.568", "0.4895", "0.5684", "0.4895", "0.583", "0.1184", "0.582" ], [ "0.5146", "0.571", "0.8867", "0.571", "0.8867", "0.5845", "0.5146", "0.5845" ], [ "0.1184", "0.5845", "0.4895", "0.5845", "0.4895", "0.597", "0.1184", "0.597" ], [ "0.5146", "0.586", "0.885", "0.586", "0.885", "0.5996", "0.5146", "0.5996" ], [ "0.12006", "0.601", "0.4895", "0.601", "0.4895", "0.6143", "0.12006", "0.6143" ], [ "0.513", "0.6025", "0.8833", "0.6016", "0.8833", "0.615", "0.513", "0.616" ], [ "0.12006", "0.618", "0.4902", "0.618", "0.4902", "0.632", "0.12006", "0.632" ], [ "0.513", "0.6177", "0.8833", "0.618", "0.8833", "0.632", "0.513", "0.6313" ], [ "0.12006", "0.6343", "0.4902", "0.6343", "0.4902", "0.648", "0.12006", "0.648" ], [ "0.5146", "0.6343", "0.882", "0.6343", "0.882", "0.6484", "0.5146", "0.6484" ], [ "0.12006", "0.651", "0.4902", "0.651", "0.4902", "0.6646", "0.12006", "0.6646" ], [ "0.513", "0.6484", "0.881", "0.6484", "0.881", "0.662", "0.513", "0.662" ], [ "0.12006", "0.665", "0.4902", "0.665", "0.4902", "0.6787", "0.12006", "0.6787" ], [ "0.513", "0.6665", "0.882", "0.6665", "0.882", "0.6797", "0.513", "0.6797" ], [ "0.12006", "0.683", "0.4895", "0.683", "0.4895", "0.697", "0.12006", "0.697" ], [ "0.5156", "0.684", "0.882", "0.684", "0.882", "0.695", "0.5156", "0.695" ], [ "0.11755", "0.6978", "0.4895", "0.6978", "0.4895", "0.7114", "0.11755", "0.7114" ], [ "0.513", "0.6987", "0.8833", "0.6987", "0.8833", "0.712", "0.513", "0.712" ], [ "0.1184", "0.7153", "0.492", "0.7153", "0.492", "0.7285", "0.1184", "0.7285" ], [ "0.513", "0.7153", "0.8833", "0.7153", "0.8833", "0.7285", "0.513", "0.7285" ], [ "0.1184", "0.731", "0.492", "0.731", "0.492", "0.7446", "0.1184", "0.7446" ], [ "0.511", "0.731", "0.8833", "0.73", "0.8833", "0.7446", "0.511", "0.745" ], [ "0.11755", "0.745", "0.4895", "0.7466", "0.4895", "0.761", "0.11755", "0.7603" ], [ "0.513", "0.7476", "0.6265", "0.7476", "0.6265", "0.7627", "0.513", "0.7627" ], [ "0.12006", "0.764", "0.2284", "0.764", "0.2284", "0.7754", "0.12006", "0.7754" ], [ "0.531", "0.7656", "0.8833", "0.7656", "0.8833", "0.78", "0.531", "0.78" ], [ "0.1351", "0.7793", "0.4902", "0.78", "0.4902", "0.7944", "0.1351", "0.7935" ], [ "0.511", "0.782", "0.8833", "0.781", "0.8833", "0.796", "0.511", "0.7964" ], [ "0.11755", "0.7944", "0.4895", "0.7944", "0.4895", "0.808", "0.11755", "0.808" ], [ "0.511", "0.7964", "0.8833", "0.7974", "0.8833", "0.8125", "0.511", "0.811" ], [ "0.1184", "0.81", "0.4937", "0.811", "0.4937", "0.826", "0.1184", "0.8247" ], [ "0.511", "0.814", "0.8833", "0.814", "0.8833", "0.8276", "0.511", "0.8276" ], [ "0.12006", "0.8276", "0.492", "0.8276", "0.492", "0.8413", "0.12006", "0.8413" ], [ "0.511", "0.8306", "0.882", "0.8306", "0.882", "0.844", "0.511", "0.844" ], [ "0.12006", "0.844", "0.4902", "0.844", "0.4902", "0.858", "0.12006", "0.858" ], [ "0.513", "0.847", "0.8867", "0.847", "0.8867", "0.861", "0.513", "0.861" ], [ "0.513", "0.8633", "0.885", "0.8633", "0.885", "0.877", "0.513", "0.877" ], [ "0.1394", "0.8706", "0.2822", "0.8706", "0.2822", "0.8843", "0.1394", "0.8843" ], [ "0.1377", "0.882", "0.492", "0.883", "0.492", "0.898", "0.1377", "0.8965" ], [ "0.511", "0.8774", "0.8867", "0.879", "0.8867", "0.8936", "0.511", "0.8926" ], [ "0.1184", "0.898", "0.361", "0.898", "0.361", "0.9077", "0.1184", "0.9077" ], [ "0.511", "0.8936", "0.888", "0.891", "0.888", "0.909", "0.511", "0.911" ] ]

[ "(Tomkins, 1978)) are often regarded as the most", "predictor in SSPlanner predicts the plan keywords", "salient topical content in generation. Using off-", " only from the given context, by learning the topica", "the-shelf Part-of-speech (PoS) tagger?, we ex-", "relationship between context and content in target", "tract three types of syntactic features: nouns,", "from training data.", "verbs,andnouns+verbs.", "Given I length of a paragraph si..Si where each", ": Attention weights are used to capture context-", "sentence s consists of a n number of words s =", "aware keywords. We use the pre-trained BERT", "Wi..Wn, PARCoM splits it into the context sen", "(Devlin et al., 2019) to encode context and target", "tences x=si..Sj-1, Sj+r..Sn and t target sentences to", "text, then average the attention weights of context", "predict y=sj..Sj+t-1. For each target sentence, p", "words with respect to each target word. We only", "number of plan keywords kj,1..kj,p for arbitrary tar-", "use the first head's attentions, then average them", "get sentence s, are given only at training time. The", "over all 12 layers3. We finally choose words", "plan keywords are chosen from the entire vocabu-", "with the maximum weight except for the special", "lary ww and later combined with word distribution", "tokens (e.g., [CLS]) and punctuation marks.", "from the language model. We describe each self-", "We set the maximum number of keywords per", "supervision module in SSPlanner as follows:", "sentence (nkps) to 5. Some extractors output an", "Surface realization with pre-trained lan-", "empty keyword list, so the number of keywords", "guage models. We use two different types of", "acrossthesystemsisdifferent.Ourkeywordsare", "transformer-based language models: BERT (De-", "alwavs uni-grams. In case they arenotuni-grams", "vlin et al., 2019) and GPT2 (Radford et al., 2019).", "we split them by whitespaces and use individual", "While GPT2 is trained on bidirectionally tied lan-", "unigrams as unique keywords. If the target text", "guage modeling, BERT is trained on masked lan-", "has multiple sentences, we combine all keywords", "guage modeling. For BERT, we use the sequential", "from the sentences and randomly shuffle them. The", "sampling method (Wang and Cho, 2019). Using", "plan keywords extracted are only provided while", "them, we encode context x and output the hidden", "training our plan predictor, but not at test time. At", "representation hj,i = f(hj-1,i,xk<(j.) for jth word", "testing time, we explicitly predict the keywords", "in ith sentence, where f E [BERT, GPT2} is the", "given context.", "transformer language model. We then output the", "4 Self-supervised Text Planning", "sentence vector h; by averaging all word vectors in", "a sentence.", "(SSPlanner)", "Sentence position embedding. We concatenate", "SSPlanner has various self-supervision modules", "the encoded sentence representation with its sen-", "that learn coherence signals from a paragraph it-", "tence position embedding. By adding the sentence", "self: surface realizer (language model) by learn-", "position embeddings into context encoding, the", "ing from a sequence of words, next sentence pre-", "model is aware of where the context sentence came", "dictor by learning from a sequence of two consec-", "from (e.g., from the first or last). Compared to the", "utive sentences, sentence position embeddings", "simple concatenation of them (Kang et al.. 2019)", "by learning from an order of context sentences.", "our sentence position embedding helps better learn", "plan predictor by learning from the relationship", "thebi-directional coherence.The context vector's", "between the given context and important keywords", "final representation is then hc=,h:poswhere", "used in the generation of the target text, and con-", "n is the number of sentences in a text and pos, is", "tent guidance by learning from whether the pre-", "the position embedding of ith sentence in the con-", "dicted plan keywords are used or not in the target", "text paragraph.", "(See Figure 2).", "Plan prediction. This work assumes that high-", "Our planner is motivated by the two-stage gener-", "level plan words consist of bag-of-words (Fu et al.", "ation framework (Moryossef et al., 2019; Miculi-", "2019), so that the model directly predicts the plan", "cich et al., 2019; Fu et al., 2019; Hua and Wang,", "keywords from the vocabulary used in surface re", "2019). While in prior works, the content is explic-", "alization. We calculate the plan probabilities over", "itly given from the dataset or task itself, our plan", "the entire vocabularies V given the context vec-", "tor hc and choose the p number of keywords with", "2https://spacy.io/", "3Vig (2019) observed that which layers or heads are im-", "maximum probability estimates over vocabulary:", "portantfor syntactic and semantic tasks", "pkev = softmax(hcwcvV) where V is the vocabu-" ]

{ "arxiv_id": "2010.05141", "language": "en", "page_idx": 3 }

[ [ "0.0801", "0.09216", "0.468", "0.09216", "0.468", "0.1061", "0.0801", "0.1061" ], [ "0.5024", "0.09216", "0.892", "0.09216", "0.892", "0.1061", "0.5024", "0.1061" ], [ "0.0801", "0.108", "0.4697", "0.108", "0.4697", "0.1218", "0.0801", "0.1218" ], [ "0.501", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.501", "0.1218" ], [ "0.0801", "0.1225", "0.4697", "0.1225", "0.4697", "0.1364", "0.0801", "0.1364" ], [ "0.5024", "0.1225", "0.8936", "0.1225", "0.8936", "0.1364", "0.5024", "0.1364" ], [ "0.0801", "0.1383", "0.4697", "0.1383", "0.4697", "0.1515", "0.0801", "0.1515" ], [ "0.5024", "0.1383", "0.892", "0.1383", "0.892", "0.1515", "0.5024", "0.1515" ], [ "0.0784", "0.1515", "0.468", "0.1528", "0.468", "0.1674", "0.0784", "0.1666" ], [ "0.5024", "0.1528", "0.8555", "0.1515", "0.8555", "0.1666", "0.5024", "0.1674" ], [ "0.0784", "0.1666", "0.468", "0.1674", "0.468", "0.1818", "0.0784", "0.1812" ], [ "0.5215", "0.1686", "0.841", "0.1686", "0.841", "0.1818", "0.5215", "0.1818" ], [ "0.0801", "0.1831", "0.468", "0.1831", "0.468", "0.1963", "0.0801", "0.1963" ], [ "0.5215", "0.1812", "0.892", "0.1818", "0.892", "0.1963", "0.5215", "0.1957" ], [ "0.0809", "0.1995", "0.468", "0.1995", "0.468", "0.2102", "0.0809", "0.2102" ], [ "0.536", "0.1995", "0.8906", "0.1995", "0.8906", "0.2102", "0.536", "0.2102" ], [ "0.0801", "0.2134", "0.4714", "0.2134", "0.4714", "0.2267", "0.0801", "0.2267" ], [ "0.536", "0.214", "0.892", "0.214", "0.892", "0.2247", "0.536", "0.2247" ], [ "0.0801", "0.2279", "0.4697", "0.2279", "0.4697", "0.2412", "0.0801", "0.2412" ], [ "0.536", "0.2299", "0.8906", "0.2299", "0.8906", "0.2405", "0.536", "0.2405" ], [ "0.0801", "0.2437", "0.4697", "0.2437", "0.4697", "0.2568", "0.0801", "0.2568" ], [ "0.535", "0.2437", "0.8228", "0.2437", "0.8228", "0.2568", "0.535", "0.2568" ], [ "0.0801", "0.2588", "0.357", "0.2588", "0.357", "0.2727", "0.0801", "0.2727" ], [ "0.5215", "0.2583", "0.892", "0.2583", "0.892", "0.2715", "0.5215", "0.2715" ], [ "0.0972", "0.276", "0.4673", "0.276", "0.4673", "0.289", "0.0972", "0.289" ], [ "0.5337", "0.2734", "0.8936", "0.2734", "0.8936", "0.2874", "0.5337", "0.2874" ], [ "0.0801", "0.2917", "0.468", "0.2917", "0.468", "0.305", "0.0801", "0.305" ], [ "0.535", "0.289", "0.8936", "0.289", "0.8936", "0.303", "0.535", "0.303" ], [ "0.0784", "0.307", "0.468", "0.3062", "0.468", "0.3208", "0.0784", "0.3213" ], [ "0.535", "0.3037", "0.892", "0.3037", "0.892", "0.3176", "0.535", "0.3176" ], [ "0.0801", "0.3213", "0.4722", "0.3213", "0.4722", "0.3352", "0.0801", "0.3352" ], [ "0.535", "0.318", "0.892", "0.318", "0.892", "0.332", "0.535", "0.332" ], [ "0.5337", "0.332", "0.892", "0.3333", "0.892", "0.348", "0.5337", "0.3467" ], [ "0.0809", "0.3384", "0.468", "0.3384", "0.468", "0.3484", "0.0809", "0.3484" ], [ "0.0801", "0.3516", "0.468", "0.3516", "0.468", "0.3655", "0.0801", "0.3655" ], [ "0.5215", "0.3484", "0.892", "0.3484", "0.892", "0.3623", "0.5215", "0.3623" ], [ "0.0809", "0.3687", "0.4673", "0.3687", "0.4673", "0.379", "0.0809", "0.379" ], [ "0.535", "0.3643", "0.8936", "0.3643", "0.8936", "0.3774", "0.535", "0.3774" ], [ "0.0809", "0.3838", "0.468", "0.3838", "0.468", "0.3945", "0.0809", "0.3945" ], [ "0.5376", "0.3801", "0.892", "0.3801", "0.892", "0.3901", "0.5376", "0.3901" ], [ "0.0801", "0.3977", "0.4697", "0.3977", "0.4697", "0.411", "0.0801", "0.411" ], [ "0.535", "0.3945", "0.892", "0.3945", "0.892", "0.4077", "0.535", "0.4077" ], [ "0.0801", "0.4136", "0.4697", "0.4136", "0.4697", "0.4268", "0.0801", "0.4268" ], [ "0.5337", "0.4104", "0.721", "0.4104", "0.721", "0.4236", "0.5337", "0.4236" ], [ "0.0801", "0.428", "0.468", "0.428", "0.468", "0.4414", "0.0801", "0.4414" ], [ "0.501", "0.4343", "0.6284", "0.4363", "0.6284", "0.4539", "0.501", "0.452" ], [ "0.0801", "0.443", "0.4697", "0.443", "0.4697", "0.457", "0.0801", "0.457" ], [ "0.0801", "0.4583", "0.468", "0.4583", "0.468", "0.4717", "0.0801", "0.4717" ], [ "0.5034", "0.461", "0.759", "0.461", "0.759", "0.476", "0.5034", "0.476" ], [ "0.0801", "0.4734", "0.468", "0.4734", "0.468", "0.4873", "0.0801", "0.4873" ], [ "0.0801", "0.488", "0.4697", "0.488", "0.4697", "0.502", "0.0801", "0.502" ], [ "0.501", "0.4873", "0.8936", "0.488", "0.8936", "0.5024", "0.501", "0.502" ], [ "0.0801", "0.5024", "0.4714", "0.5024", "0.4714", "0.5166", "0.0801", "0.5166" ], [ "0.501", "0.504", "0.892", "0.5024", "0.892", "0.5176", "0.501", "0.518" ], [ "0.0784", "0.518", "0.4697", "0.518", "0.4697", "0.532", "0.0784", "0.532" ], [ "0.5034", "0.5195", "0.892", "0.5195", "0.892", "0.5327", "0.5034", "0.5327" ], [ "0.0801", "0.535", "0.1544", "0.535", "0.1544", "0.547", "0.0801", "0.547" ], [ "0.5024", "0.535", "0.892", "0.535", "0.892", "0.549", "0.5024", "0.549" ], [ "0.0972", "0.55", "0.468", "0.5503", "0.468", "0.565", "0.0972", "0.5645" ], [ "0.5024", "0.55", "0.892", "0.55", "0.892", "0.563", "0.5024", "0.563" ], [ "0.0784", "0.565", "0.468", "0.5664", "0.468", "0.5806", "0.0784", "0.58" ], [ "0.5034", "0.565", "0.584", "0.565", "0.584", "0.577", "0.5034", "0.577" ], [ "0.0801", "0.582", "0.4697", "0.582", "0.4697", "0.595", "0.0801", "0.595" ], [ "0.6445", "0.586", "0.6846", "0.586", "0.6846", "0.603", "0.6445", "0.603" ], [ "0.6724", "0.5923", "0.8057", "0.5947", "0.8057", "0.6123", "0.6724", "0.61" ], [ "0.8677", "0.5947", "0.8945", "0.5947", "0.8945", "0.6113", "0.8677", "0.6113" ], [ "0.5923", "0.595", "0.63", "0.595", "0.63", "0.6094", "0.5923", "0.6094" ], [ "0.0801", "0.595", "0.468", "0.595", "0.468", "0.61", "0.0801", "0.61" ], [ "0.0801", "0.6123", "0.4697", "0.6123", "0.4697", "0.6255", "0.0801", "0.6255" ], [ "0.6543", "0.608", "0.669", "0.608", "0.669", "0.617", "0.6543", "0.617" ], [ "0.0801", "0.627", "0.4697", "0.627", "0.4697", "0.64", "0.0801", "0.64" ], [ "0.5024", "0.6255", "0.892", "0.6255", "0.892", "0.64", "0.5024", "0.64" ], [ "0.0801", "0.6426", "0.4697", "0.6426", "0.4697", "0.656", "0.0801", "0.656" ], [ "0.5024", "0.6416", "0.892", "0.6416", "0.892", "0.655", "0.5024", "0.655" ], [ "0.0801", "0.657", "0.468", "0.657", "0.468", "0.6704", "0.0801", "0.6704" ], [ "0.5195", "0.655", "0.892", "0.656", "0.892", "0.6704", "0.5195", "0.6694" ], [ "0.0801", "0.6724", "0.468", "0.6724", "0.468", "0.686", "0.0801", "0.686" ], [ "0.501", "0.6694", "0.642", "0.672", "0.642", "0.686", "0.501", "0.6846" ], [ "0.0784", "0.6875", "0.468", "0.6875", "0.468", "0.7007", "0.0784", "0.7007" ], [ "0.5156", "0.6978", "0.5474", "0.6997", "0.546", "0.711", "0.5146", "0.709" ], [ "0.569", "0.6934", "0.8936", "0.696", "0.8936", "0.7163", "0.569", "0.714" ], [ "0.0784", "0.702", "0.468", "0.7007", "0.468", "0.7153", "0.0784", "0.7163" ], [ "0.0801", "0.7173", "0.468", "0.7173", "0.468", "0.731", "0.0801", "0.731" ], [ "0.5034", "0.726", "0.8906", "0.726", "0.8906", "0.7407", "0.5034", "0.7407" ], [ "0.0801", "0.733", "0.282", "0.733", "0.282", "0.747", "0.0801", "0.747" ], [ "0.5024", "0.742", "0.892", "0.742", "0.892", "0.756", "0.5024", "0.756" ], [ "0.0989", "0.7505", "0.4697", "0.7505", "0.4697", "0.765", "0.0989", "0.765" ], [ "0.501", "0.756", "0.8906", "0.756", "0.8906", "0.7695", "0.501", "0.7695" ], [ "0.0801", "0.7646", "0.4697", "0.765", "0.4697", "0.78", "0.0801", "0.779" ], [ "0.5024", "0.773", "0.892", "0.773", "0.892", "0.786", "0.5024", "0.786" ], [ "0.0801", "0.781", "0.4697", "0.781", "0.4697", "0.7944", "0.0801", "0.7944" ], [ "0.5024", "0.787", "0.8447", "0.787", "0.8447", "0.8003", "0.5024", "0.8003" ], [ "0.0801", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0801", "0.81" ], [ "0.0801", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.0801", "0.8247" ], [ "0.5156", "0.811", "0.892", "0.811", "0.892", "0.829", "0.5156", "0.829" ], [ "0.0801", "0.827", "0.4697", "0.827", "0.4697", "0.8403", "0.0801", "0.8403" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.5024", "0.837", "0.8896", "0.837", "0.8896", "0.8516", "0.5024", "0.8516" ], [ "0.0801", "0.857", "0.4714", "0.857", "0.4714", "0.8706", "0.0801", "0.8706" ], [ "0.5024", "0.854", "0.697", "0.854", "0.697", "0.867", "0.5024", "0.867" ], [ "0.0801", "0.872", "0.4722", "0.872", "0.4722", "0.886", "0.0801", "0.886" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0801", "0.886", "0.4714", "0.886", "0.4714", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ] ]

[ "Visible image Fusion (IVF)and Medical Image Fusion (MIF)", "details that are difficult to control by manually designed", "are particularly challenging in Multi-Modality Image Fusion", "loss functions, resulting in visually perceptible images. As", "(MMIF) since they focus on modeling cross-modality fea-", "a generative method, DDFM achieves stable and control-", "tures and preserving critical information from all sensors and", "lable generation of fused images without discriminator, by", "modalities. Specifically, in IVF, fused images aim to retain", "applying likelihood rectification to the DDPM output.", "both thermal radiation from infrared images and detailed", "Our contributions are organized in three aspects:", "texture information from visible images, thereby avoiding", ": We introduce a DDPM-based posterior sampling model", "the limitations of visible imagesbeing sensitiveto illumi-", "forMMIF,consisting of anunconditional generation", "nation conditions and infrared images being noisy and low-", "moduleandaconditionallikelihoodrectificationmod-", "resolution. While MIF can assist in diagnosis and treatment", "ule.The sampling offused images is achieved solely", "by fusing multiple medical imaging modalities for precise", "by a pre-trained DDPM without fine-tuning", "detection of abnormality locations [16, 9].", ": In likelihood rectification, since obtaining the likeli-", " There have been numerous methods devised recently to", "hood explicitly is not feasible, we formulate the op-", "address the challenges posed by MMIF [26, 65, 29], and", "timization loss as a probability inference problem in-", "generative models [7, 38] have been extensively utilized to", "volving latent variables, which can be solved by the", "model the distribution of fused images and achieve satis-", "EM algorithm. Then the solution is integrated into the", "DDPM loop to complete conditional image generation.", "factoryfusioneffects.Amongthem,modelsbasedonGen", "erative Adversarial Networks (GANs) [34, 35, 33, 26] are", ": Extensive evaluation of IVF and MIF tasks shows that", "dominant.The workflow ofGAN-based models.illustrated", "DDFM consistently delivers favorable fusion results,", "in Fig.la,involves a generator that creates images contain", "effectivelypreservingboththestructureanddetailin-", "ing information from source images, and a discriminator that", "formation from the source images, while also satisfying", "determines whether the generated images are in a similar", "visual fidelity requirements.", "manifold to the source images. Although GAN-based meth.", "2.Background", "ods have the ability to generate high-quality fused images,", "they suffer from unstable training, lack of interpretability", "2.1. Score-based diffusion models", "and mode collapse, which seriously affect the quality of the", "generated samples. Moreover, as a black-box model, it is", "Score SDE formulation. Diffusion models aim to gen-", "difficult to comprehend the internal mechanisms and behav-", "erate samples by reversing a predefined forward process", "iors of GANs, making it challenging to achieve controllable", "that converts a clean sample co to almost Gaussian sig-", "generation.", "nal T by gradually adding noise. This forward process", "Recently, Denoising Diffusion Probabilistic Models", "can be described by an Ito Stochastic Differential Equation", "(DDPM) [13] has garnered attention in the machine learning", "(SDE) [49]:", "community, which generates high-quality images by mod-", "β(t),", "atdt+B(t)dw,", "(1)", "da =", "eling the diffusion process of restoring a noise-corrupted", "image towards a clean image. Based on the Langevin dif-", "2", "fusion process, DDPM utilizes a series of reverse diffusion", "where dw is standard Wiener process and β(t) is predefined", "steps to generate promising synthetic samples [46]. Com-", "noise schedule that favors the variance-preserving SDE [49]", "pared to GAN, DDPM does not require the discriminator", "This forward process can be reversed in time and still in", "network, thus mitigating common issues such as unstable", "the form of SDE [1]:", "training and mode collapse in GAN. Moreover, its generation", "d=", "Bg)at - β(t)V. log p(at) dt + VB(t)d, (2)", "process is interpretable, as it is based on denoising diffusion", "to generate images, enabling a better understanding of the", "where dw corresponds to the standard Wiener process run-", "image generation process [57]", "ning backward and the only unknown part Va, log pt(at)", "Therefore, we propose a Denoising Diffusion image", "can be modeled as the so-called score function se(t, t) with", "Fusion Model (DDFM), as shown in Fig. Ic. We formulate", "denoising score matching methods, and this score function", "the conditional generation task as a DDPM-based posterior", "can be trained with the following objective [15, 48]:", "sampling model, which can be further decomposed into an", "unconditional generation diffusion problem and a maximum", "EEaEao[lse(at,t)-Va, logPot(atco)l2],(3)", "likelihood estimation problem. The former satisfies natu-", "ral image prior while the latter is inferred to restrict the", "where t is uniformly sampled over [o, T and the data pair", "similarity with source images via likelihood rectification.", "(αo,t) ~po(a)pot(atαo).", "Compared to discriminative approaches, modeling the nat-", "Sampling with diffusion models. Specifically, an uncon-", "ural image prior with DDPM enables better generation of", "ditional diffusion generation process starts with a random" ]

{ "arxiv_id": "2303.06840", "language": "en", "page_idx": 1 }

[ [ "0.0801", "0.09155", "0.468", "0.09155", "0.468", "0.1061", "0.0801", "0.1061" ], [ "0.5024", "0.09155", "0.892", "0.09155", "0.892", "0.1061", "0.5024", "0.1061" ], [ "0.0825", "0.1092", "0.4673", "0.1092", "0.4673", "0.1193", "0.0825", "0.1193" ], [ "0.5034", "0.1092", "0.8906", "0.1092", "0.8906", "0.1193", "0.5034", "0.1193" ], [ "0.0784", "0.1225", "0.4697", "0.1218", "0.4697", "0.1364", "0.0784", "0.137" ], [ "0.5034", "0.125", "0.892", "0.125", "0.892", "0.1351", "0.5034", "0.1351" ], [ "0.0809", "0.1395", "0.468", "0.1395", "0.468", "0.1497", "0.0809", "0.1497" ], [ "0.5024", "0.1383", "0.892", "0.1383", "0.892", "0.1515", "0.5024", "0.1515" ], [ "0.0801", "0.1528", "0.452", "0.1528", "0.452", "0.1674", "0.0801", "0.1674" ], [ "0.5024", "0.1528", "0.8906", "0.1515", "0.8906", "0.1666", "0.5024", "0.1674" ], [ "0.5024", "0.1674", "0.892", "0.1674", "0.892", "0.1812", "0.5024", "0.1812" ], [ "0.1462", "0.1843", "0.1871", "0.1843", "0.1871", "0.1989", "0.1462", "0.1989" ], [ "0.2296", "0.1843", "0.4019", "0.1843", "0.4019", "0.1976", "0.2296", "0.1976" ], [ "0.4468", "0.1843", "0.4714", "0.1843", "0.4714", "0.1963", "0.4468", "0.1963" ], [ "0.5034", "0.1831", "0.892", "0.1831", "0.892", "0.1963", "0.5034", "0.1963" ], [ "0.1978", "0.1932", "0.2303", "0.1932", "0.2303", "0.2052", "0.1978", "0.2052" ], [ "0.5034", "0.1976", "0.892", "0.1976", "0.892", "0.2108", "0.5034", "0.2108" ], [ "0.0801", "0.214", "0.4697", "0.214", "0.4697", "0.2291", "0.0801", "0.2291" ], [ "0.5024", "0.2134", "0.892", "0.2134", "0.892", "0.2267", "0.5024", "0.2267" ], [ "0.0801", "0.2311", "0.468", "0.2311", "0.468", "0.2456", "0.0801", "0.2456" ], [ "0.5034", "0.2299", "0.8906", "0.2299", "0.8906", "0.2405", "0.5034", "0.2405" ], [ "0.0801", "0.2468", "0.4697", "0.2468", "0.4697", "0.2615", "0.0801", "0.2615" ], [ "0.5024", "0.2437", "0.892", "0.2437", "0.892", "0.2568", "0.5024", "0.2568" ], [ "0.0784", "0.262", "0.2484", "0.26", "0.2484", "0.2747", "0.0784", "0.2766" ], [ "0.5024", "0.2583", "0.892", "0.2583", "0.892", "0.2715", "0.5024", "0.2715" ], [ "0.5024", "0.2734", "0.892", "0.2734", "0.892", "0.2874", "0.5024", "0.2874" ], [ "0.1478", "0.2827", "0.2607", "0.2827", "0.2607", "0.3005", "0.1478", "0.3005" ], [ "0.286", "0.2854", "0.361", "0.2827", "0.361", "0.2974", "0.286", "0.2998" ], [ "0.3586", "0.2935", "0.4697", "0.2935", "0.4697", "0.307", "0.3586", "0.307" ], [ "0.5024", "0.289", "0.8936", "0.289", "0.8936", "0.303", "0.5024", "0.303" ], [ "0.09314", "0.2954", "0.1299", "0.2954", "0.1299", "0.3062", "0.09314", "0.3062" ], [ "0.179", "0.2998", "0.2303", "0.3018", "0.2296", "0.3164", "0.1772", "0.3145" ], [ "0.259", "0.2966", "0.281", "0.2966", "0.281", "0.3037", "0.259", "0.3037" ], [ "0.312", "0.3037", "0.3506", "0.3037", "0.3506", "0.3145", "0.312", "0.3145" ], [ "0.5034", "0.3037", "0.8936", "0.3037", "0.8936", "0.3176", "0.5034", "0.3176" ], [ "0.0784", "0.3213", "0.4697", "0.3225", "0.4697", "0.3372", "0.0784", "0.336" ], [ "0.505", "0.3208", "0.8906", "0.3208", "0.8906", "0.3308", "0.505", "0.3308" ], [ "0.0801", "0.3384", "0.4714", "0.3384", "0.4714", "0.3528", "0.0801", "0.3528" ], [ "0.501", "0.3333", "0.8906", "0.3333", "0.8906", "0.3467", "0.501", "0.3467" ], [ "0.0809", "0.3542", "0.4697", "0.3542", "0.4697", "0.3687", "0.0809", "0.3687" ], [ "0.5034", "0.351", "0.8906", "0.351", "0.8906", "0.361", "0.5034", "0.361" ], [ "0.5024", "0.3643", "0.892", "0.3643", "0.892", "0.3774", "0.5024", "0.3774" ], [ "0.0809", "0.3706", "0.4673", "0.3706", "0.4673", "0.3806", "0.0809", "0.3806" ], [ "0.0801", "0.3838", "0.406", "0.3833", "0.406", "0.3977", "0.0801", "0.3984" ], [ "0.5034", "0.3801", "0.8936", "0.3801", "0.8936", "0.3933", "0.5034", "0.3933" ], [ "0.5024", "0.3945", "0.6357", "0.3945", "0.6357", "0.4077", "0.5024", "0.4077" ], [ "0.0809", "0.4016", "0.4673", "0.4016", "0.4673", "0.4153", "0.0809", "0.4153" ], [ "0.0801", "0.4172", "0.468", "0.4172", "0.468", "0.4312", "0.0801", "0.4312" ], [ "0.5024", "0.4268", "0.826", "0.428", "0.826", "0.4458", "0.5024", "0.4443" ], [ "0.0801", "0.433", "0.468", "0.433", "0.468", "0.4463", "0.0801", "0.4463" ], [ "0.0809", "0.4495", "0.468", "0.4495", "0.468", "0.4602", "0.0809", "0.4602" ], [ "0.5215", "0.4558", "0.892", "0.4558", "0.892", "0.4697", "0.5215", "0.4697" ], [ "0.0801", "0.4634", "0.4697", "0.4634", "0.4697", "0.4766", "0.0801", "0.4766" ], [ "0.5024", "0.4702", "0.8936", "0.4702", "0.8936", "0.4841", "0.5024", "0.4841" ], [ "0.0801", "0.4778", "0.4697", "0.4778", "0.4697", "0.4912", "0.0801", "0.4912" ], [ "0.5024", "0.486", "0.892", "0.486", "0.892", "0.5", "0.5024", "0.5" ], [ "0.0801", "0.4937", "0.468", "0.4937", "0.468", "0.507", "0.0801", "0.507" ], [ "0.5034", "0.502", "0.892", "0.502", "0.892", "0.515", "0.5034", "0.515" ], [ "0.0801", "0.509", "0.468", "0.509", "0.468", "0.523", "0.0801", "0.523" ], [ "0.0801", "0.5234", "0.4697", "0.5234", "0.4697", "0.537", "0.0801", "0.537" ], [ "0.5034", "0.518", "0.8906", "0.518", "0.8906", "0.529", "0.5034", "0.529" ], [ "0.5024", "0.532", "0.892", "0.532", "0.892", "0.5454", "0.5024", "0.5454" ], [ "0.0801", "0.539", "0.4697", "0.539", "0.4697", "0.553", "0.0801", "0.553" ], [ "0.501", "0.547", "0.892", "0.547", "0.892", "0.56", "0.501", "0.56" ], [ "0.0801", "0.5537", "0.468", "0.5537", "0.468", "0.5674", "0.0801", "0.5674" ], [ "0.0801", "0.5684", "0.468", "0.5684", "0.468", "0.582", "0.0801", "0.582" ], [ "0.5024", "0.5625", "0.8936", "0.5625", "0.8936", "0.5757", "0.5024", "0.5757" ], [ "0.5024", "0.577", "0.8906", "0.577", "0.8906", "0.5903", "0.5024", "0.5903" ], [ "0.0809", "0.5854", "0.4673", "0.5854", "0.4673", "0.595", "0.0809", "0.595" ], [ "0.501", "0.5913", "0.892", "0.5903", "0.892", "0.605", "0.501", "0.606" ], [ "0.0801", "0.5986", "0.468", "0.5986", "0.468", "0.6123", "0.0801", "0.6123" ], [ "0.5024", "0.607", "0.892", "0.607", "0.892", "0.6206", "0.5024", "0.6206" ], [ "0.0801", "0.6143", "0.4697", "0.6143", "0.4697", "0.6274", "0.0801", "0.6274" ], [ "0.5024", "0.6226", "0.892", "0.6226", "0.892", "0.636", "0.5024", "0.636" ], [ "0.0809", "0.631", "0.468", "0.631", "0.468", "0.6416", "0.0809", "0.6416" ], [ "0.501", "0.637", "0.826", "0.636", "0.826", "0.651", "0.501", "0.6514" ], [ "0.0801", "0.643", "0.468", "0.643", "0.468", "0.6577", "0.0801", "0.6577" ], [ "0.0801", "0.66", "0.468", "0.66", "0.468", "0.674", "0.0801", "0.674" ], [ "0.0801", "0.675", "0.4697", "0.675", "0.4697", "0.688", "0.0801", "0.688" ], [ "0.5024", "0.677", "0.5884", "0.677", "0.5884", "0.688", "0.5024", "0.688" ], [ "0.0801", "0.69", "0.3015", "0.69", "0.3015", "0.704", "0.0801", "0.704" ], [ "0.0784", "0.709", "0.3137", "0.7104", "0.3137", "0.728", "0.0784", "0.7266" ], [ "0.5215", "0.705", "0.892", "0.705", "0.892", "0.7197", "0.5215", "0.7197" ], [ "0.5034", "0.72", "0.8936", "0.72", "0.8936", "0.7344", "0.5034", "0.7344" ], [ "0.0989", "0.735", "0.468", "0.735", "0.468", "0.7485", "0.0989", "0.7485" ], [ "0.5024", "0.735", "0.892", "0.7344", "0.892", "0.7485", "0.5024", "0.75" ], [ "0.0801", "0.7505", "0.4714", "0.7505", "0.4714", "0.7646", "0.0801", "0.7646" ], [ "0.5024", "0.7505", "0.8906", "0.7505", "0.8906", "0.765", "0.5024", "0.765" ], [ "0.0801", "0.7666", "0.4697", "0.7666", "0.4697", "0.78", "0.0801", "0.78" ], [ "0.5024", "0.7666", "0.892", "0.7666", "0.892", "0.78", "0.5024", "0.78" ], [ "0.0809", "0.7827", "0.468", "0.7827", "0.468", "0.7935", "0.0809", "0.7935" ], [ "0.5024", "0.781", "0.892", "0.781", "0.892", "0.7944", "0.5024", "0.7944" ], [ "0.0801", "0.797", "0.4697", "0.797", "0.4697", "0.81", "0.0801", "0.81" ], [ "0.5034", "0.7974", "0.8906", "0.7974", "0.8906", "0.808", "0.5034", "0.808" ], [ "0.0801", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.0801", "0.8247" ], [ "0.5034", "0.811", "0.892", "0.811", "0.892", "0.8247", "0.5034", "0.8247" ], [ "0.0801", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0801", "0.8403" ], [ "0.501", "0.8247", "0.892", "0.8257", "0.892", "0.8403", "0.501", "0.839" ], [ "0.0801", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0801", "0.855" ], [ "0.501", "0.8413", "0.8906", "0.8413", "0.8906", "0.855", "0.501", "0.855" ], [ "0.0801", "0.857", "0.4697", "0.857", "0.4697", "0.8706", "0.0801", "0.8706" ], [ "0.5024", "0.857", "0.892", "0.857", "0.892", "0.8706", "0.5024", "0.8706" ], [ "0.0801", "0.872", "0.4714", "0.872", "0.4714", "0.8853", "0.0801", "0.8853" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0809", "0.888", "0.468", "0.888", "0.468", "0.8984", "0.0809", "0.8984" ], [ "0.501", "0.886", "0.634", "0.8853", "0.634", "0.8994", "0.501", "0.901" ] ]

[ "noise vector T ~ N(o, I) and updates according to the", "a high-dimensional manifold. Algorithm unfolding mod-", "discretizationofEq.(2).Alternatively.wecanunderstand", "els[4,6.73.63.75.24lcombinetraditionaloptimization", "the sampling process in the DDIM fashion [46], where the", "methodsandneuralnetworks,balancingefficiencyandin-", "scorefunctioncanalsobeconsideredto beadenoiserand", "terpretability. Unified models [59, 66, 58, 68, 18] avoid the", "predict the denoised &colt from any state &t at iteration t:", "problem of lacking training data and ground truth for specific", "tasks. Recently, CDDFuse [69] addresses cross-modality fea-", "5olt :", "(at +(1 -at)se(at,t),", "(4)", "ture modeling and extracts modality-specific/shared features", "Vat", "through a dual-branch Transformer-CNN architecture and", "and &olt denotes the estimation of aco given at. We use the", "correlation-driven loss, achieving promising fusion results", "same notation Qt = 1 - βt and Qt = IIt-- Qg following", "in multiple fusion tasks. On the other hand,fusion meth", "Ho et al. [13]. With this predicted 5olt and the current state", "ods have been combined with pattern recognition tasks such", "at, at-1 is updated from", "as semantic segmentation [50] and object detection [26] to", "explore the interactions with downstream tasks. Specifi-", "Vat(1 -Qt-1)", "Vat-iβt", "colt+otz,(5)", "cally, TarDAL [26] demonstrates an obvious advantage in", "t-1", "1-αt", "rt", "-at", "dealing with challenge scenarios with high efficiency. Self-", "where z ~ N(o,I) and a? is the variance which is usu-", "supervisedlearning[25Jisemployedtotrainfusionnetworks", "ally set to 0. This sampled at-1 is then fed into the next", "without paired images. Moreover, the pre-processing regis", "sampling iteration until the final image co is generated. Fur-", "trationmodule[60,14.54,61|canenhancetherobustness", "for unregistered input images. Benefiting from the multi-", "therdetailsaboutthissamplingprocesscanbefoundinthe", "supplementary material or the original paper [46]", "modality data, MSIS [17] achieves realizable and outstand-", "ing stitching results.", "Diffusion models applications. Recently, diffusion models", "have been improved to generate images with better quality", "2.3. Comparison with existing approaches", "than previous generative models like GANs [5, 39]. More-", "over,diffusion models can betreated as apowerful genera", "The methods most relevant to our model are optimization-", "tive prior and be applied to numerous conditional generation", "based methods and GAN-based generative methods. Con-", "tasks. One representative work with diffusion models is sta-", "ventional optimization-based methods are often limited by", "ble diffusion which can generate images according to given", "manually designed loss functions, which may not be flexible", "text prompts [44]. Diffusion models are also applied to", "many low-level vision tasks. For instance, DDRM [19] per-", "enough to capture all relevant aspects and are sensitive to", "changes in the data distribution. While incorporating nat-", "forms diffusion sampling in the spectral space of degradation", "ural imagepriors can provide extra knowledge that cannot", "operator A to reconstruct the missing information in the ob", "servation y. DDNM [64] shares a similar idea with DDRM", "be modeled by the generation loss function alone. Then,", "in contrast to GAN-based generative methods, where un-", "byrefiningthenull-spaceoftheoperatorAiterativelyfor", "stable training and pattern collapse may occur, our DDFM", "image restoration tasks. DPS [3] endorses Laplacian ap-", "achieves more stable and controllable fusion by rectifying", "proximation to calculate the gradient of log-likelihood for", "the generation process towards source images and perform-", "posteriorsampling anditis capableof manynoisynon-linear", "ing likelihood-based refinement in each iteration.", "inverse problems. In IIGDM [47], the authors employ few", "approximations to make the log-likelihood term tractable", "and hence make it able to solve inverse problems with even", "3.Method", "non-differentiable measurements.", "2.2. Multi-modal image fusion", "In this section, we first present a novel approach for ob-", "taining a fusion image by leveraging DDPM posterior sam-", "The deep learning-based multi-modality image fusion", "pling. Then, starting from the well-established loss function", "algorithms achieve effective feature extraction and infor-", "for image fusion, we derive a likelihood rectification ap-", "mation fusion through the powerful fitting ability of neu-", "proach for the unconditional DDPM sampling. Finally, we", "ral networks.Fusion algorithms are primarilydividedinto", "propose the DDFM algorithm, which embeds the solution of", "two branches: generative methods and discriminative meth-", "the hierarchical Bayesian inference into the diffusion sam", "ods. For generative methods [34, 31, 35], particularly the", "pling. In addition, the rationality of the proposed algorithm", "GAN family, adversarial training [7, 36, 38] is employed", "will be demonstrated. For brevity, we omit the derivations", "to generate fusion images following the same distribution", "of some equations and refer interested readers to the sup", "as the source images. For discriminative methods, auto", "plementary material. It is worth noting that we use IVF as", "encoder-based models [72, 23, 21, 28, 53, 22, 65] use en-", "a case to illustrate our DDFM, and MIF can be carried out", "coders and decoders to extractfeatures and fuse them on", "analogously to IVF" ]

{ "arxiv_id": "2303.06840", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.0341", "0.4795", "0.0341", "0.4795", "0.04736", "0.1748", "0.04736" ], [ "0.1748", "0.101", "0.3945", "0.1016", "0.3945", "0.1193", "0.1748", "0.1187" ], [ "0.384", "0.1016", "0.814", "0.1016", "0.814", "0.1193", "0.384", "0.1193" ], [ "0.808", "0.1092", "0.8145", "0.1092", "0.8145", "0.11426", "0.808", "0.11426" ], [ "0.1772", "0.1288", "0.492", "0.1288", "0.492", "0.1433", "0.1772", "0.1433" ], [ "0.4836", "0.1288", "0.8228", "0.1288", "0.8228", "0.1433", "0.4836", "0.1433" ], [ "0.1748", "0.1509", "0.2876", "0.154", "0.2876", "0.1718", "0.1748", "0.1686" ], [ "0.183", "0.1945", "0.255", "0.1963", "0.255", "0.2122", "0.183", "0.2102" ], [ "0.4028", "0.1925", "0.5146", "0.1945", "0.5146", "0.2122", "0.4028", "0.2102" ], [ "0.6006", "0.1945", "0.714", "0.1945", "0.714", "0.2122", "0.6006", "0.2122" ], [ "0.183", "0.2102", "0.326", "0.209", "0.326", "0.2235", "0.183", "0.2247" ], [ "0.4043", "0.2102", "0.576", "0.209", "0.576", "0.2235", "0.4043", "0.2247" ], [ "0.5986", "0.2102", "0.7417", "0.209", "0.7417", "0.2235", "0.5986", "0.2247" ], [ "0.1847", "0.226", "0.3816", "0.226", "0.3816", "0.2393", "0.1847", "0.2393" ], [ "0.4043", "0.226", "0.5483", "0.226", "0.5483", "0.2393", "0.4043", "0.2393" ], [ "0.601", "0.2267", "0.8037", "0.2267", "0.8037", "0.2374", "0.601", "0.2374" ], [ "0.0196", "0.2676", "0.0678", "0.2676", "0.0678", "0.3801", "0.0196", "0.3801" ], [ "0.4534", "0.276", "0.5474", "0.276", "0.5474", "0.2903", "0.4534", "0.2903" ], [ "0.2345", "0.308", "0.7646", "0.308", "0.7646", "0.3213", "0.2345", "0.3213" ], [ "0.232", "0.3208", "0.7646", "0.3208", "0.7646", "0.3352", "0.232", "0.3352" ], [ "0.2336", "0.336", "0.7656", "0.336", "0.7656", "0.3499", "0.2336", "0.3499" ], [ "0.232", "0.3484", "0.7656", "0.3499", "0.7656", "0.3643", "0.232", "0.363" ], [ "0.02696", "0.3662", "0.05392", "0.3662", "0.05392", "0.3958", "0.02696", "0.3958" ], [ "0.2336", "0.363", "0.7656", "0.363", "0.7656", "0.377", "0.2336", "0.377" ], [ "0.2336", "0.3774", "0.7646", "0.3774", "0.7646", "0.3914", "0.2336", "0.3914" ], [ "0.02287", "0.3914", "0.0621", "0.3914", "0.0621", "0.4849", "0.02287", "0.4849" ], [ "0.04166", "0.3965", "0.04984", "0.3965", "0.04984", "0.3997", "0.04166", "0.3997" ], [ "0.2336", "0.3914", "0.7656", "0.3914", "0.7656", "0.4048", "0.2336", "0.4048" ], [ "0.2336", "0.4048", "0.7656", "0.4048", "0.7656", "0.4185", "0.2336", "0.4185" ], [ "0.2336", "0.4185", "0.7656", "0.4185", "0.7656", "0.432", "0.2336", "0.432" ], [ "0.2336", "0.433", "0.763", "0.433", "0.763", "0.4463", "0.2336", "0.4463" ], [ "0.2336", "0.4463", "0.7656", "0.4463", "0.7656", "0.4602", "0.2336", "0.4602" ], [ "0.2336", "0.4602", "0.767", "0.4602", "0.767", "0.4734", "0.2336", "0.4734" ], [ "0.2345", "0.476", "0.7646", "0.476", "0.7646", "0.486", "0.2345", "0.486" ], [ "0.02287", "0.4873", "0.05963", "0.4873", "0.05963", "0.7056", "0.02287", "0.7056" ], [ "0.2345", "0.488", "0.7656", "0.488", "0.7656", "0.502", "0.2345", "0.502" ], [ "0.2336", "0.502", "0.7656", "0.502", "0.7656", "0.515", "0.2336", "0.515" ], [ "0.232", "0.5166", "0.7646", "0.5166", "0.7646", "0.53", "0.232", "0.53" ], [ "0.232", "0.529", "0.759", "0.53", "0.759", "0.5444", "0.232", "0.5435" ], [ "0.1814", "0.5713", "0.1896", "0.5776", "0.183", "0.5835", "0.1748", "0.577" ], [ "0.2035", "0.5713", "0.3381", "0.5713", "0.3381", "0.586", "0.2035", "0.586" ], [ "0.1748", "0.603", "0.822", "0.603", "0.822", "0.616", "0.1748", "0.616" ], [ "0.1748", "0.617", "0.824", "0.617", "0.824", "0.632", "0.1748", "0.632" ], [ "0.1748", "0.632", "0.4983", "0.632", "0.4983", "0.646", "0.1748", "0.646" ], [ "0.4885", "0.63", "0.824", "0.631", "0.824", "0.646", "0.4885", "0.6445" ], [ "0.1748", "0.646", "0.8228", "0.646", "0.8228", "0.659", "0.1748", "0.659" ], [ "0.1732", "0.6577", "0.619", "0.659", "0.619", "0.674", "0.1732", "0.6724" ], [ "0.1724", "0.6787", "0.8228", "0.6797", "0.8228", "0.6943", "0.1724", "0.6934" ], [ "0.1748", "0.6943", "0.822", "0.6943", "0.822", "0.7085", "0.1748", "0.7085" ], [ "0.1732", "0.707", "0.8228", "0.707", "0.8228", "0.7217", "0.1732", "0.7217" ], [ "0.1732", "0.7217", "0.824", "0.7217", "0.824", "0.7363", "0.1732", "0.7363" ], [ "0.1748", "0.7363", "0.8228", "0.7363", "0.8228", "0.75", "0.1748", "0.75" ], [ "0.1748", "0.75", "0.824", "0.75", "0.824", "0.763", "0.1748", "0.763" ], [ "0.1732", "0.7646", "0.826", "0.765", "0.826", "0.78", "0.1732", "0.779" ], [ "0.1748", "0.78", "0.824", "0.78", "0.824", "0.7935", "0.1748", "0.7935" ], [ "0.1732", "0.7935", "0.824", "0.7935", "0.824", "0.8066", "0.1732", "0.8066" ], [ "0.1732", "0.806", "0.7607", "0.8066", "0.7607", "0.8213", "0.1732", "0.821" ], [ "0.1748", "0.8276", "0.824", "0.8276", "0.824", "0.8413", "0.1748", "0.8413" ], [ "0.1732", "0.8423", "0.8228", "0.8423", "0.8228", "0.856", "0.1732", "0.856" ], [ "0.1748", "0.856", "0.8228", "0.856", "0.8228", "0.869", "0.1748", "0.869" ], [ "0.1748", "0.8706", "0.824", "0.8706", "0.824", "0.884", "0.1748", "0.884" ], [ "0.1732", "0.884", "0.824", "0.884", "0.824", "0.898", "0.1732", "0.898" ], [ "0.1732", "0.898", "0.8228", "0.898", "0.8228", "0.912", "0.1732", "0.912" ], [ "0.1748", "0.912", "0.824", "0.912", "0.824", "0.9253", "0.1748", "0.9253" ], [ "0.4944", "0.9487", "0.505", "0.9487", "0.505", "0.961", "0.4944", "0.961" ] ]

[ "Published as a conference paper at ICLR 2022", "MEASURING THE", "E INTERPRETABILITYOF UNSUPER", "", "VISED REPRESENTATIONS", "SVIAOUANTIZEDREVERSE", "PROBING", "Iro Laina", "Yuki M. Asano", "Andrea Vedaldi", "University of Oxford", "University of Amsterdam", "University of Oxford", "iro.laina@eng.ox.ac.uk", "y.m.asano@uva.nl", "vedaldi@robots.ox.ac.uk", " des ", "ABSTRACT", "Self-supervised visual representation learning has recently attracted significant re-", "search interest. While a common way to evaluate self-supervised representations", "is through transfer to various downstream tasks, we instead investigate the prob-", "lem of measuring their interpretability, i.e. understanding the semantics encoded in", "7", "raw representations. We formulate the latter as estimating the mutual information", "between the representation and a space of manually labelled concepts. To quantify", "['s]", "", "this we introduce a decoding bottleneck: information must be captured by simple", "predictors, mapping concepts to clusters in representation space. This approach,", "which we call reverse linear probing, provides a single number sensitive to the se-", "manticity of the representation. This measure is also able to detect when the rep-", "resentation contains combinations of concepts (e.g.,\"red apple\") instead of just", "individual attributes (\"red\"and“\"apple\"independently).Finally,we propose to", "usesupervisedclassifierstoautomaticallylabellargedatasetsinordertoenrichthe", "18920621", "space of concepts used for probing. We use our method to evaluate a large number", "of self-supervised representations, ranking them by interpretability, highlight the", "differences that emerge compared to the standard evaluation with linear probes and", "discuss several qualitative insights. Code at: https: //github.com/iro-cp/ssl-qrp.", "1", "INTRODUCTION", "Relying on black-box models such as deep networks comes sometimes with significant methodolog", "ical and ethical challenges. This is particularly true for unsupervised and self-supervised models", "which are learned without human supervision.", "While these models perform increasingly well in", "downstream applications, often outperforming supervised counterparts, there is very little under-", "standing of what they learn, making their real-life deployment risky.", "In this paper, we thus consider the problem of characterizing the meaning of data representations,", "with a particular focus on unsupervised and self-supervised representations of images. Given a rep-", "resentation f mapping images r to representation vectors f(α) e IRD, our goal is to find whether", "these vectors contain human-interpretable information. This is usually done by finding a relationship", "or correspondence between f() and human-provided descriptions y(α) of the images, essentially", "translating the information from the representation space to a concept space. Consider for example", "the popular linear probing method (Alain and Bengio, 2017). Given a large dataset of images", "with corresponding manual annotations y(), one learns linear classifiers (probes) to map the fea-", "ture vectors f() to labels y(), measuring the resulting classification accuracy. If the predictor is", "accurate, then one can argue that the representation captures the corresponding concept y.", "One possible issue with this type of approaches is that one can only discover in the representation", "concepts that are represented in the available data annotations. In order to maximize the semantic", "coverage of the analysis, it is thus customary to combine annotations for several types of attributes", "such as object classes, textures, colour, etc. (Bau et al., 2017). In this case, y is a vector of image", "attributes, and one can compute several linear probes to predict each individual attribute yi. By", "doing so, however, attributes are treated independently, which may be unsatisfactory. In order to", "obtain a single number assessing the overall semanticity of a representation, one must combine", "1" ]

{ "arxiv_id": "2209.03268", "language": "en", "page_idx": 0 }

[ [ "0.1732", "0.03284", "0.4805", "0.0322", "0.4805", "0.04672", "0.1732", "0.04736" ], [ "0.1732", "0.1042", "0.8228", "0.1048", "0.8228", "0.1193", "0.1732", "0.1187" ], [ "0.1748", "0.1193", "0.824", "0.1193", "0.824", "0.1332", "0.1748", "0.1332" ], [ "0.1732", "0.132", "0.8228", "0.132", "0.8228", "0.1459", "0.1732", "0.1459" ], [ "0.1732", "0.1445", "0.8228", "0.1465", "0.8228", "0.161", "0.1732", "0.159" ], [ "0.1748", "0.1604", "0.8228", "0.1604", "0.8228", "0.1736", "0.1748", "0.1736" ], [ "0.1748", "0.1736", "0.8145", "0.1736", "0.8145", "0.1875", "0.1748", "0.1875" ], [ "0.1765", "0.1957", "0.824", "0.1957", "0.824", "0.209", "0.1765", "0.209" ], [ "0.1748", "0.2102", "0.8228", "0.2102", "0.8228", "0.2235", "0.1748", "0.2235" ], [ "0.1748", "0.2235", "0.824", "0.2235", "0.824", "0.2374", "0.1748", "0.2374" ], [ "0.1732", "0.238", "0.7485", "0.238", "0.7485", "0.252", "0.1732", "0.252" ], [ "0.1732", "0.2568", "0.824", "0.2583", "0.824", "0.2727", "0.1732", "0.2715" ], [ "0.1748", "0.2727", "0.8228", "0.2727", "0.8228", "0.286", "0.1748", "0.286" ], [ "0.1732", "0.2874", "0.824", "0.286", "0.824", "0.2998", "0.1732", "0.3005" ], [ "0.1765", "0.3018", "0.8228", "0.3018", "0.8228", "0.3125", "0.1765", "0.3125" ], [ "0.1732", "0.3132", "0.8228", "0.3132", "0.8228", "0.327", "0.1732", "0.327" ], [ "0.1748", "0.3276", "0.824", "0.3276", "0.824", "0.3416", "0.1748", "0.3416" ], [ "0.1748", "0.3423", "0.6177", "0.3423", "0.6177", "0.356", "0.1748", "0.356" ], [ "0.1748", "0.361", "0.8203", "0.361", "0.8203", "0.3757", "0.1748", "0.3757" ], [ "0.1748", "0.377", "0.8228", "0.377", "0.8228", "0.3901", "0.1748", "0.3901" ], [ "0.1748", "0.3914", "0.8228", "0.3914", "0.8228", "0.4048", "0.1748", "0.4048" ], [ "0.1748", "0.4048", "0.824", "0.4048", "0.824", "0.4185", "0.1748", "0.4185" ], [ "0.1765", "0.4185", "0.822", "0.4185", "0.822", "0.432", "0.1765", "0.432" ], [ "0.1748", "0.4312", "0.8228", "0.432", "0.8228", "0.4463", "0.1748", "0.4458" ], [ "0.1732", "0.4463", "0.824", "0.4463", "0.824", "0.4602", "0.1732", "0.4602" ], [ "0.1748", "0.4602", "0.826", "0.4602", "0.826", "0.4734", "0.1748", "0.4734" ], [ "0.1748", "0.4734", "0.7573", "0.4734", "0.7573", "0.4873", "0.1748", "0.4873" ], [ "0.1765", "0.5103", "0.1896", "0.5103", "0.1896", "0.5215", "0.1765", "0.5215" ], [ "0.2018", "0.509", "0.3464", "0.509", "0.3464", "0.5234", "0.2018", "0.5234" ], [ "0.1748", "0.5405", "0.822", "0.5405", "0.822", "0.5537", "0.1748", "0.5537" ], [ "0.1732", "0.5547", "0.8228", "0.5547", "0.8228", "0.5693", "0.1732", "0.5693" ], [ "0.1765", "0.571", "0.3179", "0.571", "0.3179", "0.5806", "0.1765", "0.5806" ], [ "0.3105", "0.5693", "0.8228", "0.5693", "0.8228", "0.5835", "0.3105", "0.5835" ], [ "0.1748", "0.5835", "0.824", "0.5835", "0.824", "0.5967", "0.1748", "0.5967" ], [ "0.1748", "0.5967", "0.824", "0.5967", "0.824", "0.61", "0.1748", "0.61" ], [ "0.1732", "0.61", "0.824", "0.6094", "0.824", "0.6235", "0.1732", "0.625" ], [ "0.1748", "0.625", "0.8228", "0.625", "0.8228", "0.638", "0.1748", "0.638" ], [ "0.1748", "0.638", "0.826", "0.638", "0.826", "0.6514", "0.1748", "0.6514" ], [ "0.1765", "0.654", "0.8228", "0.654", "0.8228", "0.664", "0.1765", "0.664" ], [ "0.1732", "0.6665", "0.8228", "0.6665", "0.8228", "0.6797", "0.1732", "0.6797" ], [ "0.1748", "0.6797", "0.8228", "0.6797", "0.8228", "0.6934", "0.1748", "0.6934" ], [ "0.1748", "0.6934", "0.824", "0.6934", "0.824", "0.707", "0.1748", "0.707" ], [ "0.1765", "0.707", "0.824", "0.707", "0.824", "0.72", "0.1765", "0.72" ], [ "0.1748", "0.7197", "0.824", "0.72", "0.824", "0.735", "0.1748", "0.7344" ], [ "0.1748", "0.735", "0.7983", "0.735", "0.7983", "0.7485", "0.1748", "0.7485" ], [ "0.1748", "0.7603", "0.8228", "0.7603", "0.8228", "0.7734", "0.1748", "0.7734" ], [ "0.1732", "0.7734", "0.8228", "0.7734", "0.8228", "0.787", "0.1732", "0.787" ], [ "0.1748", "0.787", "0.8228", "0.787", "0.8228", "0.8003", "0.1748", "0.8003" ], [ "0.1765", "0.8037", "0.8228", "0.8037", "0.8228", "0.8145", "0.1765", "0.8145" ], [ "0.1748", "0.8164", "0.824", "0.8164", "0.824", "0.83", "0.1748", "0.83" ], [ "0.1732", "0.83", "0.8228", "0.83", "0.8228", "0.8433", "0.1732", "0.8433" ], [ "0.1748", "0.8433", "0.824", "0.8433", "0.824", "0.857", "0.1748", "0.857" ], [ "0.1748", "0.858", "0.824", "0.858", "0.824", "0.872", "0.1748", "0.872" ], [ "0.1732", "0.872", "0.7485", "0.872", "0.7485", "0.8853", "0.1732", "0.8853" ], [ "0.1732", "0.896", "0.824", "0.8965", "0.824", "0.911", "0.1732", "0.91" ], [ "0.1748", "0.911", "0.824", "0.911", "0.824", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2022", "the prediction accuracies of several independent classifiers, and there is no natural way of doing", "so (e.g., a simple average would not account for the different complexity of the different prediction", "tasks). Furthermore, the representation might be predictive of combinations of attributes, i.e. it might", "understand the concept of “red apple\" without necessarily understanding the individual concepts,", "\"red\" and “apple\". While it is in principle possible to test any combination of attributes via linear", "probing, most attribute combinations are too rare to generate significant statistics for this analysis.", "We propose a complementary assessment strategy to address these shortcomings. In contrast to", "linear probes, we start by considering the reverse prediction problem, mapping label vectors y()", "to representation vectors f(c). A key advantage is that the entire attribute vector is used for this", "mapping, which, as we show later, accounts for attribute combinations more effectively.", "Next, we consider the challenge of deriving a quantity that allows to compare representations based", "on the performance of these reverse predictors. Obviously a simple metric such as the average L2", "prediction error is meaningless as its magnitude would be affected by irrelevant factors such as", "thescaleof therepresentationvectors.To solvethisproblemin aprincipledmanner, weconsider", "instead the mutual information between the concepts and quantized representation vectors. This", "approach, which we justify from the viewpoint of information theory, essentially measures whether", "the representation groups the data in a human-interpretable manner.", "We use our approach to evaluate a large number of self-supervised and supervised image represen.", "tations. Importantly, we show that (a) all methods capture interpretable concepts that extend well", "beyond the underlying label distribution (e.g., ImageNet labels, which are not used for training), (b)", "while some clusters that form in the representation space are purely semantic (object-centric), others", "carry information about scenery, material, textures, or combined concepts, and (c) context matters.", "We also show that more performant methods recover more of the original label distribution, i.e. learn", "better \"ImageNet concepts\", and rely less on lower-level concepts. Quantitatively, we observe that", "our interpretability measure results in a similar but not identical ranking for state-of-the-art methods", "with clustering-based approaches generally producing more interpretable representations.", "2", "RELATEDWORK", "Self-supervised representation learning (ssL)In this paper, we focus on self-supervised meth-", "ods that learn representations from image data. Early self-supervised learning approaches devise nu-", "merouspretexttasks-", "colorization, predicting image rotations or solving image puzzlesto learn", "useful representations from unlabeled images (Gidaris et al., 2018; Pathak et al.,2016;Noroozi and", "Favaro, 2016; Doersch et al., 2015; Larsson et al., 2016; Zhang et al., 2017; 2016).More recent", "approaches follow the contrastive learning paradigm (Chen et al., 2020a; Frankle et al., 2020; He", "et al.,2020; Chen et al.,2020b;Henaff et al.,2019; Oord et al., 2018; Misra and Maaten,2020;", "Tian et al., 2019; 2020; Wu et al., 2018), where the goal is to discriminate instances belonging to", "thesameimagefromnegativesamples.Latermethodstakeacloserlookattheeffectofnegatives", "samples (Mitrovic et al., 2020; Robinson et al.,2020; Chuang et al., 2020; Kalantidis et al., 2020)", "or eliminate the need for negatives altogether (Zbontar et al., 2021; Chen and He, 2020; Grill et al.", "2020; Caron et al., 2021), while others consider nearest neighbors (Dwibedi et al., 2021; Assran", "et al., 2021). Although several studies aim to explain why constrastive learning works (Arora et al.,", "2019; Wang and Isola,2020; Tschannen et al.,2020; Tsai et al., 2020; Purushwalkam and Gupta,", "2020), in most cases, the focus lies on measuring empirical improvements in downstream tasks.", "ClusteringA related part of literature deals with unsupervised image clustering. Early approaches", "include using autoencoders (Hinton and Salakhutdinov,2006), agglomerative clustering (Bautista", "et al., 2016) and partially ordered sets of hand-crafted features (Bautista et al., 2017).More recent", "methodscombinerepresentationlearning withclustering,usingmutual information(Jietal.,2019:", "Hu et al., 2017), K-means (Caron et al., 2018; Zhan et al., 2020), prototypes (Li et al., 2021) or", "optimal transport (Asano et al., 2020; Caron et al., 2020). Other methods build on strong feature", "representations and, at a second stage, cluster and further refine the network (Yan et al., 2020;", "Van Gansbeke et al., 2020; Lee et al., 2019). These methods produce pseudo-labels which can be", "used to evaluate representations by measuring the correlation to the ground-truth labels.", "Evaluation and analysis of SSL There are two main directions in evaluating self-supervised", "learning methods. The first one uses pre-trained representations as an initialization for subsequent", "2" ]

{ "arxiv_id": "2209.03268", "language": "en", "page_idx": 1 }

[ [ "0.1732", "0.03284", "0.4805", "0.0322", "0.4805", "0.04672", "0.1732", "0.04736" ], [ "0.1732", "0.1048", "0.822", "0.1048", "0.822", "0.1193", "0.1732", "0.1193" ], [ "0.1748", "0.1193", "0.822", "0.1193", "0.822", "0.1332", "0.1748", "0.1332" ], [ "0.1748", "0.1332", "0.822", "0.1332", "0.822", "0.1465", "0.1748", "0.1465" ], [ "0.1732", "0.1465", "0.8228", "0.1465", "0.8228", "0.1604", "0.1732", "0.1604" ], [ "0.1748", "0.1604", "0.8228", "0.1604", "0.8228", "0.1736", "0.1748", "0.1736" ], [ "0.1748", "0.1749", "0.824", "0.1749", "0.824", "0.1881", "0.1748", "0.1881" ], [ "0.1732", "0.1875", "0.8228", "0.1875", "0.8228", "0.2008", "0.1732", "0.2008" ], [ "0.1748", "0.202", "0.822", "0.202", "0.822", "0.2153", "0.1748", "0.2153" ], [ "0.1748", "0.2166", "0.8228", "0.2166", "0.8228", "0.2299", "0.1748", "0.2299" ], [ "0.1748", "0.2299", "0.826", "0.2299", "0.826", "0.2437", "0.1748", "0.2437" ], [ "0.1748", "0.2437", "0.63", "0.2437", "0.63", "0.2568", "0.1748", "0.2568" ], [ "0.1732", "0.2708", "0.822", "0.2695", "0.822", "0.2842", "0.1732", "0.2854" ], [ "0.1748", "0.2854", "0.8228", "0.2854", "0.8228", "0.2986", "0.1748", "0.2986" ], [ "0.1732", "0.2998", "0.824", "0.2998", "0.824", "0.3132", "0.1732", "0.3132" ], [ "0.1748", "0.3132", "0.824", "0.3132", "0.824", "0.327", "0.1748", "0.327" ], [ "0.1748", "0.3257", "0.8096", "0.3257", "0.8096", "0.339", "0.1748", "0.339" ], [ "0.1748", "0.3484", "0.822", "0.3484", "0.822", "0.3623", "0.1748", "0.3623" ], [ "0.1748", "0.3623", "0.824", "0.3623", "0.824", "0.377", "0.1748", "0.377" ], [ "0.1748", "0.3757", "0.8228", "0.3757", "0.8228", "0.3901", "0.1748", "0.3901" ], [ "0.1732", "0.3882", "0.824", "0.3894", "0.824", "0.404", "0.1732", "0.4028" ], [ "0.1748", "0.404", "0.824", "0.404", "0.824", "0.4172", "0.1748", "0.4172" ], [ "0.1732", "0.4172", "0.824", "0.4172", "0.824", "0.4312", "0.1732", "0.4312" ], [ "0.1732", "0.4312", "0.824", "0.4312", "0.824", "0.4443", "0.1732", "0.4443" ], [ "0.1748", "0.4458", "0.824", "0.4458", "0.824", "0.459", "0.1748", "0.459" ], [ "0.1732", "0.459", "0.824", "0.459", "0.824", "0.473", "0.1732", "0.473" ], [ "0.1765", "0.4749", "0.819", "0.4749", "0.819", "0.4849", "0.1765", "0.4849" ], [ "0.1748", "0.4873", "0.824", "0.4873", "0.824", "0.5005", "0.1748", "0.5005" ], [ "0.1732", "0.5005", "0.822", "0.5005", "0.822", "0.5146", "0.1732", "0.5146" ], [ "0.1732", "0.513", "0.824", "0.5146", "0.824", "0.529", "0.1732", "0.528" ], [ "0.1765", "0.529", "0.734", "0.529", "0.734", "0.5425", "0.1765", "0.5425" ], [ "0.1748", "0.549", "0.8228", "0.549", "0.8228", "0.563", "0.1748", "0.563" ], [ "0.1748", "0.563", "0.824", "0.563", "0.824", "0.577", "0.1748", "0.577" ], [ "0.1765", "0.577", "0.8228", "0.577", "0.8228", "0.5903", "0.1765", "0.5903" ], [ "0.1748", "0.5913", "0.824", "0.5913", "0.824", "0.605", "0.1748", "0.605" ], [ "0.1748", "0.605", "0.824", "0.605", "0.824", "0.6187", "0.1748", "0.6187" ], [ "0.1748", "0.6187", "0.7646", "0.6187", "0.7646", "0.632", "0.1748", "0.632" ], [ "0.1748", "0.655", "0.1929", "0.655", "0.1929", "0.6685", "0.1748", "0.6685" ], [ "0.2018", "0.655", "0.2834", "0.655", "0.2834", "0.6694", "0.2018", "0.6694" ], [ "0.1732", "0.686", "0.8228", "0.6875", "0.8228", "0.702", "0.1732", "0.7007" ], [ "0.1748", "0.7007", "0.824", "0.7007", "0.824", "0.7153", "0.1748", "0.7153" ], [ "0.1724", "0.711", "0.818", "0.711", "0.818", "0.732", "0.1724", "0.732" ], [ "0.1748", "0.735", "0.822", "0.7363", "0.822", "0.7505", "0.1748", "0.75" ], [ "0.1732", "0.7505", "0.822", "0.7505", "0.822", "0.7646", "0.1732", "0.7646" ], [ "0.1748", "0.765", "0.8228", "0.765", "0.8228", "0.779", "0.1748", "0.779" ], [ "0.1748", "0.779", "0.8228", "0.779", "0.8228", "0.7925", "0.1748", "0.7925" ], [ "0.1748", "0.7925", "0.8228", "0.7925", "0.8228", "0.806", "0.1748", "0.806" ], [ "0.1748", "0.8066", "0.8228", "0.8066", "0.8228", "0.821", "0.1748", "0.821" ], [ "0.1748", "0.821", "0.824", "0.821", "0.824", "0.834", "0.1748", "0.834" ], [ "0.1732", "0.834", "0.822", "0.834", "0.822", "0.8477", "0.1732", "0.8477" ], [ "0.1748", "0.855", "0.8228", "0.855", "0.8228", "0.8687", "0.1748", "0.8687" ], [ "0.1748", "0.869", "0.824", "0.869", "0.824", "0.8833", "0.1748", "0.8833" ], [ "0.1732", "0.8833", "0.824", "0.8833", "0.824", "0.8965", "0.1732", "0.8965" ], [ "0.1732", "0.898", "0.8228", "0.898", "0.8228", "0.911", "0.1732", "0.911" ], [ "0.1748", "0.911", "0.822", "0.911", "0.822", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2022", "supervised tasks, e.g., for object detection. The second approach is to train a linear classifier on the", "representation space which is kept frozen after pre-training. Some concerns have been voiced re-", "garding the generality of ImageNet linear classification accuracy as the main metric to evaluate rep-", "resentations (Kotar et al., 2021) and, as a result, several benchmarking suites with various datasets", "and tasks have been proposed (Zhai et al., 2019; Goyal et al., 2019; Van Horn et al., 2021; Kotar", "et al., 2021). Complementary to this, clustering in the frozen feature space has also been proposed", "as an evaluation metric (Zheltonozhskii et al., 2020; Sariyildiz et al.,2020).In addition, downstream", "dataset (Ericsson et al., 2021) and pretraining dataset (Zha0 et al., 2021; Cole et al., 2021) depen-", "dencies have been evaluated. Finally, recent investigations of self-supervised feature spaces aim to", "understand separability (Sehwag et al., 2020), concept generalization (Sariyildiz et al., 2020) and", "the effect of balanced vs. long-tailed training data (Kang et al., 2020).", "InterpretabilityAlthough a large amount of work has studied feature representations in Convo-", "lutional Neural Networks (CNNs), it is heavily focused on models trained with full supervision", "ZeilerandFergus(2014)andZhouet al.(2014)analyzenetworksbyvisualizingthemostactivating", "patches, while activation maximization methods (Mahendran and Vedaldi, 2016a;b; Nguyen et al.,", "2017; 2016; Olah et al., 2017; Simonyan et al., 2014) generate inputs to activate specific neurons.", "Another line of work focuses on understanding what information is present in intermediate repre-", "sentations of CNNs, usually mapping activations to high-level concepts. Alain and Bengio (2017)", "introduce linear probes as a means to understanding the dynamics of intermediate layers, by pre-", "dicting the target labels from these layers. Escorcia et al. (2015) also use a linear formulation to", "study the relationship of mid-level features and visual attributes, while Oramas et al. (2019) adopt", "this to predict task-specific classes from the features (similar to Alain and Bengio (2017)), select", "relevant features and generate a visual explanation. Similarly, Zhou et al. (2018) decompose feature", "vectors into a set of elementary and interpretable components and show decomposed Grad-CAM", "heat maps (Selvaraju et al., 2017) for concepts that contribute to the prediction. Furthermore, Kim", "etal.(2o18)relatefeaturevectorsandhuman-interpretableconceptsusingasetofuser-providedex", "amples for each concept and Bau et al. (2017)propose to quantify the interpretability of individual", "units by measuring the overlap between each unit and a set of densely annotated concepts, Finally", "Ghorbani et al. (2019) propose a method to automatically assign concepts to image segments and", "Yeh et al. (2020) analyze the completeness of these concepts for explaining the CNN.", "With the exception of (Laina et al., 2020; Bau et al., 2017; Fong and Vedaldi, 2018), relatively little", "work has been done to understand the emergence of interpretable visual concepts in self-supervised", "representations specifically.In particular, Laina et al.(2020)quantify the learnability and describa-", "bility of unsupervised image groupings, by measuring how successful humans are at understanding", "the learned concept from a set of provided examples. Instead, our approach does not depend on", "human input and is thus easily scalable to a wide range of methods and number of classes.", "3", "METHOD", "We are interested in measuring the semanticity of data representations. A representation f is a map", "f : X → RD that encodes data samples e X as vectors f() e R D. In this paper, we assume that", "the data are images X = R3xHxW and f is a deep neural network, but other choices are possible.", "The semantic content of an image r can be summarized by obtaining a label or description y(r) e )", "of the image from a human annotator. If the representation f(r) captures the meaning of the image", "well, then it should be predictive of the description y(α). The mutual information I(f(α), y(α))", "provides a natural measure of predictivity, and it is thus tempting to use this quantity as a measure", "of the overall semanticity of the representation. Note that, due to the data processing inequality", "(I(f(r), y(r)) ≤ I(,y(r))), the information is maximized by observing the raw image, i.e. by", "the identity representation f(r) = . Since data processing cannot increase information content, a", "useful representation must preserve information while also making it easier to decode and act on it.", "There are many possible definitions of what constitutes“easy decoding\".Common in literature is the", "transfer through a simple predictor (e.g., linear) to new tasks. Here we propose a definition that stems", "from the interpretation of differential entropy as the limit of discrete entropy for quantized versions", "of the corresponding variables (Cover and Thomas, 2006). In our case, the description y(α) is", "already discrete, but the representation vectors are continuous. We thus propose to discretize (vector-", "3" ]

{ "arxiv_id": "2209.03268", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1748", "0.04736" ], [ "0.1953", "0.1111", "0.3782", "0.1111", "0.3782", "0.1218", "0.1953", "0.1218" ], [ "0.4722", "0.11554", "0.5293", "0.11554", "0.5293", "0.1187", "0.4722", "0.1187" ], [ "0.5312", "0.11554", "0.5596", "0.11554", "0.5596", "0.1187", "0.5312", "0.1187" ], [ "0.693", "0.11554", "0.7173", "0.11554", "0.7173", "0.1187", "0.693", "0.1187" ], [ "0.7363", "0.113", "0.8037", "0.113", "0.8037", "0.1206", "0.7363", "0.1206" ], [ "0.1765", "0.2443", "0.8203", "0.2443", "0.8203", "0.2583", "0.1765", "0.2583" ], [ "0.1765", "0.26", "0.638", "0.26", "0.638", "0.2708", "0.1765", "0.2708" ], [ "0.6465", "0.2568", "0.822", "0.2588", "0.822", "0.2727", "0.6465", "0.2708" ], [ "0.1732", "0.2715", "0.4387", "0.2727", "0.4387", "0.286", "0.1732", "0.2854" ], [ "0.5", "0.2727", "0.824", "0.2727", "0.824", "0.286", "0.5", "0.286" ], [ "0.1765", "0.2886", "0.824", "0.2886", "0.824", "0.2986", "0.1765", "0.2986" ], [ "0.1748", "0.3005", "0.71", "0.3005", "0.71", "0.3145", "0.1748", "0.3145" ], [ "0.1732", "0.3372", "0.822", "0.336", "0.822", "0.351", "0.1732", "0.3516" ], [ "0.1748", "0.3516", "0.8228", "0.3516", "0.8228", "0.3655", "0.1748", "0.3655" ], [ "0.1765", "0.3674", "0.822", "0.3674", "0.822", "0.3774", "0.1765", "0.3774" ], [ "0.1732", "0.379", "0.822", "0.379", "0.822", "0.3926", "0.1732", "0.3926" ], [ "0.1732", "0.3933", "0.824", "0.3926", "0.824", "0.4072", "0.1732", "0.4077" ], [ "0.1748", "0.4077", "0.3245", "0.4077", "0.3245", "0.4216", "0.1748", "0.4216" ], [ "0.1748", "0.4287", "0.822", "0.4287", "0.822", "0.443", "0.1748", "0.443" ], [ "0.1724", "0.4426", "0.3423", "0.4395", "0.3423", "0.457", "0.1724", "0.4602" ], [ "0.367", "0.4458", "0.3962", "0.4458", "0.3962", "0.4539", "0.367", "0.4539" ], [ "0.3945", "0.4475", "0.407", "0.4475", "0.407", "0.4539", "0.3945", "0.4539" ], [ "0.407", "0.443", "0.824", "0.443", "0.824", "0.457", "0.407", "0.457" ], [ "0.1732", "0.457", "0.5034", "0.457", "0.5034", "0.4717", "0.1732", "0.4717" ], [ "0.2573", "0.4749", "0.7417", "0.4749", "0.7417", "0.488", "0.2573", "0.488" ], [ "0.8022", "0.4749", "0.826", "0.4749", "0.826", "0.488", "0.8022", "0.488" ], [ "0.1732", "0.4912", "0.8228", "0.4924", "0.8228", "0.507", "0.1732", "0.506" ], [ "0.1732", "0.505", "0.8228", "0.506", "0.8228", "0.521", "0.1732", "0.5195" ], [ "0.1748", "0.521", "0.824", "0.521", "0.824", "0.535", "0.1748", "0.535" ], [ "0.1748", "0.534", "0.4453", "0.534", "0.4453", "0.5474", "0.1748", "0.5474" ], [ "0.246", "0.552", "0.751", "0.552", "0.751", "0.5664", "0.246", "0.5664" ], [ "0.1732", "0.5684", "0.8203", "0.5684", "0.8203", "0.582", "0.1732", "0.582" ], [ "0.1732", "0.5835", "0.822", "0.5835", "0.822", "0.5967", "0.1732", "0.5967" ], [ "0.1748", "0.5977", "0.824", "0.5977", "0.824", "0.6113", "0.1748", "0.6113" ], [ "0.1732", "0.6123", "0.7793", "0.6123", "0.7793", "0.6255", "0.1732", "0.6255" ], [ "0.1732", "0.6333", "0.8228", "0.6333", "0.8228", "0.6465", "0.1732", "0.6465" ], [ "0.1748", "0.6465", "0.824", "0.6465", "0.824", "0.661", "0.1748", "0.661" ], [ "0.1732", "0.66", "0.8228", "0.662", "0.8228", "0.677", "0.1732", "0.675" ], [ "0.1748", "0.677", "0.826", "0.677", "0.826", "0.69", "0.1748", "0.69" ], [ "0.1765", "0.6924", "0.808", "0.6924", "0.808", "0.7026", "0.1765", "0.7026" ], [ "0.1748", "0.712", "0.822", "0.712", "0.822", "0.7266", "0.1748", "0.7266" ], [ "0.1748", "0.728", "0.8228", "0.728", "0.8228", "0.741", "0.1748", "0.741" ], [ "0.1765", "0.7437", "0.8228", "0.7437", "0.8228", "0.754", "0.1765", "0.754" ], [ "0.1732", "0.755", "0.6772", "0.755", "0.6772", "0.768", "0.1732", "0.768" ], [ "0.1748", "0.778", "0.8228", "0.778", "0.8228", "0.7915", "0.1748", "0.7915" ], [ "0.1748", "0.7925", "0.822", "0.7925", "0.822", "0.806", "0.1748", "0.806" ], [ "0.1748", "0.806", "0.8228", "0.806", "0.8228", "0.8193", "0.1748", "0.8193" ], [ "0.1732", "0.8193", "0.8228", "0.821", "0.8228", "0.834", "0.1732", "0.8335" ], [ "0.1732", "0.834", "0.808", "0.834", "0.808", "0.8477", "0.1732", "0.8477" ], [ "0.1765", "0.855", "0.824", "0.855", "0.824", "0.8687", "0.1765", "0.8687" ], [ "0.1772", "0.8706", "0.8228", "0.8706", "0.8228", "0.881", "0.1772", "0.881" ], [ "0.1748", "0.8833", "0.8228", "0.8833", "0.8228", "0.8965", "0.1748", "0.8965" ], [ "0.1732", "0.898", "0.826", "0.898", "0.826", "0.911", "0.1732", "0.911" ], [ "0.1748", "0.911", "0.7075", "0.911", "0.7075", "0.925", "0.1748", "0.925" ] ]

[ "Published as a conference paper at ICLR 2022", "Featureextractionand quantization", "OnCeOTY", "aTe", "a", "linearprobing", "Figure l: Overview of our approach. (1) we evaluate pre-trained SSL models on an image collec", "tion,extracting and quantizingfeaturevectors to obtain clusters", "),(2)welabeltheimage", "data with a diverse set of concepts (", ") from expert models trained with supervision on", "externaldata sources,and(3)wetrainalinearmodel hetomapconceptsto clusters,measuringthe", "mutual information between the representation and human-interpretable concepts.", "quantize) the representation and to use its information as a measure of semanticity. Intuitively, in the", "limit of infinitely-fine quantization, this quantity reduces (up to a normalizing term) to the mutual", "information above.For afinitenumber of clusters.however,alarge valueof information means", "that the representation groups images in a way that makes sense to a human observer. In our case", "the quantization amounts to grouping images based on the L2-distance between their representation", "vectors, i.e. clustering.", "Formally, we use a vector quantization algorithm such as K-means (Lloyd, 1982) to learn a quan", "tizer function μk : RD", "31.", "", "., K} for the representation vectors, and estimate the mutual", "information I(μk(f(α), y(c)) by rewriting it as:", "I(fk(α),y()) =H(fk(r)) -H(fk(r) y()), fk(α) =μk(f(r)).", "(1)", "In the above equation, the first term denotes the entropy of the cluster assignments. In practice", "given a sample dataset x of images, we firstrun K-means to compute the quantizer μk,and then", "compute the frequency of cluster assignments fk(α), E X to calculate the entropy. The second", "term in Eq. (1) is the conditional entropy:", "H(fk() [y(r)) = Ep[-Inp(fk(α) / y(α)] ≤ Ep[-Inq(fk(r) Iy(r))]", "where p is the true joint distribution between variables fk (α) and y(). While this is difficult tc", "compute, we can upper-bound it by considering an auxiliary posterior distribution g interpreting the", "conditional entropy as the cross-entropy loss of a predictor he : → K, parametrized by , that", "maps labels y(α) to clusters fk(). The gap can be minimized by learning the parameters .", "In practice, learning the predictor may result in overfitting, which would lead to an over-optimistic", "estimate of the mutual information. We address this issue by learning the predictor he on a subset", " C X of the data and evaluating its cross-entropy loss on the remaining subset X -X. Importantly", "we consider a linear predictor (probe)for he, to furtherreduce the risk of overfitting.We refer to", "this asguantizedreverseprobing,as itmaps semantic concepts toquantized datarepresentations.", "Number of clusters The number of clusters K controls the size of the bottleneck. Mutual in-", "formation increases monotonically as K increases, as we show in the Appendix (Fig. 6); when", "comparingrepresentations,itis thus importanttofixavalueofK and useitconsistently.Foradded", "convenience, in practice we report a normalised value of information (NMI).", "Obtaining labelled data via automatic experts Similar to prior work on interpretation, our", "method requires a dataset E X of images equipped with manually-provided labels y(r) y)", "Because we do not know a-priori which concepts may be captured by a given representation f()", "under analysis, these labels should provide a good coveragefor a number of different concepts (e.g.,", "textures, object types, parts, etc.). A good example of such a dataset is Broden (Bau et al., 2017).", "Because it would be cumbersome, and maybe even infeasible, to annotate any target dataset with all", "presentvisualconcepts.acost-effectivewaytoincreasethecoverageofthelabelspaceistopredict", "the labels y() automatically via a battery of expert classifiers learned using full supervision. The", "noise of these automated predictions is small compared to the noise in the correlation between the", "concepts and the unsupervised/self-supervised visual representations under study." ]

{ "arxiv_id": "2209.03268", "language": "en", "page_idx": 3 }

[ [ "0.1732", "0.03284", "0.4805", "0.0322", "0.4805", "0.04672", "0.1732", "0.04736" ], [ "0.1748", "0.2437", "0.482", "0.2437", "0.482", "0.2568", "0.1748", "0.2568" ], [ "0.517", "0.2437", "0.8228", "0.2437", "0.8228", "0.2568", "0.517", "0.2568" ], [ "0.1748", "0.2556", "0.286", "0.2556", "0.286", "0.2695", "0.1748", "0.2695" ], [ "0.517", "0.2556", "0.661", "0.2556", "0.661", "0.2695", "0.517", "0.2695" ], [ "0.1732", "0.281", "0.822", "0.281", "0.822", "0.2954", "0.1732", "0.2954" ], [ "0.1732", "0.2966", "0.8228", "0.2966", "0.8228", "0.31", "0.1732", "0.31" ], [ "0.1732", "0.3093", "0.8228", "0.3093", "0.8228", "0.3225", "0.1732", "0.3225" ], [ "0.1748", "0.3257", "0.822", "0.3257", "0.822", "0.336", "0.1748", "0.336" ], [ "0.1748", "0.339", "0.8228", "0.339", "0.8228", "0.3499", "0.1748", "0.3499" ], [ "0.1748", "0.3516", "0.8022", "0.3516", "0.8022", "0.3655", "0.1748", "0.3655" ], [ "0.1748", "0.3926", "0.822", "0.3926", "0.822", "0.406", "0.1748", "0.406" ], [ "0.1748", "0.4072", "0.824", "0.4072", "0.824", "0.4204", "0.1748", "0.4204" ], [ "0.1748", "0.4204", "0.824", "0.4204", "0.824", "0.4343", "0.1748", "0.4343" ], [ "0.1732", "0.4343", "0.824", "0.433", "0.824", "0.4475", "0.1732", "0.449" ], [ "0.1748", "0.449", "0.822", "0.449", "0.822", "0.4622", "0.1748", "0.4622" ], [ "0.1748", "0.461", "0.822", "0.461", "0.822", "0.4749", "0.1748", "0.4749" ], [ "0.1732", "0.476", "0.824", "0.476", "0.824", "0.4893", "0.1732", "0.4893" ], [ "0.1732", "0.4893", "0.824", "0.4893", "0.824", "0.5024", "0.1732", "0.5024" ], [ "0.1748", "0.504", "0.8228", "0.504", "0.8228", "0.5176", "0.1748", "0.5176" ], [ "0.1748", "0.5176", "0.7305", "0.5176", "0.7305", "0.531", "0.1748", "0.531" ], [ "0.1748", "0.5386", "0.8228", "0.5386", "0.8228", "0.552", "0.1748", "0.552" ], [ "0.1748", "0.553", "0.824", "0.553", "0.824", "0.5664", "0.1748", "0.5664" ], [ "0.1748", "0.5664", "0.824", "0.5664", "0.824", "0.58", "0.1748", "0.58" ], [ "0.1748", "0.579", "0.8228", "0.579", "0.8228", "0.5923", "0.1748", "0.5923" ], [ "0.1748", "0.5947", "0.824", "0.5947", "0.824", "0.608", "0.1748", "0.608" ], [ "0.1748", "0.608", "0.824", "0.608", "0.824", "0.622", "0.1748", "0.622" ], [ "0.1748", "0.6206", "0.8228", "0.6206", "0.8228", "0.634", "0.1748", "0.634" ], [ "0.1748", "0.6353", "0.8228", "0.6353", "0.8228", "0.6484", "0.1748", "0.6484" ], [ "0.1732", "0.6494", "0.824", "0.6494", "0.824", "0.6636", "0.1732", "0.6636" ], [ "0.1748", "0.6636", "0.822", "0.6636", "0.822", "0.677", "0.1748", "0.677" ], [ "0.1748", "0.677", "0.8228", "0.677", "0.8228", "0.69", "0.1748", "0.69" ], [ "0.1748", "0.6914", "0.824", "0.6914", "0.824", "0.705", "0.1748", "0.705" ], [ "0.1732", "0.705", "0.822", "0.705", "0.822", "0.7183", "0.1732", "0.7183" ], [ "0.1748", "0.7173", "0.824", "0.7173", "0.824", "0.731", "0.1748", "0.731" ], [ "0.1748", "0.7344", "0.671", "0.7344", "0.671", "0.744", "0.1748", "0.744" ], [ "0.1765", "0.768", "0.1912", "0.768", "0.1912", "0.779", "0.1765", "0.779" ], [ "0.2051", "0.7676", "0.3284", "0.7676", "0.3284", "0.781", "0.2051", "0.781" ], [ "0.1748", "0.7974", "0.826", "0.7974", "0.826", "0.811", "0.1748", "0.811" ], [ "0.1748", "0.8125", "0.824", "0.8125", "0.824", "0.8257", "0.1748", "0.8257" ], [ "0.1732", "0.8257", "0.8228", "0.8257", "0.8228", "0.839", "0.1732", "0.839" ], [ "0.1748", "0.8403", "0.8022", "0.8403", "0.8022", "0.854", "0.1748", "0.854" ], [ "0.1765", "0.872", "0.2141", "0.872", "0.2141", "0.882", "0.1765", "0.882" ], [ "0.2059", "0.869", "0.406", "0.8706", "0.406", "0.8853", "0.2059", "0.884" ], [ "0.1748", "0.8965", "0.822", "0.8965", "0.822", "0.911", "0.1748", "0.911" ], [ "0.1748", "0.911", "0.822", "0.911", "0.822", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.961", "0.492", "0.961" ] ]

[ "Published as a conference paper at ICLR 2022", "(a) All attributes are linearly separable (Acc[hi] =", "(b) Color attribute not linearly separable (Acc[hi] =", "Acc[h2] = 100%).", "50%, Acc[h2]= 100%].", "Figure 2: Forward vs. reverse probing. A dataset X embedded in R2 by two different represen-", "tations, with meaningful clusters in representation space. Each data point has two binary attributes,", "color yi (r): red or blue and shape y2(α): or O. Both representations separate these attributes:", "however,colorisnotlinearlyseparablein(b),soforwardlinearprobingcannotrecoverthisrela-", "tionship.Decisionboundariesare shownfortheforward linear probes.Onthecontraryourreverse", "probing easily discovers that all combinations of shape and color map to well separated clusters.", "Relation to linear probing Linear probing is a common approach for assessing representations", "A linear probe is a simple predictor hi o f(α) ~ y;(α) that maps the representation f(r) to a", "specific concept yi (e.g., a binary attribute or a class label). The idea is that the representation", "captures concept yi if the probe performs well. As a result, linear probing has become a standard", "evaluation protocol in self-supervised representation to measure the (linear) classification accuracy", "of the learned representations against a labelled set, commonly ImageNet (Russakovsky et al., 2015)", "Ininterpretabilitystudies,differentformulationsoflinearprobing(Alain andBengio,2017)areused", "to understand intermediate layers of neural networks. Similar to our approach, the simplicity of the", "probes (imposed via linearity and sometimes sparsity) is a necessary bottleneck for interpretability:", "without such a constraint, the best possible probe would take the raw image as input.", "Our method is complementary to linear probes, with some advantages. To discuss this point further,", "we show an example in Fig. 2, where data points with color and shape attributes are encoded by two", "different representations, such that they form well-defined nameable clusters, i.e. blue square, red", "square, blue circle, and red circle. Forward linear probes (X → ) can be trained as binary classifiers", "hi, h2 for each attribute, mapping a feature vector (in this case, coordinates) to the corresponding", "attribute value for each data point. If attributes are linearly separable (Fig. 2a), then forward probes", "do well at separating the representation space. If an attribute is not linearly separable (color in", "Fig. 2b), the predictive accuracy of the corresponding probe hi reduces to chance, and consequently", "reduces the average score. As a result, linear probes cannot always assess the meaningfulness of", "clusters or, in other words, whether clusters in representation space respond to well-defined concepts", "On the contrary, if all data points in a cluster have consistent attributes, the reverse probe (→ X)", "achieves a high score for, e.g., \"red square\" without requiring the concept \"red\" to be also encoded", "in the features. In this case, we can say that the model has discovered the concept of a \"red square\"", "without identifying,in isolation, the concept\"red.Therefore,reverse probing handles combinations", "ofattributesnaturallyandallowstobetterassessthesemanticityofclusters.", "4 ", "EXPERIMENTS", "We use our method to evaluate a wide range of recent self-supervised representation learning and", "clustering techniques. In the following, we compare these techniques according to the proposed", "criterion, which results in a different ranking for state-of-the-art approaches. Further, we show", "qualitatively how our method discovers elementary concepts encoded in the raw representations", "4.1", "IMPLEMENTATIONDETAILS", "Attributes and expert models As a first step in our approach, we collect a set of attributes which", "we expect to find encoded in the self-supervised representations; these include semantic object cat-", "5" ]

{ "arxiv_id": "2209.03268", "language": "en", "page_idx": 4 }

[ [ "0.1316", "0.1332", "0.839", "0.1332", "0.839", "0.1477", "0.1316", "0.1477" ], [ "0.134", "0.1812", "0.2712", "0.1768", "0.273", "0.1976", "0.1356", "0.202" ], [ "0.3064", "0.1812", "0.4592", "0.1787", "0.4592", "0.1976", "0.3064", "0.1995" ], [ "0.4958", "0.1812", "0.642", "0.1799", "0.642", "0.1976", "0.4958", "0.1989" ], [ "0.683", "0.1831", "0.831", "0.1831", "0.831", "0.1976", "0.683", "0.1976" ], [ "0.183", "0.1963", "0.601", "0.1989", "0.601", "0.2166", "0.183", "0.214" ], [ "0.619", "0.1989", "0.786", "0.2008", "0.786", "0.2153", "0.619", "0.2134" ], [ "0.1609", "0.2197", "0.8037", "0.2197", "0.8037", "0.2343", "0.1609", "0.2343" ], [ "0.0196", "0.2676", "0.058", "0.2664", "0.0637", "0.3599", "0.02533", "0.361" ], [ "0.2361", "0.2676", "0.3145", "0.2676", "0.3145", "0.2854", "0.2361", "0.2854" ], [ "0.1013", "0.303", "0.468", "0.303", "0.468", "0.3164", "0.1013", "0.3164" ], [ "0.818", "0.31", "0.8267", "0.308", "0.83", "0.3125", "0.819", "0.3145" ], [ "0.0809", "0.318", "0.468", "0.318", "0.468", "0.332", "0.0809", "0.332" ], [ "0.0809", "0.3333", "0.4673", "0.3333", "0.4673", "0.3467", "0.0809", "0.3467" ], [ "0.03595", "0.3528", "0.04822", "0.3528", "0.04822", "0.363", "0.03595", "0.363" ], [ "0.0801", "0.3484", "0.4697", "0.3484", "0.4697", "0.3623", "0.0801", "0.3623" ], [ "0.0784", "0.363", "0.468", "0.3623", "0.468", "0.377", "0.0784", "0.3774" ], [ "0.02942", "0.3718", "0.05228", "0.3718", "0.05228", "0.3865", "0.02942", "0.3865" ], [ "0.0801", "0.379", "0.468", "0.379", "0.468", "0.3926", "0.0801", "0.3926" ], [ "0.02533", "0.3977", "0.05554", "0.3977", "0.05554", "0.4463", "0.02533", "0.4463" ], [ "0.0809", "0.3945", "0.468", "0.3945", "0.468", "0.4077", "0.0809", "0.4077" ], [ "0.0801", "0.4092", "0.468", "0.4092", "0.468", "0.4224", "0.0801", "0.4224" ], [ "0.0809", "0.4236", "0.468", "0.4236", "0.468", "0.4382", "0.0809", "0.4382" ], [ "0.0335", "0.4414", "0.05392", "0.4414", "0.05392", "0.4734", "0.0335", "0.4734" ], [ "0.0784", "0.4395", "0.4697", "0.4395", "0.4697", "0.4526", "0.0784", "0.4526" ], [ "0.0801", "0.4539", "0.468", "0.4539", "0.468", "0.4673", "0.0801", "0.4673" ], [ "0.0801", "0.4673", "0.468", "0.4685", "0.468", "0.483", "0.0801", "0.4817" ], [ "0.0237", "0.4873", "0.05637", "0.4873", "0.05637", "0.577", "0.0237", "0.577" ], [ "0.0801", "0.4841", "0.29", "0.4841", "0.29", "0.4976", "0.0801", "0.4976" ], [ "0.0825", "0.532", "0.2075", "0.532", "0.2075", "0.5454", "0.0825", "0.5454" ], [ "0.8423", "0.537", "0.859", "0.537", "0.859", "0.5415", "0.8423", "0.5415" ], [ "0.0972", "0.558", "0.468", "0.558", "0.468", "0.5713", "0.0972", "0.5713" ], [ "0.0784", "0.5728", "0.468", "0.5737", "0.468", "0.587", "0.0784", "0.586" ], [ "0.0335", "0.5903", "0.03757", "0.5903", "0.03757", "0.5996", "0.0335", "0.5996" ], [ "0.0801", "0.5884", "0.4697", "0.5884", "0.4697", "0.6016", "0.0801", "0.6016" ], [ "0.02777", "0.601", "0.05554", "0.601", "0.05554", "0.629", "0.02777", "0.629" ], [ "0.0801", "0.604", "0.468", "0.604", "0.468", "0.617", "0.0801", "0.617" ], [ "0.831", "0.604", "0.871", "0.604", "0.871", "0.619", "0.831", "0.619" ], [ "0.02942", "0.622", "0.05392", "0.622", "0.05392", "0.6396", "0.02942", "0.6396" ], [ "0.0801", "0.617", "0.4673", "0.617", "0.4673", "0.631", "0.0801", "0.631" ], [ "0.8384", "0.622", "0.8735", "0.622", "0.8735", "0.629", "0.8384", "0.629" ], [ "0.03513", "0.634", "0.05148", "0.634", "0.05148", "0.66", "0.03513", "0.66" ], [ "0.0801", "0.634", "0.468", "0.634", "0.468", "0.648", "0.0801", "0.648" ], [ "0.0784", "0.6484", "0.468", "0.648", "0.468", "0.662", "0.0784", "0.6636" ], [ "0.03186", "0.6636", "0.03757", "0.6636", "0.03757", "0.6704", "0.03186", "0.6704" ], [ "0.03757", "0.6655", "0.04984", "0.6655", "0.04984", "0.672", "0.03757", "0.672" ], [ "0.0801", "0.664", "0.4697", "0.664", "0.4697", "0.678", "0.0801", "0.678" ], [ "0.0335", "0.6704", "0.05392", "0.6704", "0.05392", "0.6997", "0.0335", "0.6997" ], [ "0.5024", "0.6675", "0.892", "0.6675", "0.892", "0.681", "0.5024", "0.681" ], [ "0.0801", "0.6797", "0.2091", "0.6797", "0.2091", "0.6934", "0.0801", "0.6934" ], [ "0.5024", "0.681", "0.892", "0.681", "0.892", "0.6943", "0.5024", "0.6943" ], [ "0.10046", "0.696", "0.4673", "0.696", "0.4673", "0.709", "0.10046", "0.709" ], [ "0.5034", "0.6963", "0.8906", "0.6963", "0.8906", "0.707", "0.5034", "0.707" ], [ "0.0809", "0.711", "0.468", "0.711", "0.468", "0.7246", "0.0809", "0.7246" ], [ "0.501", "0.7085", "0.8906", "0.7085", "0.8906", "0.7217", "0.501", "0.7217" ], [ "0.0809", "0.726", "0.4697", "0.726", "0.4697", "0.7393", "0.0809", "0.7393" ], [ "0.5034", "0.7246", "0.8906", "0.7246", "0.8906", "0.735", "0.5034", "0.735" ], [ "0.0809", "0.7407", "0.4697", "0.7407", "0.4697", "0.754", "0.0809", "0.754" ], [ "0.5034", "0.7373", "0.8906", "0.7373", "0.8906", "0.7505", "0.5034", "0.7505" ], [ "0.501", "0.7485", "0.892", "0.75", "0.892", "0.7646", "0.501", "0.763" ], [ "0.0801", "0.755", "0.4656", "0.755", "0.4656", "0.768", "0.0801", "0.768" ], [ "0.5024", "0.7646", "0.8145", "0.7646", "0.8145", "0.778", "0.5024", "0.778" ], [ "0.0784", "0.771", "0.468", "0.7695", "0.468", "0.784", "0.0784", "0.785" ], [ "0.0784", "0.786", "0.4019", "0.785", "0.4019", "0.799", "0.0784", "0.8" ], [ "0.0989", "0.802", "0.468", "0.802", "0.468", "0.815", "0.0989", "0.815" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.0784", "0.8174", "0.468", "0.8164", "0.468", "0.83", "0.0784", "0.8306" ], [ "0.5024", "0.827", "0.8896", "0.827", "0.8896", "0.8403", "0.5024", "0.8403" ], [ "0.0825", "0.834", "0.4656", "0.834", "0.4656", "0.8447", "0.0825", "0.8447" ], [ "0.5024", "0.8403", "0.8906", "0.839", "0.8906", "0.854", "0.5024", "0.855" ], [ "0.0801", "0.8477", "0.468", "0.8477", "0.468", "0.8623", "0.0801", "0.8623" ], [ "0.501", "0.856", "0.892", "0.855", "0.892", "0.869", "0.501", "0.8706" ], [ "0.0989", "0.877", "0.2075", "0.877", "0.2075", "0.887", "0.0989", "0.887" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0801", "0.887", "0.4028", "0.887", "0.4028", "0.9014", "0.0801", "0.9014" ], [ "0.5024", "0.886", "0.812", "0.886", "0.812", "0.901", "0.5024", "0.901" ], [ "0.4836", "0.928", "0.4878", "0.928", "0.4878", "0.933", "0.4836", "0.933" ] ]

[ "SinGRAF: Learning a 3D Generative Radiance Field for a Single Scene", "Minjung Son*1,2", "Jeong Joon Park*2", "Leonidas Guibas?", "Gordon Wetzstein", "1Samsung Advanced Institute of Technology (SAIT)", "2StanfordUniversity", "minjungs.son@samsung.com,(jjpark3d,gordon.wetzstein,guibas)@stanford.edu", "", "Abstract", "Generative models have shown great promise in syn-", "", "thesizing photorealistic 3D objects, but they require large", "amounts of training data. We introduce SinGRAF a 3D-", "A", "aware generative model that is trained with a few input", "images of a single scene. Once trained, SinGRAF gener-", "2", "ates different realizations of this 3D scene that preserve the", "LAD", "appearance of the input while varying scene layout. For", "this purpose, we build on recent progress in 3D GAN ar-", "chitectures and introduce a novel progressive-scale patch", "'SO", "discrimination approach during training. With several ex-", "periments, we demonstrate that the results produced by Sin-", "GRAF outperform the closest related works in both quality", "1109211", "and diversity by a large margin.", "1.Introduction", "", "Creating a new 3D asset is a laborious task, which often", "requires manual design of triangle meshes, texture maps,", "", "and object placements. As such, numerous methods were", "2", "proposed to automatically create diverse and realistic varia-", "", "2", "tions of existing 3D assets. For example, procedural model-", "", ".A", "ing techniques [11,27] produce variations in 3D assets given", "predefined rules and grammars, and example-based model-", "", "", "ing methods [13, 21] combine different 3D components to", "XP", "Figure 1. SinGRAF generates different plausible realizations of a", "generate new ones.", "single 3D scene from a few unposed input images of that scene.", "With our work, we propose a different, generative strat-", "In this example,i.e.,the\"office.3\"scene,we use100 input im-", "egy that is able to create realistic variations of a single 3D", "ages, four of which are shown in the top row. Next, we visualize", "scene from a small number of photographs. Unlike existing", "fourrealizationsofthe3Dsceneaspanoramas,renderedusingthe", "3D generative models, which typically require 3D assets as", "generated neural radiance fields. Note the variations in scene lay-", "out, including chairs, tables, lamps, and other parts, while staying", "input [13, 52], our approach only takes a set of unposed im-", "faithful to the structure and style of the input images.", "ages as input and outputs a generative model of a single 3D", "scene, represented as a neural radiance field [3]].", "Our method, dubbed SinGRAF, builds on recent", "is challenging, because they typically require a large train-", "progress in unconditional 3D-aware GANs [5,41] that train", "ing set of diverse images and often limit their optimal oper-", "generativeradiancefieldsfromasetofsingle-viewimages", "ating ranges to objects, rather than entire scenes. SinGRAF", "However, directly applying these 3D GANs to our problem", "makes a first attempt to train a 3D generative radiance field", "*Equalcontribution.", "for individual indoor 3D scenes, creating realistic 3D varia-", "Project page: computationalimaging.org/publications/singraf/", "tions in scene layout from unposed 2D images.", "-" ]

{ "arxiv_id": "2211.17260", "language": "en", "page_idx": 0 }

[ [ "0.6274", "0.09155", "0.66", "0.09155", "0.66", "0.0985", "0.6274", "0.0985" ], [ "0.8145", "0.0884", "0.8447", "0.0884", "0.8447", "0.101", "0.8145", "0.101" ], [ "0.687", "0.1048", "0.7837", "0.1067", "0.7837", "0.1187", "0.687", "0.11615" ], [ "0.0989", "0.09216", "0.468", "0.09216", "0.468", "0.1061", "0.0989", "0.1061" ], [ "0.0801", "0.108", "0.468", "0.108", "0.468", "0.1218", "0.0801", "0.1218" ], [ "0.831", "0.1111", "0.851", "0.1111", "0.851", "0.1237", "0.831", "0.1237" ], [ "0.0809", "0.1225", "0.4697", "0.1225", "0.4697", "0.1364", "0.0809", "0.1364" ], [ "0.7427", "0.1206", "0.7812", "0.1206", "0.7812", "0.1338", "0.7427", "0.1338" ], [ "0.0801", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0801", "0.1515" ], [ "0.5024", "0.1477", "0.5596", "0.1477", "0.5596", "0.1578", "0.5024", "0.1578" ], [ "0.0809", "0.1547", "0.468", "0.1547", "0.468", "0.1654", "0.0809", "0.1654" ], [ "0.505", "0.1604", "0.558", "0.1604", "0.558", "0.1674", "0.505", "0.1674" ], [ "0.0809", "0.1692", "0.4656", "0.1692", "0.4656", "0.1799", "0.0809", "0.1799" ], [ "0.6587", "0.1674", "0.6895", "0.1674", "0.6895", "0.1799", "0.6587", "0.1799" ], [ "0.0801", "0.1831", "0.468", "0.1831", "0.468", "0.1963", "0.0801", "0.1963" ], [ "0.0809", "0.1995", "0.4673", "0.1995", "0.4673", "0.2102", "0.0809", "0.2102" ], [ "0.6562", "0.1963", "0.6855", "0.1963", "0.6855", "0.2052", "0.6562", "0.2052" ], [ "0.0801", "0.2134", "0.468", "0.2134", "0.468", "0.2267", "0.0801", "0.2267" ], [ "0.634", "0.2203", "0.66", "0.2203", "0.66", "0.2317", "0.634", "0.2317" ], [ "0.0784", "0.2291", "0.468", "0.2279", "0.468", "0.2412", "0.0784", "0.2424" ], [ "0.835", "0.2299", "0.8906", "0.2299", "0.8906", "0.2405", "0.835", "0.2405" ], [ "0.664", "0.2393", "0.693", "0.2412", "0.6914", "0.2524", "0.6626", "0.2507" ], [ "0.818", "0.238", "0.8906", "0.2405", "0.8906", "0.252", "0.818", "0.25" ], [ "0.7563", "0.2405", "0.789", "0.2405", "0.789", "0.252", "0.7563", "0.252" ], [ "0.0801", "0.2437", "0.468", "0.2437", "0.468", "0.2568", "0.0801", "0.2568" ], [ "0.527", "0.2405", "0.596", "0.2405", "0.596", "0.2507", "0.527", "0.2507" ], [ "0.0801", "0.2583", "0.468", "0.2583", "0.468", "0.2715", "0.0801", "0.2715" ], [ "0.501", "0.2615", "0.8906", "0.2615", "0.8906", "0.2747", "0.501", "0.2747" ], [ "0.0801", "0.2734", "0.468", "0.2734", "0.468", "0.2874", "0.0801", "0.2874" ], [ "0.5034", "0.2766", "0.8906", "0.2766", "0.8906", "0.2874", "0.5034", "0.2874" ], [ "0.0801", "0.289", "0.468", "0.289", "0.468", "0.303", "0.0801", "0.303" ], [ "0.501", "0.2886", "0.892", "0.2886", "0.892", "0.3018", "0.501", "0.3018" ], [ "0.0801", "0.3037", "0.4697", "0.3037", "0.4697", "0.3176", "0.0801", "0.3176" ], [ "0.5024", "0.303", "0.892", "0.303", "0.892", "0.3164", "0.5024", "0.3164" ], [ "0.0801", "0.3193", "0.468", "0.3193", "0.468", "0.3333", "0.0801", "0.3333" ], [ "0.5024", "0.3176", "0.892", "0.3176", "0.892", "0.3308", "0.5024", "0.3308" ], [ "0.0801", "0.334", "0.468", "0.334", "0.468", "0.348", "0.0801", "0.348" ], [ "0.5024", "0.3308", "0.892", "0.3308", "0.892", "0.3447", "0.5024", "0.3447" ], [ "0.0809", "0.351", "0.468", "0.351", "0.468", "0.361", "0.0809", "0.361" ], [ "0.5034", "0.3467", "0.8906", "0.3467", "0.8906", "0.3567", "0.5034", "0.3567" ], [ "0.0784", "0.3643", "0.3882", "0.3643", "0.3882", "0.3774", "0.0784", "0.3774" ], [ "0.5034", "0.3599", "0.8906", "0.3599", "0.8906", "0.3706", "0.5034", "0.3706" ], [ "0.5024", "0.3726", "0.8906", "0.3726", "0.8906", "0.3865", "0.5024", "0.3865" ], [ "0.0989", "0.3801", "0.468", "0.3801", "0.468", "0.3945", "0.0989", "0.3945" ], [ "0.5024", "0.387", "0.683", "0.387", "0.683", "0.401", "0.5024", "0.401" ], [ "0.0801", "0.3945", "0.4673", "0.3945", "0.4673", "0.4077", "0.0801", "0.4077" ], [ "0.0784", "0.4092", "0.4697", "0.411", "0.4697", "0.4248", "0.0784", "0.4224" ], [ "0.0801", "0.4255", "0.468", "0.4255", "0.468", "0.44", "0.0801", "0.44" ], [ "0.501", "0.4287", "0.8906", "0.428", "0.8906", "0.4426", "0.501", "0.443" ], [ "0.0801", "0.4414", "0.4697", "0.4414", "0.4697", "0.455", "0.0801", "0.455" ], [ "0.5024", "0.4443", "0.8906", "0.4443", "0.8906", "0.4583", "0.5024", "0.4583" ], [ "0.0801", "0.4558", "0.4697", "0.4558", "0.4697", "0.4697", "0.0801", "0.4697" ], [ "0.5024", "0.459", "0.892", "0.459", "0.892", "0.473", "0.5024", "0.473" ], [ "0.0801", "0.4717", "0.4697", "0.4717", "0.4697", "0.4849", "0.0801", "0.4849" ], [ "0.5034", "0.4734", "0.8906", "0.4734", "0.8906", "0.4873", "0.5034", "0.4873" ], [ "0.0784", "0.4841", "0.286", "0.4849", "0.286", "0.5", "0.0784", "0.4988" ], [ "0.5024", "0.488", "0.8906", "0.4893", "0.8906", "0.504", "0.5024", "0.5024" ], [ "0.5024", "0.505", "0.8906", "0.505", "0.8906", "0.518", "0.5024", "0.518" ], [ "0.0801", "0.518", "0.2197", "0.518", "0.2197", "0.532", "0.0801", "0.532" ], [ "0.5034", "0.5215", "0.5474", "0.5215", "0.5474", "0.5327", "0.5034", "0.5327" ], [ "0.0801", "0.5444", "0.468", "0.5444", "0.468", "0.558", "0.0801", "0.558" ], [ "0.0801", "0.56", "0.468", "0.56", "0.468", "0.5737", "0.0801", "0.5737" ], [ "0.5034", "0.5566", "0.892", "0.5566", "0.892", "0.571", "0.5034", "0.571" ], [ "0.0784", "0.5747", "0.468", "0.5737", "0.468", "0.5884", "0.0784", "0.589" ], [ "0.5024", "0.5737", "0.892", "0.5737", "0.892", "0.587", "0.5024", "0.587" ], [ "0.0801", "0.589", "0.4673", "0.589", "0.4673", "0.603", "0.0801", "0.603" ], [ "0.5024", "0.5884", "0.892", "0.5884", "0.892", "0.6016", "0.5024", "0.6016" ], [ "0.0784", "0.605", "0.4673", "0.605", "0.4673", "0.6187", "0.0784", "0.6187" ], [ "0.5024", "0.603", "0.8906", "0.6016", "0.8906", "0.616", "0.5024", "0.617" ], [ "0.0801", "0.6206", "0.468", "0.6206", "0.468", "0.634", "0.0801", "0.634" ], [ "0.501", "0.617", "0.8906", "0.617", "0.8906", "0.631", "0.501", "0.631" ], [ "0.0809", "0.6353", "0.4697", "0.6353", "0.4697", "0.6484", "0.0809", "0.6484" ], [ "0.5024", "0.632", "0.892", "0.632", "0.892", "0.646", "0.5024", "0.646" ], [ "0.0801", "0.651", "0.4697", "0.651", "0.4697", "0.664", "0.0801", "0.664" ], [ "0.5024", "0.6484", "0.892", "0.6484", "0.892", "0.662", "0.5024", "0.662" ], [ "0.0801", "0.6665", "0.468", "0.6665", "0.468", "0.6797", "0.0801", "0.6797" ], [ "0.5024", "0.6636", "0.892", "0.6636", "0.892", "0.677", "0.5024", "0.677" ], [ "0.0801", "0.6797", "0.468", "0.6797", "0.468", "0.6934", "0.0801", "0.6934" ], [ "0.5034", "0.6797", "0.8896", "0.6797", "0.8896", "0.69", "0.5034", "0.69" ], [ "0.0801", "0.6963", "0.468", "0.6963", "0.468", "0.7104", "0.0801", "0.7104" ], [ "0.5024", "0.6934", "0.892", "0.6934", "0.892", "0.707", "0.5024", "0.707" ], [ "0.0801", "0.711", "0.4697", "0.711", "0.4697", "0.7246", "0.0801", "0.7246" ], [ "0.5024", "0.709", "0.8906", "0.709", "0.8906", "0.7227", "0.5024", "0.7227" ], [ "0.0801", "0.726", "0.468", "0.726", "0.468", "0.7393", "0.0801", "0.7393" ], [ "0.5024", "0.7236", "0.8906", "0.7236", "0.8906", "0.7373", "0.5024", "0.7373" ], [ "0.0784", "0.7393", "0.468", "0.7407", "0.468", "0.755", "0.0784", "0.754" ], [ "0.501", "0.738", "0.8906", "0.738", "0.8906", "0.752", "0.501", "0.752" ], [ "0.0784", "0.756", "0.452", "0.755", "0.452", "0.7695", "0.0784", "0.771" ], [ "0.5034", "0.755", "0.8936", "0.755", "0.8936", "0.768", "0.5034", "0.768" ], [ "0.5034", "0.7695", "0.892", "0.7695", "0.892", "0.7827", "0.5034", "0.7827" ], [ "0.5034", "0.786", "0.8896", "0.786", "0.8896", "0.797", "0.5034", "0.797" ], [ "0.0809", "0.7954", "0.468", "0.7954", "0.468", "0.81", "0.0809", "0.81" ], [ "0.0801", "0.811", "0.4697", "0.811", "0.4697", "0.8247", "0.0801", "0.8247" ], [ "0.0784", "0.8257", "0.468", "0.8257", "0.468", "0.839", "0.0784", "0.839" ], [ "0.5034", "0.83", "0.705", "0.83", "0.705", "0.8447", "0.5034", "0.8447" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.0801", "0.856", "0.4697", "0.856", "0.4697", "0.869", "0.0801", "0.869" ], [ "0.5215", "0.856", "0.892", "0.856", "0.892", "0.869", "0.5215", "0.869" ], [ "0.0801", "0.872", "0.4714", "0.872", "0.4714", "0.8853", "0.0801", "0.8853" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0801", "0.887", "0.4714", "0.887", "0.4714", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ], [ "0.4805", "0.924", "0.4927", "0.924", "0.4927", "0.936", "0.4805", "0.936" ] ]

[ "in-planes", "Patches", "NeuralRendering", "Intuitively, our method is supervised to capture the inter-", "nal statistics of image patches at various scales and generate", "", "3D scenes whose patch-based projections follow the input", "Rendemng", "image statistics. At the core of our method lies continuous-", "StyleGAN2", "scale patch-based adversarial [I4training.Ourradiance", "Generator", "fieldsare represented astriplanefeature maps[4,44l pro-", "Input", "duced by a StyleGAN2 [22] generator. We volume-render", "ourgeneratedscenesfromrandomlysampledcameraswith", "", "varying fields of view, to simulate the appearance of im-", "", "age patches at various scales. A scale-aware discrimina-", "StyleGAN2", "Images", "Discriminator", "Patches", "tor is then used to compute an adversarial loss to the real", "Optional 3D Scene", "and generated 2D patches to enforce realistic patch dis-", "Figure 2. SinGRAF pipeline. The framework takes as input a few", "tributions across all sampled views. Notably, our design", "images of a single scenefor examplerendered froma3D scan", "of continuous-scale patch-based generator and discrimina-", "or photographed (bottom). The 3D-aware generator (top left) is", "tor allows patch-level adversarial training without expen-", "then trained to generate 2D feature planes that are arranged in a", "sive hierarchical training [42, 52, 55]. During the training,", "triplane configuration and rendered into patches of varying scale", "we find applying perspective augmentations to the image", "(top right). These rendered patches along with patches cropped", "patchesand optimizingthecamera samplingdistributionto", "from the input images arethen comparedbya discriminator.Once", "be important for high-quality scene generation.", "trained,the SinGRAF generator synthesizes different realizations", "of the 3D scene that resemble the appearance of the training im-", "The resulting system is able to create plausible 3D vari-", "ages while varying the layout.", "ations of a given scene trained only from a set of unposed", "2Dimagesofthatscene.Wedemonstrateourmethodon", "two challenging indoor datasets of Replica [48] and Matter-", "apply their discriminators on full-resolution images dur-", "port3D [6] as well as a captured outdoor scene. We evaluate", "ing training, but some also employ patch-based discrimina-", "SinGRAF against the state-of-the-art 3D scene generation", "tors [28,41,47]. The resulting 3D GANs can create diverse", "methods, demonstrating its unique ability to induce realis-", "3D radiance fields that enable view-consistent NVS. Exist-", "tic and diverse3Dgenerations.", "ing 3D GANs, however, rely on a large amount of training", "data, while our model only uses a few images of a single 3D", "2. Related Work", "scene.", "Synthesis from 3D Supervision.A large body of prior", "work aims at creating variations of scenes or objects. Pro-", "Few-Shot Generative Models. Recently, researchers", "cedural modeling approaches [11,27,33,36] are widely used", "have started applying generative modeling techniques to", "for auto-generating repetitive scenes such as terrains, build-", "few-shot settings, where only a few or single examples are", "ings, or plants. These methods typically require manually", "given. In the 2D image domain, SinGAN and its exten-", "designing the rules and grammars to procedurally add new", "sions [19,42,55] explored the idea of training a CNN-based", "3D elements. Example-based methods [13, 21, 54] aim at", "hierarchical generator on a single image, supervised using", "extracting patterns from 3D asset examples to synthesize", "patch discrimination at multiple scales.The strategy of", "new models. This line of data-driven approaches learns how", "learning the internal patch distribution of a single example", "to mix and match different components to create a plausible", "totrainagenerativemodel hasbeenwidelyadoptedforvar-", "3D asset. Similarly, scene synthesis techniques [12, 38, 50]", "ious tasks, including the synthesis of videos [16], motion", "learn a distribution of plausible object arrangements from", "sequences [23], 3D textures [17, 39], or 3D shapes [18,52]", "professionally designed scene datasets. All of these meth-", "However, none of these works applied 3D generative mod-", "ods require datasets of 3D assets, part segmentation, or ob-", "els from single-scene images. Concurrently to our work,", "ject arrangement designs, which are expensive to collect.", "[51] trains generative radiance fields from single-scene im-", "ages but focuses mainly on stochastic and repetitive syn-", "thetic scenes, rather than structured, human-made scenes.", "3D-aware GANs. Leveraging the recent developments of", "neural implicit representations [8, 30, 37, 45] and radiance", "fields [2,24, 31,32, 46], 3D GANs [1,3-5,9, 10, 15,34,", "3.Single Scene3DGAN", "35,43,47,49,53,56-59] train generative 3D radiance fields", "from a set of single-view images. These methods render the", "Our system takes as input an unposed set of images taken", "sampled scenes from various viewpoints via volume render-", "from a single scene and outputs a 3D generative model G", "ing and supervise adversarially. Many of these approaches", "that can generate diverse 3D radiance fields. We assume the", "2" ]

{ "arxiv_id": "2211.17260", "language": "en", "page_idx": 1 }

[ [ "0.0801", "0.09216", "0.468", "0.09216", "0.468", "0.1061", "0.0801", "0.1061" ], [ "0.0809", "0.1092", "0.4673", "0.1092", "0.4673", "0.1193", "0.0809", "0.1193" ], [ "0.0801", "0.1237", "0.468", "0.1237", "0.468", "0.137", "0.0801", "0.137" ], [ "0.0801", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0801", "0.1515" ], [ "0.0801", "0.1528", "0.468", "0.1528", "0.468", "0.1666", "0.0801", "0.1666" ], [ "0.5312", "0.1642", "0.5786", "0.1642", "0.5786", "0.1755", "0.5312", "0.1755" ], [ "0.626", "0.1642", "0.6733", "0.1642", "0.6733", "0.1755", "0.626", "0.1755" ], [ "0.721", "0.1642", "0.7686", "0.1642", "0.7686", "0.1755", "0.721", "0.1755" ], [ "0.812", "0.1654", "0.8667", "0.1654", "0.8667", "0.1755", "0.812", "0.1755" ], [ "0.5376", "0.1881", "0.8545", "0.1881", "0.8545", "0.202", "0.5376", "0.202" ], [ "0.0801", "0.1674", "0.468", "0.1674", "0.468", "0.1812", "0.0801", "0.1812" ], [ "0.0809", "0.1901", "0.2467", "0.1901", "0.2467", "0.2052", "0.0809", "0.2052" ], [ "0.0989", "0.2153", "0.4697", "0.2153", "0.4697", "0.2291", "0.0989", "0.2291" ], [ "0.0801", "0.2291", "0.4697", "0.2299", "0.4697", "0.2443", "0.0801", "0.2437" ], [ "0.505", "0.2279", "0.8906", "0.2279", "0.8906", "0.2412", "0.505", "0.2412" ], [ "0.0801", "0.2468", "0.4656", "0.2468", "0.4656", "0.26", "0.0801", "0.26" ], [ "0.5034", "0.2437", "0.8906", "0.2437", "0.8906", "0.2568", "0.5034", "0.2568" ], [ "0.5024", "0.2583", "0.8896", "0.2583", "0.8896", "0.2715", "0.5024", "0.2715" ], [ "0.0809", "0.279", "0.468", "0.279", "0.468", "0.2922", "0.0809", "0.2922" ], [ "0.5024", "0.2734", "0.892", "0.2734", "0.892", "0.2874", "0.5024", "0.2874" ], [ "0.501", "0.2874", "0.8906", "0.2874", "0.8906", "0.3018", "0.501", "0.3018" ], [ "0.0801", "0.2942", "0.4697", "0.2942", "0.4697", "0.308", "0.0801", "0.308" ], [ "0.0801", "0.3093", "0.4697", "0.3093", "0.4697", "0.3225", "0.0801", "0.3225" ], [ "0.5024", "0.3037", "0.6787", "0.3037", "0.6787", "0.3145", "0.5024", "0.3145" ], [ "0.523", "0.318", "0.8906", "0.318", "0.8906", "0.332", "0.523", "0.332" ], [ "0.0801", "0.3245", "0.4697", "0.3245", "0.4697", "0.3384", "0.0801", "0.3384" ], [ "0.5024", "0.332", "0.892", "0.3333", "0.892", "0.348", "0.5024", "0.3467" ], [ "0.0801", "0.3384", "0.4673", "0.3384", "0.4673", "0.3528", "0.0801", "0.3528" ], [ "0.0801", "0.3547", "0.4697", "0.3547", "0.4697", "0.3687", "0.0801", "0.3687" ], [ "0.501", "0.3499", "0.892", "0.3484", "0.892", "0.3623", "0.501", "0.363" ], [ "0.0768", "0.3643", "0.4697", "0.3643", "0.4697", "0.385", "0.0768", "0.385" ], [ "0.5024", "0.3643", "0.892", "0.3643", "0.892", "0.3774", "0.5024", "0.3774" ], [ "0.0784", "0.3838", "0.4697", "0.3833", "0.4697", "0.3977", "0.0784", "0.3984" ], [ "0.5024", "0.379", "0.892", "0.379", "0.892", "0.3926", "0.5024", "0.3926" ], [ "0.0768", "0.3977", "0.4453", "0.3945", "0.4453", "0.4153", "0.0768", "0.4185" ], [ "0.5034", "0.3945", "0.8906", "0.3945", "0.8906", "0.4077", "0.5034", "0.4077" ], [ "0.501", "0.4077", "0.892", "0.4092", "0.892", "0.4236", "0.501", "0.4224" ], [ "0.0989", "0.414", "0.468", "0.4136", "0.468", "0.428", "0.0989", "0.4287" ], [ "0.5024", "0.4248", "0.8896", "0.4248", "0.8896", "0.4382", "0.5024", "0.4382" ], [ "0.0801", "0.4312", "0.4697", "0.4312", "0.4697", "0.4443", "0.0801", "0.4443" ], [ "0.0801", "0.4458", "0.4656", "0.4458", "0.4656", "0.459", "0.0801", "0.459" ], [ "0.501", "0.44", "0.8906", "0.44", "0.8906", "0.4539", "0.501", "0.4539" ], [ "0.0801", "0.459", "0.3145", "0.459", "0.3145", "0.4734", "0.0801", "0.4734" ], [ "0.5024", "0.455", "0.6445", "0.455", "0.6445", "0.4685", "0.5024", "0.4685" ], [ "0.5195", "0.4666", "0.8906", "0.464", "0.8906", "0.4817", "0.5195", "0.4841" ], [ "0.0907", "0.488", "0.468", "0.488", "0.468", "0.5024", "0.0907", "0.5024" ], [ "0.5034", "0.4841", "0.8906", "0.4841", "0.8906", "0.4976", "0.5034", "0.4976" ], [ "0.5024", "0.5", "0.8896", "0.5", "0.8896", "0.513", "0.5024", "0.513" ], [ "0.0801", "0.5166", "0.415", "0.5166", "0.415", "0.53", "0.0801", "0.53" ], [ "0.505", "0.5166", "0.8906", "0.5166", "0.8906", "0.5264", "0.505", "0.5264" ], [ "0.5024", "0.528", "0.8896", "0.53", "0.8896", "0.5444", "0.5024", "0.5425" ], [ "0.0801", "0.55", "0.468", "0.55", "0.468", "0.563", "0.0801", "0.563" ], [ "0.5034", "0.5454", "0.892", "0.5454", "0.892", "0.5596", "0.5034", "0.5596" ], [ "0.0801", "0.5645", "0.468", "0.5645", "0.468", "0.579", "0.0801", "0.579" ], [ "0.5034", "0.56", "0.892", "0.56", "0.892", "0.5737", "0.5034", "0.5737" ], [ "0.0801", "0.58", "0.468", "0.58", "0.468", "0.5947", "0.0801", "0.5947" ], [ "0.505", "0.577", "0.8896", "0.577", "0.8896", "0.587", "0.505", "0.587" ], [ "0.0801", "0.595", "0.2303", "0.595", "0.2303", "0.6094", "0.0801", "0.6094" ], [ "0.5024", "0.5903", "0.892", "0.5903", "0.892", "0.604", "0.5024", "0.604" ], [ "0.5034", "0.606", "0.892", "0.606", "0.892", "0.619", "0.5034", "0.619" ], [ "0.5024", "0.6206", "0.892", "0.619", "0.892", "0.634", "0.5024", "0.6353" ], [ "0.2336", "0.629", "0.4004", "0.629", "0.4004", "0.6465", "0.2336", "0.6465" ], [ "0.4468", "0.631", "0.4697", "0.631", "0.4697", "0.646", "0.4468", "0.646" ], [ "0.5024", "0.6353", "0.8906", "0.6353", "0.8906", "0.6484", "0.5024", "0.6484" ], [ "0.5034", "0.651", "0.8896", "0.651", "0.8896", "0.664", "0.5034", "0.664" ], [ "0.1528", "0.632", "0.1887", "0.632", "0.1887", "0.6445", "0.1528", "0.6445" ], [ "0.1487", "0.662", "0.2385", "0.6655", "0.2378", "0.6846", "0.1478", "0.681" ], [ "0.2754", "0.664", "0.3677", "0.664", "0.3677", "0.683", "0.2754", "0.683" ], [ "0.4468", "0.6685", "0.4697", "0.6685", "0.4697", "0.6816", "0.4468", "0.6816" ], [ "0.5034", "0.6665", "0.892", "0.6665", "0.892", "0.6797", "0.5034", "0.6797" ], [ "0.5034", "0.681", "0.7812", "0.681", "0.7812", "0.6943", "0.5034", "0.6943" ], [ "0.5215", "0.696", "0.892", "0.696", "0.892", "0.709", "0.5215", "0.709" ], [ "0.0809", "0.7026", "0.4673", "0.7026", "0.4673", "0.7163", "0.0809", "0.7163" ], [ "0.5024", "0.7104", "0.892", "0.711", "0.892", "0.726", "0.5024", "0.7246" ], [ "0.0809", "0.7183", "0.4673", "0.7183", "0.4673", "0.732", "0.0809", "0.732" ], [ "0.0809", "0.732", "0.4673", "0.733", "0.4673", "0.7476", "0.0809", "0.747" ], [ "0.501", "0.726", "0.8906", "0.7246", "0.8906", "0.7393", "0.501", "0.7407" ], [ "0.5034", "0.741", "0.892", "0.741", "0.892", "0.755", "0.5034", "0.755" ], [ "0.0801", "0.7476", "0.3857", "0.7476", "0.3857", "0.761", "0.0801", "0.761" ], [ "0.5024", "0.757", "0.892", "0.756", "0.892", "0.7695", "0.5024", "0.771" ], [ "0.0989", "0.763", "0.4697", "0.763", "0.4697", "0.778", "0.0989", "0.778" ], [ "0.5024", "0.7715", "0.892", "0.7715", "0.892", "0.785", "0.5024", "0.785" ], [ "0.0801", "0.779", "0.468", "0.779", "0.468", "0.7935", "0.0801", "0.7935" ], [ "0.5034", "0.787", "0.8716", "0.787", "0.8716", "0.8003", "0.5034", "0.8003" ], [ "0.0801", "0.7935", "0.4697", "0.7935", "0.4697", "0.8066", "0.0801", "0.8066" ], [ "0.524", "0.803", "0.8906", "0.803", "0.8906", "0.813", "0.524", "0.813" ], [ "0.0801", "0.8096", "0.4307", "0.8096", "0.4307", "0.8228", "0.0801", "0.8228" ], [ "0.5024", "0.8164", "0.8906", "0.8164", "0.8906", "0.83", "0.5024", "0.83" ], [ "0.0801", "0.8335", "0.2418", "0.8335", "0.2418", "0.8467", "0.0801", "0.8467" ], [ "0.501", "0.832", "0.892", "0.8306", "0.892", "0.8447", "0.501", "0.845" ], [ "0.5024", "0.8467", "0.8936", "0.845", "0.8936", "0.86", "0.5024", "0.8613" ], [ "0.0784", "0.855", "0.468", "0.856", "0.468", "0.8706", "0.0784", "0.869" ], [ "0.5034", "0.8623", "0.7754", "0.8623", "0.7754", "0.8755", "0.5034", "0.8755" ], [ "0.0784", "0.8706", "0.4697", "0.872", "0.4697", "0.886", "0.0784", "0.8853" ], [ "0.0768", "0.8853", "0.468", "0.887", "0.468", "0.901", "0.0768", "0.8984" ], [ "0.614", "0.8833", "0.778", "0.884", "0.778", "0.9014", "0.614", "0.901" ], [ "0.8677", "0.8853", "0.8945", "0.8853", "0.8945", "0.9014", "0.8677", "0.9014" ], [ "0.4795", "0.925", "0.492", "0.925", "0.492", "0.9375", "0.4795", "0.9375" ] ]

[ "intrinsic parameters of the camera are available. We do not", "assume ourtargetscenetobe static.i.e.,weallowtemporal", "changes in the scene. In the experiment section, we show", "our method's behavior for dynamic scenes. We illustrate", "our image generation and discrimination processes in Fig.2.", "Scale 1.0", "Scale0.8", "Scale0.5", "Scale0.25", "Figure 3. Visualizations of patches at different scales.", "Source code and pre-trained models will be made available.", "3.1. Rendering Model", "Our generative model G(T, z) takes a set of query rays", "F and a noise vector z and outputs RGB predictions for the", "(Sec. 4.8), adversarial training at full scale leads to a col-", "rays. Below we briefly discuss the rendering process of G.", "lapse of the learned distribution to a single mode, likely", "because the joint information of the images uniquely deter-", "Generator Architecture G represents continuous radi-", "mines a single 3D structure. Therefore, we turn to learn the", "internal patch distribution of the images, inspired by some", "ance fields using tri-planes, following [4]. We adopt a", "StyleGAN2-based generator backbone [22], which consists", "existing works[42.52.55]", "Directly extending hierarchical patch-based GANs [42,", "of a mapping network that takes as input a noise vector z ~", "52,55], however, would require expensive training of a", "IR128 and transforms it into a latent code vector w ~ IR128", "Next, a synthesis network transforms w into a 2D feature", "pyramid of generators withprogressively-growingfeature", "image F e IRNxNx3C with a total of 3C feature channels.", "plane resolutions. To address the issue, we notice that our", "Following [4], we split these feature channels into three", "tri-plane features define a continuous radiance field and thus", "axis-aligned feature planes Fruy,Fa,Fy RNxNx", "are able to render patches at an arbitrary resolution and", "scale. Therefore, we use a single generator network and", "A color c and density value can now be queried at an", "continuously control the scale of the patches during train-", "arbitrary 3D coordinate x by aggregating the triplane fea-", "tures and processing them by a small multilayer perceptron-", "ing, forgoing the progressive training of multiple generator", "style decoder, MLP: R3C R4, as", "networks [42,52,55].", "Given input images {I1, ., IN|I; E RH'xH'x3] of res-", "(c(x),α (x)) = MLP(Fry (x)+Frz (x)+Fuz (x)). (I)", "olution H'× H' and field of view (FOV) e, we consider the", "image plane P of a virtual camera with the same FOV. Dur-", "Note that we do not model view-dependent effects.", "ingtraining,wevolume-renderpatcheswithfixedresolution", "H × H but with varying scale s. Here, the scale s e [0, 1]", "Neural Rendering. Using volume rendering [26,31], we", "indicates the spatial extent of a patch on the image plane of", "project the 3D neural field into 2D images. For this purpose,", "P. When s = 1.0 the patch covers the entire image plane", "the aggregated color C(r) of a ray r E T is computed by", "with H xH pixels,whichis usedforfull-resolutiontrain-", "integrating the field as", "ing of existing 3D GANs [5, 41]. With smaller s, the patch", "will cover a smaller window in the image plane with the", "same resolution, thus containing more details. The location", "T (t) α (r(t)c(r(t) dt,", "(2)", "of the patch window on P is sampled randomly. We render", "a sampled patch p with its associated rays Ip: G(Te, z).", "C(r)", "T (t) = exp (", "α(r(s)ds)", "(3)", "Similarly, a ground truth patch can be sampled by cropping", "I,'s with a given patch scale s (see Fig. 3).", "The scale value s is sampled randomly for each patch", "where tn and tf indicate near and far bounds along the ray", "from a uniform distribution s ~ U(smin(t), smax(t)), where", "r(t) = o + td pointing from its origin o into direction d.", "The continuous volume rendering equation (Eq. (3)) is typ-", "t is the current training epoch. We schedule the scale dis-", "tribution so that in early epochs, we have larger patches to", "ically computed using the quadrature rule [26]", "provide scene structural information and gradually decrease", "Our volume rendering step directly outputs RGB color", "the scales towards the end of the training to induce better", "images or patches; we do not apply a superresolution mod-", "quality and diversity. Refer to supplementary for details.", "ule on the rendered values. Moreover, we use 96 samples", "OurdiscriminatornetworkD,whose architecture", "per ray and do not use hierarchical ray sampling [31].", "closely follows that of [22], takes as input patches and out-", "3.2. Training Process", "puts a scalar, indicating the realism of thepatches.Because", "our patches have varying scales, we additionally condition", "Progressive Patch Scaling.Given a set of input images,", "D with the scale value s of each patch p:", "most existing 3D GANs generate and discriminate images", "atfull resolution.However,as shown in theablation study", "D : G(Te,z) × $+→R.", "(4)", "3" ]

{ "arxiv_id": "2211.17260", "language": "en", "page_idx": 2 }

[ [ "0.0801", "0.09155", "0.4722", "0.09155", "0.4722", "0.1061", "0.0801", "0.1061" ], [ "0.5024", "0.09216", "0.8906", "0.09216", "0.8906", "0.1061", "0.5024", "0.1061" ], [ "0.0809", "0.1092", "0.2573", "0.1092", "0.2573", "0.1193", "0.0809", "0.1193" ], [ "0.5024", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.5024", "0.1218" ], [ "0.5024", "0.1225", "0.892", "0.1225", "0.892", "0.1364", "0.5024", "0.1364" ], [ "0.0801", "0.1395", "0.468", "0.1401", "0.468", "0.1547", "0.0801", "0.154" ], [ "0.0801", "0.154", "0.468", "0.154", "0.468", "0.1686", "0.0801", "0.1686" ], [ "0.5024", "0.1622", "0.814", "0.1622", "0.814", "0.1799", "0.5024", "0.1799" ], [ "0.0784", "0.1692", "0.468", "0.1704", "0.468", "0.1849", "0.0784", "0.1843" ], [ "0.869", "0.1642", "0.892", "0.1642", "0.892", "0.178", "0.869", "0.178" ], [ "0.0801", "0.1849", "0.4697", "0.1843", "0.4697", "0.1989", "0.0801", "0.1995" ], [ "0.539", "0.1843", "0.892", "0.1843", "0.892", "0.1989", "0.539", "0.1989" ], [ "0.0801", "0.202", "0.2695", "0.202", "0.2695", "0.2153", "0.0801", "0.2153" ], [ "0.505", "0.2122", "0.6045", "0.2122", "0.6045", "0.2235", "0.505", "0.2235" ], [ "0.597", "0.2102", "0.8906", "0.2102", "0.8906", "0.2235", "0.597", "0.2235" ], [ "0.0989", "0.2166", "0.4697", "0.2166", "0.4697", "0.2299", "0.0989", "0.2299" ], [ "0.0784", "0.2299", "0.4697", "0.2311", "0.4697", "0.2456", "0.0784", "0.2443" ], [ "0.5024", "0.226", "0.892", "0.226", "0.892", "0.2393", "0.5024", "0.2393" ], [ "0.5024", "0.2393", "0.892", "0.2405", "0.892", "0.2551", "0.5024", "0.254" ], [ "0.0784", "0.2456", "0.468", "0.2468", "0.468", "0.2615", "0.0784", "0.26" ], [ "0.5034", "0.2556", "0.796", "0.2556", "0.796", "0.2695", "0.5034", "0.2695" ], [ "0.0801", "0.262", "0.468", "0.262", "0.468", "0.276", "0.0801", "0.276" ], [ "0.0809", "0.2766", "0.4673", "0.2766", "0.4673", "0.2903", "0.0809", "0.2903" ], [ "0.5034", "0.2842", "0.6313", "0.2842", "0.6313", "0.2974", "0.5034", "0.2974" ], [ "0.0809", "0.2922", "0.468", "0.2922", "0.468", "0.3062", "0.0809", "0.3062" ], [ "0.0825", "0.3093", "0.468", "0.3093", "0.468", "0.3193", "0.0825", "0.3193" ], [ "0.523", "0.308", "0.892", "0.308", "0.892", "0.3225", "0.523", "0.3225" ], [ "0.0801", "0.3208", "0.4697", "0.3208", "0.4697", "0.3352", "0.0801", "0.3352" ], [ "0.5024", "0.3245", "0.892", "0.3245", "0.892", "0.3384", "0.5024", "0.3384" ], [ "0.0801", "0.3372", "0.468", "0.3372", "0.468", "0.351", "0.0801", "0.351" ], [ "0.5024", "0.339", "0.8906", "0.339", "0.8906", "0.3528", "0.5024", "0.3528" ], [ "0.0801", "0.3516", "0.4697", "0.3516", "0.4697", "0.3655", "0.0801", "0.3655" ], [ "0.5024", "0.3547", "0.892", "0.3547", "0.892", "0.3687", "0.5024", "0.3687" ], [ "0.0801", "0.3674", "0.4697", "0.3674", "0.4697", "0.3806", "0.0801", "0.3806" ], [ "0.5034", "0.3694", "0.892", "0.3694", "0.892", "0.3833", "0.5034", "0.3833" ], [ "0.0809", "0.3818", "0.4673", "0.3818", "0.4673", "0.3958", "0.0809", "0.3958" ], [ "0.5024", "0.385", "0.892", "0.385", "0.892", "0.3984", "0.5024", "0.3984" ], [ "0.0801", "0.3977", "0.468", "0.3977", "0.468", "0.411", "0.0801", "0.411" ], [ "0.501", "0.3997", "0.8906", "0.3984", "0.8906", "0.4136", "0.501", "0.414" ], [ "0.0784", "0.411", "0.468", "0.4124", "0.468", "0.4268", "0.0784", "0.4255" ], [ "0.5024", "0.4216", "0.603", "0.4236", "0.603", "0.4382", "0.5024", "0.4363" ], [ "0.0801", "0.428", "0.4697", "0.428", "0.4697", "0.4414", "0.0801", "0.4414" ], [ "0.0801", "0.443", "0.429", "0.443", "0.429", "0.457", "0.0801", "0.457" ], [ "0.5215", "0.4463", "0.8936", "0.4463", "0.8936", "0.461", "0.5215", "0.461" ], [ "0.505", "0.4634", "0.8906", "0.4634", "0.8906", "0.4734", "0.505", "0.4734" ], [ "0.0809", "0.476", "0.468", "0.476", "0.468", "0.4893", "0.0809", "0.4893" ], [ "0.5024", "0.4766", "0.892", "0.4766", "0.892", "0.4905", "0.5024", "0.4905" ], [ "0.0809", "0.4924", "0.4697", "0.4924", "0.4697", "0.5024", "0.0809", "0.5024" ], [ "0.501", "0.4912", "0.892", "0.4905", "0.892", "0.505", "0.501", "0.506" ], [ "0.0801", "0.506", "0.4697", "0.506", "0.4697", "0.5195", "0.0801", "0.5195" ], [ "0.5", "0.507", "0.892", "0.506", "0.892", "0.521", "0.5", "0.5215" ], [ "0.0809", "0.523", "0.4673", "0.523", "0.4673", "0.5327", "0.0809", "0.5327" ], [ "0.5024", "0.523", "0.8906", "0.523", "0.8906", "0.536", "0.5024", "0.536" ], [ "0.0825", "0.536", "0.468", "0.536", "0.468", "0.5503", "0.0825", "0.5503" ], [ "0.5024", "0.537", "0.892", "0.537", "0.892", "0.552", "0.5024", "0.552" ], [ "0.0801", "0.552", "0.468", "0.552", "0.468", "0.565", "0.0801", "0.565" ], [ "0.501", "0.552", "0.892", "0.553", "0.892", "0.5664", "0.501", "0.565" ], [ "0.0801", "0.5664", "0.4697", "0.5664", "0.4697", "0.5806", "0.0801", "0.5806" ], [ "0.5024", "0.5674", "0.8906", "0.5674", "0.8906", "0.5806", "0.5024", "0.5806" ], [ "0.0784", "0.5806", "0.468", "0.582", "0.468", "0.595", "0.0784", "0.5947" ], [ "0.5034", "0.5835", "0.888", "0.5835", "0.888", "0.5977", "0.5034", "0.5977" ], [ "0.0809", "0.5967", "0.4697", "0.5967", "0.4697", "0.6113", "0.0809", "0.6113" ], [ "0.5034", "0.5996", "0.8896", "0.5996", "0.8896", "0.61", "0.5034", "0.61" ], [ "0.0784", "0.61", "0.4697", "0.6123", "0.4697", "0.627", "0.0784", "0.625" ], [ "0.5024", "0.613", "0.892", "0.613", "0.892", "0.627", "0.5024", "0.627" ], [ "0.0801", "0.627", "0.4272", "0.627", "0.4272", "0.64", "0.0801", "0.64" ], [ "0.5024", "0.6274", "0.892", "0.6274", "0.892", "0.6416", "0.5024", "0.6416" ], [ "0.0989", "0.6416", "0.4673", "0.6416", "0.4673", "0.655", "0.0989", "0.655" ], [ "0.5024", "0.643", "0.8906", "0.643", "0.8906", "0.657", "0.5024", "0.657" ], [ "0.0801", "0.657", "0.468", "0.657", "0.468", "0.6704", "0.0801", "0.6704" ], [ "0.501", "0.657", "0.8906", "0.6577", "0.8906", "0.6724", "0.501", "0.672" ], [ "0.0784", "0.6704", "0.4697", "0.672", "0.4697", "0.686", "0.0784", "0.685" ], [ "0.5024", "0.674", "0.892", "0.674", "0.892", "0.6875", "0.5024", "0.6875" ], [ "0.0801", "0.6875", "0.4697", "0.6875", "0.4697", "0.7007", "0.0801", "0.7007" ], [ "0.5024", "0.688", "0.892", "0.688", "0.892", "0.7026", "0.5024", "0.7026" ], [ "0.0801", "0.702", "0.4673", "0.702", "0.4673", "0.7153", "0.0801", "0.7153" ], [ "0.5024", "0.704", "0.8936", "0.704", "0.8936", "0.7173", "0.5024", "0.7173" ], [ "0.0801", "0.7173", "0.468", "0.7173", "0.468", "0.731", "0.0801", "0.731" ], [ "0.5034", "0.7183", "0.8423", "0.7183", "0.8423", "0.732", "0.5034", "0.732" ], [ "0.0809", "0.7344", "0.4656", "0.7344", "0.4656", "0.744", "0.0809", "0.744" ], [ "0.0801", "0.7476", "0.4697", "0.7476", "0.4697", "0.761", "0.0801", "0.761" ], [ "0.505", "0.7437", "0.607", "0.7437", "0.607", "0.755", "0.505", "0.755" ], [ "0.0801", "0.763", "0.4697", "0.763", "0.4697", "0.7764", "0.0801", "0.7764" ], [ "0.5215", "0.7666", "0.8906", "0.7666", "0.8906", "0.78", "0.5215", "0.78" ], [ "0.0801", "0.779", "0.4697", "0.779", "0.4697", "0.7925", "0.0801", "0.7925" ], [ "0.501", "0.781", "0.8906", "0.781", "0.8906", "0.7944", "0.501", "0.7944" ], [ "0.0801", "0.7935", "0.4697", "0.7935", "0.4697", "0.8066", "0.0801", "0.8066" ], [ "0.5034", "0.797", "0.892", "0.797", "0.892", "0.81", "0.5034", "0.81" ], [ "0.0801", "0.8096", "0.4697", "0.8096", "0.4697", "0.8228", "0.0801", "0.8228" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.0801", "0.8237", "0.468", "0.8237", "0.468", "0.837", "0.0801", "0.837" ], [ "0.5024", "0.827", "0.892", "0.827", "0.892", "0.8403", "0.5024", "0.8403" ], [ "0.0801", "0.8384", "0.183", "0.8384", "0.183", "0.853", "0.0801", "0.853" ], [ "0.5024", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.5024", "0.855" ], [ "0.5024", "0.856", "0.892", "0.856", "0.892", "0.869", "0.5024", "0.869" ], [ "0.0801", "0.869", "0.468", "0.872", "0.468", "0.8853", "0.0801", "0.8833" ], [ "0.5024", "0.872", "0.8936", "0.872", "0.8936", "0.8853", "0.5024", "0.8853" ], [ "0.0801", "0.887", "0.4714", "0.887", "0.4714", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.887", "0.8936", "0.887", "0.8936", "0.901", "0.5024", "0.901" ], [ "0.482", "0.927", "0.492", "0.927", "0.492", "0.9346", "0.482", "0.9346" ] ]

[ "Implementation-wise, we simply repeat the scale value to", "losses: R1 gradient penalty [29] and discriminator recon-", "matchthepatchresolution.", "struction loss [251(R). Combining the regularizations with", "the adversarial loss, we have the final objective as follows:", "Data augmentation. Our approach aims at generating a", "3D scene from a limited set of 2D observations. As such,", "8(G,D, T) = Ez,T~T,s,p [f(D(G(Tp,z) +", "it is desirable to augment the available data during training.", "(5)", "To this end, we apply data augmentation techniques for both", "E1,s,p [f(-D(p, s)) +>1/V,D(p, s)2 +入2&r(D),", "image and camera pose data.", "where f(a) =", "- log(1 + exp(-a), and 's are balancing", "In addition to the usual image augmentation techniques", "used for 2D GANs, such as translation, cropping, and", "parameters. Note that the sampled patch p is a function of", "the sampled scale s. We minimize the objective using the", "cutout, we suggest a perspective augmentation approach,", "ADAM optimizer with learning rate of 2e-3.", "which is appropriate for generating 3D scenes with patch", "discrimination. Based on the known camera pose, we re-", "4.Experiments", "project image patches so as to imitate camera rotations fol-", "lowedbyperspectiveprojections.Camerarotationswithout", "We conduct a number of experiments to test SinGRAF's", "changing the position do not induce occlusions or parallax;", "capabilities to (i) create realistic 3D scene variations across", "thus, this approach is applicable to captured content. More-", "a diverse set of challenging indoor scenes, (i) reliably pro-", "over, because our model discriminates patches instead of", "duce view-consistent 3D representations, and (ii) handle", "full images, unknown regions after rotation and perspective", "scene dynamics. At the end of the section, we show ab-", "projection can be cropped. In practice, we start our train-", "lation studies and justify our design decisions. We refer to", "ing without rotational augmentation but gradually introduce", "supplementary for additional empirical analysis and results.", "and increase this augmentation up to 15 for later training", "4.1. Datasets", "epochs where the patch scale s decreases. Additional details", "on data augmentation are included in the supplement.", "We test and compare our approach on five scenes", "from the Replica dataset and one scene from the Mat-", "Camera Pose Distribution.To render patches from ra-", "terport3D dataset, featuring realistically scanned indoor", "diance fields weneed to sampleavirtualcamera to", "scenes.Within the Replica dataset, we choose to cover", "render from.We define the camera pose distribution", "a good range of different scene types, including offices,", "non-parametrically using a set of 1.oo0 cameras:T=", "apartments, and hotel rooms. The Replica dataset, how-", "[Ti,.., Tiooo|T; E SE(3)); each virtual camera T is ran-", "ever, only contains scanned indoor scenes that are typical", "domly chosen from T during training. We reject the sam-", "of their own categories.To stress-test our algorithm, we", "pled T when the occupancy (opacity) value at the camera", "add a large ‘castle'ballroom dataset from the Matterport3D", "center is above some threshold.The T,'s are initially sam-", "dataset that has a very different appearance from 'regular", "pled from a zero-mean 2D Gaussian on a plane with shared", "indoor scenes.For each of the above6 scenes, we ren-", "heights with random rotation about the vertical axis. During", "der 100 views by sampling camera locations from a zero-", "training, we jitter the T 's on translation and rotation.", "mean Gaussian distribution with random rotations about the", "The randomly initialized camera distribution, however.", "height axis. We reject the cameras that collide with occu-", "may not be a good representation of the real camera distri-", "pied volumes. Lastly, we showcase our method on a cap-", "bution of the input images. While our method can generate", "tured image dataset where we use our own photographs of", "reasonable 3D scenes when training with random distribu-", "an outdoor scene. We capture the images using a hand-held", "tions, we find it beneficial to optimize the camera distribu-", "consumer-grade smartphone camera and pass them to our", "tion T for higher-quality outputs. Such optimization will", "algorithm without intrinsic or extrinsic calibrations.", "transformthedistributiontoaccommodateposeswithvari-", "ous rotations. Therefore, we optimize the T's via automatic", "4.2. Baselines", "differentiation using our adversarial loss in the early stage", "As a baseline, we compare SinGRAF against the cur-", "of the training.We eventually stop the optimization since", "rent state-of-the-art 3D scene generative method, GSN", "smaller patches at the later stage contain weak pose infor-", "[10]. While numerous 3D-GAN methods were proposed,", "mation.In practice, we decompose the poses into more", "we could not find any other suitable baseline that demon-", "easily optimizable forms, which we discuss further in the", "strated the modeling of apartment-scale scenes, such as", "supplementary.", "the ones from the Replica dataset.Concurrently to our", "work, GAUDI [3] extended GSN to model even larger-scale", "Training ObjectiveBesides the regular min-max adver-", "scenes with generative modeling, but the code is not pub-", "sarial loss [14], we find it useful to adopt two regularization", "licly available for us to run. To compare against GSN, we", "4" ]

{ "arxiv_id": "2211.17260", "language": "en", "page_idx": 3 }

[ [ "0.4968", "0.2393", "0.5063", "0.2393", "0.5063", "0.2798", "0.4968", "0.2798" ], [ "0.1626", "0.4539", "0.3938", "0.4539", "0.3938", "0.464", "0.1626", "0.464" ], [ "0.5645", "0.4558", "0.615", "0.4558", "0.615", "0.4634", "0.5645", "0.4634" ], [ "0.6665", "0.4539", "0.819", "0.4539", "0.819", "0.464", "0.6665", "0.464" ], [ "0.1422", "0.83", "0.415", "0.83", "0.415", "0.8403", "0.1422", "0.8403" ], [ "0.5645", "0.83", "0.8203", "0.83", "0.8203", "0.8403", "0.5645", "0.8403" ], [ "0.0825", "0.855", "0.2942", "0.855", "0.2942", "0.8657", "0.0825", "0.8657" ], [ "0.3506", "0.857", "0.4084", "0.857", "0.4084", "0.8643", "0.3506", "0.8643" ], [ "0.4182", "0.8594", "0.4355", "0.86", "0.4355", "0.863", "0.4182", "0.8623" ], [ "0.4436", "0.8594", "0.4714", "0.8594", "0.4714", "0.863", "0.4436", "0.863" ], [ "0.4673", "0.856", "0.888", "0.856", "0.888", "0.866", "0.4673", "0.866" ], [ "0.0801", "0.866", "0.892", "0.867", "0.892", "0.881", "0.0801", "0.88" ], [ "0.2974", "0.86", "0.3186", "0.86", "0.3186", "0.863", "0.2974", "0.863" ], [ "0.0801", "0.882", "0.8906", "0.882", "0.8906", "0.896", "0.0801", "0.896" ] ]

[ "", "(a)\"hotel_o\"scenefromtheReplicadataset", "(b)“aparti", "cenefromtheReplicadataset", "(c)\"frl_apartment_4\"scenefromtheReplicadataset", "(d)\"castle\"scenefromtheMatterport3Ddataset.", "Figure4.Resultsoffourdifference", "indicated", "", "", "example,weshow someoftheinputimages,theresultachievedbythe", "GSN [1o] baseline, and three realizations of SinGRAF's results for the same input images. Note that GSN is mode-collapsed and not able", "", "to generate different realizations of this scene while SinGRAF is capable of generating a diverse set of realizations with high image quality." ]

{ "arxiv_id": "2211.17260", "language": "en", "page_idx": 4 }

[ [ "0.1721", "0.114", "0.8286", "0.1134", "0.8286", "0.1312", "0.1721", "0.1318" ], [ "0.1368", "0.1372", "0.864", "0.1383", "0.864", "0.1561", "0.1368", "0.155" ], [ "0.3242", "0.1675", "0.681", "0.1687", "0.681", "0.1859", "0.3242", "0.1853" ], [ "0.2761", "0.1859", "0.7446", "0.1859", "0.7446", "0.2037", "0.2761", "0.2037" ], [ "0.3704", "0.2031", "0.635", "0.2037", "0.635", "0.2185", "0.3704", "0.2173" ], [ "0.1763", "0.2214", "0.827", "0.2214", "0.827", "0.2363", "0.1763", "0.2363" ], [ "0.2644", "0.2708", "0.3435", "0.2708", "0.3435", "0.2852", "0.2644", "0.2852" ], [ "0.147", "0.2922", "0.4592", "0.2922", "0.4592", "0.3071", "0.147", "0.3071" ], [ "0.02184", "0.3071", "0.068", "0.3071", "0.068", "0.4395", "0.02184", "0.4395" ], [ "0.147", "0.3071", "0.461", "0.3071", "0.461", "0.3206", "0.147", "0.3206" ], [ "0.147", "0.3218", "0.461", "0.3218", "0.461", "0.335", "0.147", "0.335" ], [ "0.1486", "0.3374", "0.4592", "0.3374", "0.4592", "0.3474", "0.1486", "0.3474" ], [ "0.511", "0.3403", "0.882", "0.3403", "0.882", "0.354", "0.511", "0.354" ], [ "0.147", "0.3486", "0.4592", "0.3486", "0.4592", "0.3623", "0.147", "0.3623" ], [ "0.5146", "0.3582", "0.8833", "0.3582", "0.8833", "0.3682", "0.5146", "0.3682" ], [ "0.1486", "0.364", "0.461", "0.364", "0.461", "0.3777", "0.1486", "0.3777" ], [ "0.511", "0.3706", "0.885", "0.3694", "0.885", "0.383", "0.511", "0.3843" ], [ "0.1503", "0.3801", "0.4592", "0.3801", "0.4592", "0.3901", "0.1503", "0.3901" ], [ "0.513", "0.3855", "0.8867", "0.3855", "0.8867", "0.3984", "0.513", "0.3984" ], [ "0.147", "0.3926", "0.461", "0.3926", "0.461", "0.4062", "0.147", "0.4062" ], [ "0.513", "0.3997", "0.739", "0.3997", "0.739", "0.4133", "0.513", "0.4133" ], [ "0.147", "0.4067", "0.461", "0.4067", "0.461", "0.4204", "0.147", "0.4204" ], [ "0.147", "0.4216", "0.4626", "0.4216", "0.4626", "0.4353", "0.147", "0.4353" ], [ "0.02016", "0.4358", "0.0638", "0.4353", "0.06635", "0.517", "0.02351", "0.5186" ], [ "0.147", "0.4353", "0.461", "0.4353", "0.461", "0.449", "0.147", "0.449" ], [ "0.513", "0.4424", "0.882", "0.4424", "0.882", "0.456", "0.513", "0.456" ], [ "0.147", "0.4495", "0.4592", "0.4495", "0.4592", "0.4631", "0.147", "0.4631" ], [ "0.1486", "0.4631", "0.461", "0.4631", "0.461", "0.4768", "0.1486", "0.4768" ], [ "0.513", "0.459", "0.8833", "0.459", "0.8833", "0.4727", "0.513", "0.4727" ], [ "0.1486", "0.4785", "0.4592", "0.4785", "0.4592", "0.4893", "0.1486", "0.4893" ], [ "0.513", "0.4744", "0.882", "0.4744", "0.882", "0.488", "0.513", "0.488" ], [ "0.1486", "0.4934", "0.4592", "0.4934", "0.4592", "0.5044", "0.1486", "0.5044" ], [ "0.513", "0.4912", "0.885", "0.4912", "0.885", "0.505", "0.513", "0.505" ], [ "0.147", "0.506", "0.461", "0.506", "0.461", "0.5195", "0.147", "0.5195" ], [ "0.5146", "0.5073", "0.8867", "0.5073", "0.8867", "0.521", "0.5146", "0.521" ], [ "0.02351", "0.5195", "0.05878", "0.5195", "0.05878", "0.728", "0.02351", "0.728" ], [ "0.147", "0.521", "0.461", "0.521", "0.461", "0.5337", "0.147", "0.5337" ], [ "0.5146", "0.524", "0.885", "0.524", "0.885", "0.5376", "0.5146", "0.5376" ], [ "0.147", "0.5337", "0.461", "0.5337", "0.461", "0.5474", "0.147", "0.5474" ], [ "0.5146", "0.539", "0.885", "0.539", "0.885", "0.5527", "0.5146", "0.5527" ], [ "0.147", "0.549", "0.461", "0.549", "0.461", "0.5625", "0.147", "0.5625" ], [ "0.511", "0.5547", "0.8833", "0.5547", "0.8833", "0.5684", "0.511", "0.5684" ], [ "0.1486", "0.564", "0.4592", "0.564", "0.4592", "0.5747", "0.1486", "0.5747" ], [ "0.147", "0.5767", "0.461", "0.5767", "0.461", "0.5903", "0.147", "0.5903" ], [ "0.513", "0.572", "0.8037", "0.572", "0.8037", "0.585", "0.513", "0.585" ], [ "0.147", "0.5913", "0.461", "0.5913", "0.461", "0.605", "0.147", "0.605" ], [ "0.531", "0.589", "0.8833", "0.589", "0.8833", "0.6025", "0.531", "0.6025" ], [ "0.147", "0.605", "0.2183", "0.605", "0.2183", "0.619", "0.147", "0.619" ], [ "0.511", "0.605", "0.885", "0.604", "0.885", "0.618", "0.511", "0.619" ], [ "0.1377", "0.627", "0.2603", "0.627", "0.2603", "0.641", "0.1377", "0.641" ], [ "0.513", "0.6216", "0.885", "0.6216", "0.885", "0.6353", "0.513", "0.6353" ], [ "0.12427", "0.632", "0.1301", "0.632", "0.1301", "0.637", "0.12427", "0.637" ], [ "0.511", "0.637", "0.8867", "0.6357", "0.8867", "0.651", "0.511", "0.652" ], [ "0.12006", "0.652", "0.4902", "0.652", "0.4902", "0.665", "0.12006", "0.665" ], [ "0.513", "0.654", "0.8867", "0.654", "0.8867", "0.6675", "0.513", "0.6675" ], [ "0.12006", "0.6685", "0.4902", "0.6685", "0.4902", "0.6816", "0.12006", "0.6816" ], [ "0.513", "0.6704", "0.8833", "0.6704", "0.8833", "0.681", "0.513", "0.681" ], [ "0.1184", "0.684", "0.4902", "0.684", "0.4902", "0.6978", "0.1184", "0.6978" ], [ "0.513", "0.686", "0.8833", "0.686", "0.8833", "0.6997", "0.513", "0.6997" ], [ "0.12006", "0.7007", "0.4902", "0.7007", "0.4902", "0.7144", "0.12006", "0.7144" ], [ "0.5146", "0.7017", "0.885", "0.7017", "0.885", "0.7153", "0.5146", "0.7153" ], [ "0.12177", "0.7163", "0.492", "0.7163", "0.492", "0.73", "0.12177", "0.73" ], [ "0.513", "0.7188", "0.885", "0.7188", "0.885", "0.7324", "0.513", "0.7324" ], [ "0.12006", "0.7324", "0.4902", "0.7324", "0.4902", "0.745", "0.12006", "0.745" ], [ "0.511", "0.733", "0.882", "0.733", "0.882", "0.7466", "0.511", "0.7466" ], [ "0.1184", "0.7466", "0.4902", "0.7476", "0.4902", "0.7627", "0.1184", "0.761" ], [ "0.511", "0.749", "0.882", "0.749", "0.882", "0.7627", "0.511", "0.7627" ], [ "0.1184", "0.763", "0.4937", "0.764", "0.4937", "0.7793", "0.1184", "0.778" ], [ "0.513", "0.7666", "0.8833", "0.7666", "0.8833", "0.78", "0.513", "0.78" ], [ "0.1184", "0.7793", "0.4895", "0.78", "0.4895", "0.7944", "0.1184", "0.7935" ], [ "0.511", "0.782", "0.885", "0.781", "0.885", "0.796", "0.511", "0.7964" ], [ "0.1184", "0.7974", "0.492", "0.7974", "0.492", "0.811", "0.1184", "0.811" ], [ "0.513", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.8135", "0.492", "0.8135", "0.492", "0.8267", "0.1184", "0.8267" ], [ "0.513", "0.814", "0.885", "0.814", "0.885", "0.8276", "0.513", "0.8276" ], [ "0.1184", "0.83", "0.4895", "0.83", "0.4895", "0.8433", "0.1184", "0.8433" ], [ "0.513", "0.8306", "0.885", "0.8306", "0.885", "0.844", "0.513", "0.844" ], [ "0.1184", "0.8457", "0.4937", "0.8457", "0.4937", "0.8594", "0.1184", "0.8594" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.11755", "0.86", "0.4902", "0.861", "0.4902", "0.876", "0.11755", "0.8745" ], [ "0.511", "0.861", "0.8833", "0.8633", "0.8833", "0.8774", "0.511", "0.876" ], [ "0.147", "0.882", "0.2678", "0.882", "0.2678", "0.8955", "0.147", "0.8955" ], [ "0.513", "0.879", "0.885", "0.879", "0.885", "0.8926", "0.513", "0.8926" ], [ "0.1486", "0.8955", "0.4534", "0.8955", "0.4534", "0.909", "0.1486", "0.909" ], [ "0.513", "0.8955", "0.885", "0.8955", "0.885", "0.909", "0.513", "0.909" ], [ "0.4937", "0.931", "0.509", "0.931", "0.509", "0.9424", "0.4937", "0.9424" ] ]

[ "Template-Based Question Generation from Retrieved", "Sentences for Improved Unsupervised Question Answering", "Alexander R. Fabbri*+ 1,2 Patrick Ng*'", "Zhiguo Wang*Ramesh Nallapati**Bing Xiang\"", "Yale University AWS AI Labs", "alexander.fabbri@yale.edu, (patricng, zhiguow, rnallapa, bxiang} @ amazon.com", "Abstract", "Question Answering (QA) is in increasing", "", "demand as the amount of information avail-", "able online and the desire for quick access", "to this contentgrows.A common approach", "Figure 1: Question Generation Pipeline: the original", "to QA has been to fine-tune a pretrained lan-", "context sentence containing a given answer is used as", "guage model on a task-specific labeled dataset.", "a query to retrieve a related sentence containing match-", "This paradigm,however,relies on scarce,and", "ing entities, which is input into our question-style con-", "costly to obtain, large-scale human-labeled", "verter to create QA training data.", "data. We propose an unsupervised approach", "to training QA models with generated pseudo-", "[T's]", "training data. We show that generating ques-", "desirable. This problem may be approached with", "tions for QA training by applying a simple", "template on a related, retrieved sentence rather", "domain adaptation or transfer learning techniques", "than the original context sentence improves", "(Chung et al., 2018) as well as data augmentation", "downstreamQAperformancebyallowingthe", "(Yang et al., 2017; Dhingra et al., 2018; Wang et al.", "model to learn more complex context-question", "2018; Alberti et al., 2019). However, here we ex-", "00", "relationships. Training a QA model on this", "pand upon the recently introduced task of unsuper-", "data gives a relative improvement over a pre-", "vised question answering (Lewis et al., 2019) to", "vious unsupervised model in F1 score on the", "examine the extent to which synthetic training data", "SQuAD datasetbyabout 14%,and20%when", "the answer is a named entity, achieving state-", "alone can be used to train a QA model.", "of-the-art performance on SQuAD for unsuper", "In particular, we focus on the machine reading", "vised QA.", "comprehension setting in which the context is a", "Introduction", "given paragraph, and the QA model can only access", "", "this paragraph to answer a question. Furthermore,", "Question Answering aims to answer a question", "we work on extractive QA, where the answer is as-", "based on a given knowledge source. Recent ad-", "sumed to bea contiguous sub-string of the context", "vances have driven the performance of QA sys-", "A training instance for supervised reading compre-", "tems to above or near-human performance on", "hension consists of three components: a question,", "QA datasets such as SQuAD (Rajpurkar et al.,", "a context, and an answer. For a given dataset do-", "2016) and Natural Questions (Kwiatkowski et al.,", "main, a collection of documents can usually be", "2019) thanks to pretrained language models such", " easily obtained, providing context in the form of", "as BERT (Devlin et al., 2019), XLNet (Yang et al.,", "paragraphs or sets of sentences. Answers can be", "2019) and RoBERTa (Liu et al., 2019). Fine-tuning", "gathered from keywords and phrases from the con-", "these language models, however, requires large-", "text. We focus mainly on factoid QA; the question", "scale data for fine-tuning. Creating a dataset for ev-", "concerns a concise fact. In particular, we empha-", "ery new domain is extremely costly and practically", "size questions whose answers are named entities,", "infeasible. The ability to apply QA models on out-", "the majority type of factoid questions. Entities can", "of-domain data in an efficient manner is thus very", "be extracted from text using named entity recog", "'Equal contribution", "nition (NER) techniques as the training instance's", "2 Work done during internship at the AWS AI Labs", "answer. Thus, the main challenge, and the focus", "1" ]

{ "arxiv_id": "2004.11892", "language": "en", "page_idx": 0 }

[ [ "0.12006", "0.0766", "0.4895", "0.0766", "0.4895", "0.0903", "0.12006", "0.0903" ], [ "0.11755", "0.0932", "0.4844", "0.0932", "0.4844", "0.1069", "0.11755", "0.1069" ], [ "0.529", "0.0932", "0.553", "0.0932", "0.553", "0.0974", "0.529", "0.0974" ], [ "0.1368", "0.10925", "0.4895", "0.10925", "0.4895", "0.1223", "0.1368", "0.1223" ], [ "0.1184", "0.1259", "0.4895", "0.1259", "0.4895", "0.1395", "0.1184", "0.1395" ], [ "0.11755", "0.1401", "0.4895", "0.1414", "0.4895", "0.1561", "0.11755", "0.155" ], [ "0.5283", "0.1354", "0.629", "0.1395", "0.629", "0.1467", "0.5283", "0.1426" ], [ "0.1184", "0.1567", "0.4878", "0.1567", "0.4878", "0.1704", "0.1184", "0.1704" ], [ "0.561", "0.1567", "0.79", "0.158", "0.79", "0.165", "0.561", "0.1644" ], [ "0.797", "0.152", "0.8555", "0.152", "0.8555", "0.1592", "0.797", "0.1592" ], [ "0.531", "0.1592", "0.553", "0.1609", "0.553", "0.1644", "0.531", "0.1621" ], [ "0.6465", "0.1704", "0.7173", "0.1704", "0.7173", "0.1812", "0.6465", "0.1812" ], [ "0.791", "0.1716", "0.854", "0.1716", "0.854", "0.1787", "0.791", "0.1787" ], [ "0.6206", "0.1758", "0.6484", "0.1758", "0.6484", "0.1787", "0.6206", "0.1787" ], [ "0.1184", "0.1733", "0.4895", "0.1733", "0.4895", "0.187", "0.1184", "0.187" ], [ "0.5996", "0.1776", "0.684", "0.1776", "0.684", "0.1818", "0.5996", "0.1818" ], [ "0.706", "0.1799", "0.784", "0.1799", "0.784", "0.1841", "0.706", "0.1841" ], [ "0.11755", "0.1895", "0.4895", "0.1895", "0.4895", "0.2031", "0.11755", "0.2031" ], [ "0.5977", "0.1882", "0.854", "0.1882", "0.854", "0.209", "0.5977", "0.209" ], [ "0.1184", "0.206", "0.4902", "0.206", "0.4902", "0.2197", "0.1184", "0.2197" ], [ "0.592", "0.2114", "0.733", "0.2114", "0.733", "0.2185", "0.592", "0.2185" ], [ "0.748", "0.2131", "0.8003", "0.2131", "0.8003", "0.2173", "0.748", "0.2173" ], [ "0.831", "0.2114", "0.854", "0.2114", "0.854", "0.2185", "0.831", "0.2185" ], [ "0.5425", "0.2185", "0.8345", "0.2185", "0.8345", "0.2257", "0.5425", "0.2257" ], [ "0.552", "0.2144", "0.592", "0.2144", "0.592", "0.2173", "0.552", "0.2173" ], [ "0.11755", "0.2214", "0.4902", "0.2203", "0.4902", "0.2351", "0.11755", "0.2363" ], [ "0.1184", "0.2382", "0.4895", "0.2382", "0.4895", "0.2517", "0.1184", "0.2517" ], [ "0.1184", "0.2542", "0.4895", "0.2542", "0.4895", "0.2673", "0.1184", "0.2673" ], [ "0.513", "0.2505", "0.8833", "0.2505", "0.8833", "0.2642", "0.513", "0.2642" ], [ "0.1184", "0.2708", "0.4902", "0.2708", "0.4902", "0.284", "0.1184", "0.284" ], [ "0.513", "0.2654", "0.885", "0.2654", "0.885", "0.279", "0.513", "0.279" ], [ "0.513", "0.2798", "0.6885", "0.2798", "0.6885", "0.2935", "0.513", "0.2935" ], [ "0.1184", "0.2861", "0.4902", "0.2874", "0.4902", "0.3005", "0.1184", "0.2998" ], [ "0.1184", "0.3018", "0.4895", "0.3018", "0.4895", "0.3154", "0.1184", "0.3154" ], [ "0.12006", "0.3184", "0.4785", "0.3184", "0.4785", "0.332", "0.12006", "0.332" ], [ "0.513", "0.3206", "0.885", "0.3206", "0.885", "0.335", "0.513", "0.335" ], [ "0.1368", "0.335", "0.4902", "0.335", "0.4902", "0.3486", "0.1368", "0.3486" ], [ "0.511", "0.3374", "0.885", "0.3374", "0.885", "0.3508", "0.511", "0.3508" ], [ "0.1184", "0.3516", "0.486", "0.3516", "0.486", "0.3623", "0.1184", "0.3623" ], [ "0.513", "0.354", "0.885", "0.354", "0.885", "0.3677", "0.513", "0.3677" ], [ "0.12006", "0.3682", "0.4895", "0.3682", "0.4895", "0.379", "0.12006", "0.379" ], [ "0.513", "0.3694", "0.8867", "0.3694", "0.8867", "0.383", "0.513", "0.383" ], [ "0.1184", "0.383", "0.4902", "0.383", "0.4902", "0.3967", "0.1184", "0.3967" ], [ "0.5146", "0.3872", "0.8833", "0.3872", "0.8833", "0.398", "0.5146", "0.398" ], [ "0.12006", "0.401", "0.4895", "0.401", "0.4895", "0.411", "0.12006", "0.411" ], [ "0.511", "0.401", "0.8833", "0.402", "0.8833", "0.415", "0.511", "0.4146" ], [ "0.1184", "0.415", "0.4902", "0.415", "0.4902", "0.4287", "0.1184", "0.4287" ], [ "0.513", "0.4187", "0.885", "0.4187", "0.885", "0.4316", "0.513", "0.4316" ], [ "0.1184", "0.4316", "0.4895", "0.4316", "0.4895", "0.4453", "0.1184", "0.4453" ], [ "0.5146", "0.434", "0.8833", "0.434", "0.8833", "0.4478", "0.5146", "0.4478" ], [ "0.11584", "0.4478", "0.2957", "0.4465", "0.2957", "0.4614", "0.11584", "0.462" ], [ "0.511", "0.4507", "0.8833", "0.4495", "0.8833", "0.4631", "0.511", "0.4644" ], [ "0.513", "0.466", "0.885", "0.466", "0.885", "0.4797", "0.513", "0.4797" ], [ "0.12006", "0.4768", "0.3938", "0.4768", "0.3938", "0.4912", "0.12006", "0.4912" ], [ "0.5146", "0.484", "0.8833", "0.484", "0.8833", "0.4946", "0.5146", "0.4946" ], [ "0.1184", "0.5044", "0.4902", "0.5044", "0.4902", "0.517", "0.1184", "0.517" ], [ "0.513", "0.4988", "0.885", "0.4988", "0.885", "0.5127", "0.513", "0.5127" ], [ "0.12006", "0.5195", "0.4895", "0.5195", "0.4895", "0.533", "0.12006", "0.533" ], [ "0.5146", "0.514", "0.885", "0.514", "0.885", "0.528", "0.5146", "0.528" ], [ "0.1184", "0.536", "0.4685", "0.536", "0.4685", "0.55", "0.1184", "0.55" ], [ "0.513", "0.531", "0.852", "0.531", "0.852", "0.5444", "0.513", "0.5444" ], [ "0.1351", "0.5503", "0.4895", "0.552", "0.4895", "0.5664", "0.1351", "0.5654" ], [ "0.532", "0.5464", "0.8833", "0.5464", "0.8833", "0.561", "0.532", "0.561" ], [ "0.1184", "0.5684", "0.4895", "0.5684", "0.4895", "0.582", "0.1184", "0.582" ], [ "0.513", "0.564", "0.885", "0.564", "0.885", "0.5776", "0.513", "0.5776" ], [ "0.1184", "0.585", "0.4895", "0.585", "0.4895", "0.5986", "0.1184", "0.5986" ], [ "0.513", "0.58", "0.885", "0.58", "0.885", "0.5933", "0.513", "0.5933" ], [ "0.1184", "0.5996", "0.4902", "0.5996", "0.4902", "0.6133", "0.1184", "0.6133" ], [ "0.511", "0.594", "0.8867", "0.5957", "0.8867", "0.61", "0.511", "0.6094" ], [ "0.513", "0.611", "0.885", "0.611", "0.885", "0.6245", "0.513", "0.6245" ], [ "0.1184", "0.6177", "0.4895", "0.616", "0.4895", "0.63", "0.1184", "0.6313" ], [ "0.12006", "0.633", "0.4895", "0.633", "0.4895", "0.6465", "0.12006", "0.6465" ], [ "0.513", "0.6274", "0.885", "0.6274", "0.885", "0.641", "0.513", "0.641" ], [ "0.5205", "0.6455", "0.5283", "0.651", "0.5225", "0.655", "0.5146", "0.6494" ], [ "0.546", "0.6436", "0.885", "0.6436", "0.885", "0.6567", "0.546", "0.6567" ], [ "0.12006", "0.6494", "0.4895", "0.6494", "0.4895", "0.663", "0.12006", "0.663" ], [ "0.11755", "0.6646", "0.4895", "0.6646", "0.4895", "0.6777", "0.11755", "0.6777" ], [ "0.513", "0.66", "0.882", "0.66", "0.882", "0.6733", "0.513", "0.6733" ], [ "0.5146", "0.676", "0.885", "0.676", "0.885", "0.6895", "0.5146", "0.6895" ], [ "0.1184", "0.6816", "0.4902", "0.6816", "0.4902", "0.6953", "0.1184", "0.6953" ], [ "0.1184", "0.697", "0.4902", "0.697", "0.4902", "0.7104", "0.1184", "0.7104" ], [ "0.513", "0.6924", "0.885", "0.6924", "0.885", "0.706", "0.513", "0.706" ], [ "0.511", "0.707", "0.885", "0.708", "0.885", "0.7227", "0.511", "0.7217" ], [ "0.11584", "0.7134", "0.4902", "0.712", "0.4902", "0.727", "0.11584", "0.728" ], [ "0.1184", "0.73", "0.4895", "0.73", "0.4895", "0.7437", "0.1184", "0.7437" ], [ "0.511", "0.7246", "0.885", "0.7246", "0.885", "0.7383", "0.511", "0.7383" ], [ "0.1184", "0.745", "0.4902", "0.745", "0.4902", "0.759", "0.1184", "0.759" ], [ "0.513", "0.741", "0.885", "0.741", "0.885", "0.755", "0.513", "0.755" ], [ "0.1184", "0.761", "0.492", "0.761", "0.492", "0.775", "0.1184", "0.775" ], [ "0.5146", "0.7573", "0.885", "0.7573", "0.885", "0.771", "0.5146", "0.771" ], [ "0.1184", "0.778", "0.4902", "0.777", "0.4902", "0.7915", "0.1184", "0.792" ], [ "0.513", "0.774", "0.885", "0.774", "0.885", "0.7876", "0.513", "0.7876" ], [ "0.511", "0.788", "0.8833", "0.788", "0.8833", "0.802", "0.511", "0.802" ], [ "0.1184", "0.7944", "0.4902", "0.7944", "0.4902", "0.808", "0.1184", "0.808" ], [ "0.11755", "0.81", "0.4895", "0.81", "0.4895", "0.8237", "0.11755", "0.8237" ], [ "0.513", "0.8047", "0.8867", "0.8047", "0.8867", "0.8184", "0.513", "0.8184" ], [ "0.1184", "0.826", "0.4895", "0.826", "0.4895", "0.8403", "0.1184", "0.8403" ], [ "0.511", "0.822", "0.8833", "0.821", "0.8833", "0.8345", "0.511", "0.835" ], [ "0.1184", "0.843", "0.4902", "0.843", "0.4902", "0.8555", "0.1184", "0.8555" ], [ "0.513", "0.8374", "0.8833", "0.8374", "0.8833", "0.851", "0.513", "0.851" ], [ "0.5146", "0.8525", "0.885", "0.8525", "0.885", "0.866", "0.5146", "0.866" ], [ "0.11755", "0.8594", "0.4895", "0.8594", "0.4895", "0.8726", "0.11755", "0.8726" ], [ "0.5146", "0.869", "0.8833", "0.869", "0.8833", "0.883", "0.5146", "0.883" ], [ "0.1377", "0.882", "0.3938", "0.883", "0.3938", "0.898", "0.1377", "0.8965" ], [ "0.1184", "0.8965", "0.2603", "0.8965", "0.2603", "0.91", "0.1184", "0.91" ], [ "0.532", "0.8936", "0.69", "0.8955", "0.69", "0.91", "0.532", "0.9077" ] ]

[ "of this paper, is creating a relevant question from a", "(context, answer) pair in an unsupervised manner.", "Contes", "Recent work of (Lewis et al., 2019) uses style", "transfer for generating questions for (context, an-", "swer) pairs but shows little improvement over ap-", "Quedoliveredhis\"HouseDivi", "plying a much simpler question generator which", "HStates infrontoftheOld StateCapitolbuildinginSpringfield.Ilin", "yforPresidento", "-let", "10,2007,[MASK]", "cyforPresident of", "EDruSLY", "drops, permutates and masks words. We improve", "", "IdinginSnrinefield.Ili", "upon this paper by proposing a simple, intuitive, re-", "", "trieval and template-based question generation", "ion Wh + B + A + 7: Who announced his c", "dacvforPreside", "United", "es in frontofthe Old State Capitol building in Springfield, Ilinois,onFebruary10,200", "ateooues", "approach, illustrated in Figure 1. The idea is to", "retrieve a sentence from the corpus similar to the", "current context, and then generate a question based", "Figure 2: Example of synthetically generated ques-", "on that sentence. Having created a question for", "tions using generic cloze-style questions as well as a", "template-based approach.", "all (context, answer) pairs, we then fine-tune a pre-", "trained BERT model on this data and evaluate on", "the SQuAD v1.1 dataset (Rajpurkar et al., 2016)", "make use of the spaCy? NER pipeline. Then, for", "Our contributions are as follows: we intro-", "a given context-answer pair, we query the index,", "duce a retrieval.template-based framework which", "using the original context sentence as a query, to", "achievesstate-of-the-artresultsonSOuADforun-", "return a sentence which (1) contains the answer,", "supervised models, particularly when the answer", "(2)doesnotcomefromthecontext,and(3)hasa", "isanamed entity.Weperform ablation studies", "lowerthan95%F1 score with thequery sentence", "to determine the effect of components in template", "to discard highly similar or plagiarized sentences.", "question generation. We are releasing our synthetic", "Besides ensuring that the retrieved sentence and", "training data and code.1", "query sentence share the answer entity, we require", "that at least one additional matching entity appears", "2Unsupervised QA Approach", "inboththequery sentenceand inthe entire context.", "We focus on creating high-quality, non-trivial ques-", "and we perform ablation studies on the effect of", "tions which will allow the model to learn to extract", "this matching below. These retrieved sentences are", "the proper answer from a context-question pair.", "then fed into our question-generation module.", "Sentence Retrieval: A standard cloze question", "Template-based Question Generation: We", "can be obtained by taking the original sentence", "consider several question styles (1) generic cloze-", "in which the answer appears from the context and", "style questions where the answer is replaced", "masking the answer with a chosen token. How-", "by the token “\"[MASK]\", (2) templated ques-", "tion “Wh+B+A+?\" as well as variations on the", "ever, a model trained on this data will only learn", "text matching and how to fill-in-the-blank, with", "ordering of this template, as shown in Figure", "n", "Given the retrieved sentence in the form", "little generalizability. For this reason, we chose", "to use a retrieval-based approach to obtain a sen-", "of [Fragment A][Answer] [Fragment", "B], the templated question\"Wh+B+A+?\"replaces", "tence similar to that which contains the answer,", "upon which to create a given question. For our", "the answer with a Wh-component (e.g., what, who,", "where), which depends on the entity type of the", "experiments, we focused on answers which are", "named entities, which has proven to be a useful", "answer and places the Wh-component at the be-", "prior assumption for downstream QA performance", "ginning of the question, followed by sentence", "(Lewis et al., 2019) confirmed by our initial ex-", "Fragment B and Fragment A. For the choice", "periments. First, we indexed all of the sentences", "of wh-component, we sample a bi-gram based on", "prior probabilities of that bi-gram being associated", "from a Wikipedia dump using the ElasticSearch", "search engine. We also extract named entities for", "with the named-entity type of the answer. This prior", "each sentence in both the Wikipedia corpus and the", "probability is calculated based on named-entity", "sentences used as queries. We assume access to a", "and question bi-gram starters from the SQuAD", "dataset. This information does not make use of", "named-entity recognition system, and in this work", "the full context-question-answer and can be viewed", "'https://github.com/awslabs/", "unsupervised-qa", "2https://spacy.io" ]

{ "arxiv_id": "2004.11892", "language": "en", "page_idx": 1 }

[ [ "0.12006", "0.0766", "0.486", "0.0766", "0.486", "0.0903", "0.12006", "0.0903" ], [ "0.5767", "0.0766", "0.7", "0.0766", "0.7", "0.0903", "0.5767", "0.0903" ], [ "0.7246", "0.0766", "0.754", "0.0766", "0.754", "0.08905", "0.7246", "0.08905" ], [ "0.777", "0.0766", "0.7983", "0.0766", "0.7983", "0.0903", "0.777", "0.0903" ], [ "0.1184", "0.0932", "0.4878", "0.0932", "0.4878", "0.1069", "0.1184", "0.1069" ], [ "0.723", "0.0944", "0.7676", "0.0944", "0.7676", "0.10925", "0.723", "0.10925" ], [ "0.777", "0.0932", "0.8193", "0.0956", "0.818", "0.1081", "0.775", "0.1057" ], [ "0.723", "0.1081", "0.7656", "0.1081", "0.7656", "0.1194", "0.723", "0.1194" ], [ "0.5786", "0.0956", "0.705", "0.0956", "0.705", "0.10925", "0.5786", "0.10925" ], [ "0.1184", "0.10925", "0.4895", "0.10925", "0.4895", "0.1223", "0.1184", "0.1223" ], [ "0.58", "0.10925", "0.711", "0.10925", "0.711", "0.1194", "0.58", "0.1194" ], [ "0.777", "0.1081", "0.818", "0.1081", "0.818", "0.1194", "0.777", "0.1194" ], [ "0.1184", "0.1259", "0.4895", "0.1259", "0.4895", "0.1395", "0.1184", "0.1395" ], [ "0.511", "0.1342", "0.882", "0.1342", "0.882", "0.1478", "0.511", "0.1478" ], [ "0.1184", "0.1401", "0.4895", "0.1395", "0.4895", "0.1538", "0.1184", "0.155" ], [ "0.1184", "0.1567", "0.4902", "0.1567", "0.4902", "0.1704", "0.1184", "0.1704" ], [ "0.5156", "0.152", "0.781", "0.152", "0.781", "0.1621", "0.5156", "0.1621" ], [ "0.11755", "0.1859", "0.2612", "0.187", "0.2612", "0.2019", "0.11755", "0.2007" ], [ "0.511", "0.199", "0.8833", "0.1978", "0.8833", "0.212", "0.511", "0.2131" ], [ "0.11755", "0.2131", "0.4902", "0.2144", "0.4902", "0.2286", "0.11755", "0.228" ], [ "0.51", "0.2156", "0.882", "0.2156", "0.882", "0.2286", "0.51", "0.2286" ], [ "0.11755", "0.231", "0.492", "0.2299", "0.492", "0.2446", "0.11755", "0.2452" ], [ "0.511", "0.2322", "0.882", "0.2322", "0.882", "0.2452", "0.511", "0.2452" ], [ "0.1184", "0.2477", "0.4902", "0.2477", "0.4902", "0.2612", "0.1184", "0.2612" ], [ "0.513", "0.2477", "0.8833", "0.2477", "0.8833", "0.2612", "0.513", "0.2612" ], [ "0.12006", "0.2632", "0.4937", "0.2632", "0.4937", "0.2778", "0.12006", "0.2778" ], [ "0.509", "0.2632", "0.8867", "0.2625", "0.8867", "0.2766", "0.509", "0.2778" ], [ "0.1184", "0.2798", "0.4937", "0.2798", "0.4937", "0.2935", "0.1184", "0.2935" ], [ "0.511", "0.2778", "0.8867", "0.279", "0.8867", "0.2935", "0.511", "0.2922" ], [ "0.12006", "0.2976", "0.4902", "0.2976", "0.4902", "0.308", "0.12006", "0.308" ], [ "0.513", "0.2957", "0.885", "0.2957", "0.885", "0.3088", "0.513", "0.3088" ], [ "0.12006", "0.3123", "0.4902", "0.3123", "0.4902", "0.326", "0.12006", "0.326" ], [ "0.51", "0.3123", "0.8867", "0.3123", "0.8867", "0.326", "0.51", "0.326" ], [ "0.1184", "0.328", "0.492", "0.328", "0.492", "0.3416", "0.1184", "0.3416" ], [ "0.5146", "0.3289", "0.8833", "0.3289", "0.8833", "0.339", "0.5146", "0.339" ], [ "0.11584", "0.3445", "0.4902", "0.3433", "0.4902", "0.357", "0.11584", "0.3582" ], [ "0.513", "0.3433", "0.885", "0.3433", "0.885", "0.357", "0.513", "0.357" ], [ "0.12006", "0.3599", "0.4902", "0.3599", "0.4902", "0.3735", "0.12006", "0.3735" ], [ "0.511", "0.3594", "0.882", "0.3594", "0.882", "0.3723", "0.511", "0.3723" ], [ "0.1184", "0.376", "0.4895", "0.376", "0.4895", "0.3896", "0.1184", "0.3896" ], [ "0.513", "0.3772", "0.8833", "0.3772", "0.8833", "0.3901", "0.513", "0.3901" ], [ "0.12006", "0.3926", "0.4902", "0.3926", "0.4902", "0.4062", "0.12006", "0.4062" ], [ "0.513", "0.3926", "0.881", "0.3926", "0.881", "0.4026", "0.513", "0.4026" ], [ "0.1184", "0.408", "0.4878", "0.408", "0.4878", "0.4216", "0.1184", "0.4216" ], [ "0.511", "0.4067", "0.885", "0.408", "0.885", "0.4229", "0.511", "0.4216" ], [ "0.12006", "0.4246", "0.4895", "0.4246", "0.4895", "0.4382", "0.12006", "0.4382" ], [ "0.513", "0.4246", "0.796", "0.4246", "0.796", "0.4382", "0.513", "0.4382" ], [ "0.11755", "0.4407", "0.4895", "0.4395", "0.4895", "0.4536", "0.11755", "0.4548" ], [ "0.531", "0.4436", "0.885", "0.4436", "0.885", "0.4578", "0.531", "0.4578" ], [ "0.1184", "0.4573", "0.4895", "0.4573", "0.4895", "0.4702", "0.1184", "0.4702" ], [ "0.511", "0.4602", "0.8833", "0.4602", "0.8833", "0.4739", "0.511", "0.4739" ], [ "0.12006", "0.4744", "0.4895", "0.4744", "0.4895", "0.485", "0.12006", "0.485" ], [ "0.511", "0.4768", "0.8833", "0.4768", "0.8833", "0.4905", "0.511", "0.4905" ], [ "0.1184", "0.4893", "0.4937", "0.4893", "0.4937", "0.503", "0.1184", "0.503" ], [ "0.513", "0.4922", "0.8833", "0.4922", "0.8833", "0.506", "0.513", "0.506" ], [ "0.1184", "0.505", "0.4895", "0.505", "0.4895", "0.5186", "0.1184", "0.5186" ], [ "0.511", "0.5083", "0.8867", "0.5073", "0.8867", "0.5215", "0.511", "0.5225" ], [ "0.1184", "0.521", "0.4878", "0.521", "0.4878", "0.535", "0.1184", "0.535" ], [ "0.513", "0.525", "0.8867", "0.524", "0.8867", "0.538", "0.513", "0.539" ], [ "0.12006", "0.5376", "0.4895", "0.5376", "0.4895", "0.5503", "0.12006", "0.5503" ], [ "0.513", "0.5415", "0.885", "0.5415", "0.885", "0.5547", "0.513", "0.5547" ], [ "0.1184", "0.5527", "0.2703", "0.5527", "0.2703", "0.568", "0.1184", "0.568" ], [ "0.511", "0.5557", "0.885", "0.557", "0.885", "0.572", "0.511", "0.571" ], [ "0.513", "0.5737", "0.885", "0.5737", "0.885", "0.5874", "0.513", "0.5874" ], [ "0.1184", "0.582", "0.2854", "0.582", "0.2854", "0.5967", "0.1184", "0.5967" ], [ "0.513", "0.589", "0.885", "0.589", "0.885", "0.6025", "0.513", "0.6025" ], [ "0.12006", "0.605", "0.4937", "0.605", "0.4937", "0.618", "0.12006", "0.618" ], [ "0.511", "0.605", "0.8867", "0.604", "0.8867", "0.618", "0.511", "0.619" ], [ "0.1184", "0.6216", "0.4954", "0.6216", "0.4954", "0.6353", "0.1184", "0.6353" ], [ "0.51", "0.6216", "0.8867", "0.6216", "0.8867", "0.6353", "0.51", "0.6353" ], [ "0.11755", "0.6357", "0.4937", "0.6357", "0.4937", "0.651", "0.11755", "0.651" ], [ "0.511", "0.637", "0.8867", "0.637", "0.8867", "0.651", "0.511", "0.651" ], [ "0.12006", "0.654", "0.4902", "0.654", "0.4902", "0.6675", "0.12006", "0.6675" ], [ "0.511", "0.652", "0.882", "0.652", "0.882", "0.665", "0.511", "0.665" ], [ "0.12006", "0.6704", "0.4937", "0.6704", "0.4937", "0.684", "0.12006", "0.684" ], [ "0.513", "0.6694", "0.885", "0.6694", "0.885", "0.683", "0.513", "0.683" ], [ "0.12006", "0.686", "0.4937", "0.686", "0.4937", "0.6997", "0.12006", "0.6997" ], [ "0.513", "0.686", "0.885", "0.686", "0.885", "0.6997", "0.513", "0.6997" ], [ "0.12177", "0.7036", "0.492", "0.7036", "0.492", "0.7144", "0.12177", "0.7144" ], [ "0.5146", "0.7036", "0.882", "0.7036", "0.882", "0.7144", "0.5146", "0.7144" ], [ "0.12006", "0.7188", "0.4937", "0.7188", "0.4937", "0.7324", "0.12006", "0.7324" ], [ "0.511", "0.7173", "0.882", "0.7173", "0.882", "0.731", "0.511", "0.731" ], [ "0.1184", "0.733", "0.4895", "0.733", "0.4895", "0.7466", "0.1184", "0.7466" ], [ "0.511", "0.733", "0.881", "0.733", "0.881", "0.7466", "0.511", "0.7466" ], [ "0.12006", "0.752", "0.4878", "0.752", "0.4878", "0.7627", "0.12006", "0.7627" ], [ "0.513", "0.7505", "0.8833", "0.7505", "0.8833", "0.764", "0.513", "0.764" ], [ "0.1184", "0.7656", "0.4895", "0.7656", "0.4895", "0.7793", "0.1184", "0.7793" ], [ "0.513", "0.7666", "0.885", "0.7666", "0.885", "0.78", "0.513", "0.78" ], [ "0.11755", "0.781", "0.4895", "0.782", "0.4895", "0.7964", "0.11755", "0.796" ], [ "0.5146", "0.782", "0.885", "0.782", "0.885", "0.796", "0.5146", "0.796" ], [ "0.11755", "0.799", "0.4902", "0.799", "0.4902", "0.8125", "0.11755", "0.8125" ], [ "0.511", "0.799", "0.885", "0.799", "0.885", "0.8125", "0.511", "0.8125" ], [ "0.1184", "0.8164", "0.4878", "0.8164", "0.4878", "0.8267", "0.1184", "0.8267" ], [ "0.511", "0.814", "0.8833", "0.8135", "0.8833", "0.8276", "0.511", "0.829" ], [ "0.12006", "0.8306", "0.4902", "0.8306", "0.4902", "0.844", "0.12006", "0.844" ], [ "0.51", "0.832", "0.885", "0.832", "0.885", "0.8457", "0.51", "0.8457" ], [ "0.1184", "0.847", "0.4937", "0.847", "0.4937", "0.861", "0.1184", "0.861" ], [ "0.513", "0.847", "0.885", "0.847", "0.885", "0.861", "0.513", "0.861" ], [ "0.1184", "0.8633", "0.4937", "0.8633", "0.4937", "0.877", "0.1184", "0.877" ], [ "0.51", "0.8633", "0.885", "0.8633", "0.885", "0.877", "0.51", "0.877" ], [ "0.1184", "0.88", "0.4937", "0.88", "0.4937", "0.8936", "0.1184", "0.8936" ], [ "0.511", "0.879", "0.8833", "0.879", "0.8833", "0.8936", "0.511", "0.8936" ], [ "0.1184", "0.8936", "0.4895", "0.8955", "0.4895", "0.91", "0.1184", "0.9077" ], [ "0.513", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.513", "0.909" ] ]

[ "as prior information, not disturbing the integrity", "Training procedure", "EM", "F1", "of our unsupervised approach. Additionally, the", "17.36", "25.90", "30.53", "Cloze-style original", "choice of wh component does not significantly af-", "Cloze-style retrieved", "39.61", "fect results. For template-based approaches, we", "Table I: Effect of original vs retrieved sentences for", "also experimented with clause-based templates but", "did not find significant differences in performance.", "generic cloze-style questiongeneration.", "3Experiments", "semantically matching sentences. Additionally, we", "Settings: For all downstream question answering", " believe the sentences which have loosely related", "models, we fine-tune a pretrained BERT model us-", " meaning may act as a regularization factor which", "ing the Transformers repository (Wolf et al., 2019)", "prevent the downstream QA model from learning", "and report ablation study numbers using the base-", "only string matching patterns. Along these lines,", "uncased version of BERT, consistent with (Lewis", "(Lewis et al., 2019) found that a simple noise func-", "et al.,2019).All models are trained and validated", "tion of dropping, masking and permuting words", "on generated pairs of questions and answers along", " was a strong question generation baseline. We be-", "with their contexts tested on the SQuAD develop-", "lievethatlooselyrelatedcontextsentencescanact", "ment set. The training set differs for each ablation", "as a more intuitive noise function, and investigat-", "study and will be described below, while the valida-", "ing the role of the semantic match of the retrieved", "tion dataset is a random set of 1,000 template-based", "sentences is an important direction for future work", "generated data points, which is consistent across", "Forthe sections whichfollow.we only showresults", "all ablation studies. We train all QA models for 2", "of retrieved sentences, as the trend of improved per-", "epochs, checkpointing the models every 500 steps", "formance held across all experiments.", "and choosing the checkpoint with the highest F1", "Effect of template components: We evaluate", "score on the validation set as the best model. All", "the effect of individual template components on", "ablation studies areaveraged overtwotrainingruns", " downstream QA performance. Results are shown", "with different seeds. Unless otherwise stated, ex-", "in Table 2. Wh template methods improve largely", "periments are performed using 50,000 synthetic", "over the simple cloze templates.\"Wh + B + A +?\"", "QA training examples, as initial models performed", "performs best among the template-based methods,", "best with this amount. We will make this generated", "as having the Wh word at the beginning most re-", "training data public.", " sembles the target SQuAD domain and switching", "the order of Fragment B and Fragment A may force", "3.1 Model Analysis", "the model to learn more complex relationships from", "Effect of retrieved sentences: We test the effect", "the question. We additionally test the effect of the", "of retrieved vs original sentences as input to ques-", " wh-component and the question mark added at the", "tion generation when using generic cloze questions.", " end of the sentence. Using the same data as \"Wh +", "As shown in Table l, using retrieved sentences", "B + A + ?\" but removing the wh-component results", "improves over using the original sentence, rein-", "in a large decrease in performance. We believe that", "forcing our motivation that a retrieved sentence,", "this is because the wh-component signals the type", "whichmaynotmatchtriviallythe current context,", "of possible answerentities,whichhelps narrow", "forces the QA model to learn more complex re-", "down the space of possible answers. Removing the", "lationships than just simple entity matching. The", " question mark at the end of the template also results", "retrievalprocessmayreturnsentenceswhichdonot", "in decreased performance, but not as large as re-", "match the original context. On a random sample,", "moving the wh-component. This may be a result of", "15/18 retrieved sentences were judged as entirely", "BERT pretraining which expects certain punctua-", "relevant to the original sentence. This retrieval", "tion based on sentence structure. We note that these", "is already quite good, as we use a high quality", "questions may not be grammatical, which may have", "ElasticSearch retrieval and use the original context", "an impact on performance. Improving the ques-", "sentence as the query, not just the answer word.", "tion quality makes a difference in performance as", "While we do not explicitly ensure that the retrieved", " seen from the jump from cloze-style questions to", "sentence has the same meaning, we find that the", "template questions. The ablation studies suggest", "search results with entity matching gives largely", "that a combination of question relevance, though" ]

{ "arxiv_id": "2004.11892", "language": "en", "page_idx": 2 }

[ [ "0.126", "0.076", "0.2024", "0.076", "0.2024", "0.0873", "0.126", "0.0873" ], [ "0.3938", "0.076", "0.4836", "0.076", "0.4836", "0.0873", "0.3938", "0.0873" ], [ "0.529", "0.076", "0.8833", "0.076", "0.8833", "0.08905", "0.529", "0.08905" ], [ "0.529", "0.08905", "0.885", "0.0885", "0.885", "0.1027", "0.529", "0.10394" ], [ "0.3015", "0.0766", "0.3855", "0.0766", "0.3855", "0.0873", "0.3015", "0.0873" ], [ "0.126", "0.0903", "0.2225", "0.0903", "0.2225", "0.10156", "0.126", "0.10156" ], [ "0.3005", "0.0903", "0.3333", "0.0903", "0.3333", "0.10156", "0.3005", "0.10156" ], [ "0.3938", "0.0903", "0.4282", "0.0903", "0.4282", "0.1027", "0.3938", "0.1027" ], [ "0.1276", "0.0986", "0.2913", "0.0986", "0.2913", "0.11224", "0.1276", "0.11224" ], [ "0.3005", "0.10156", "0.3315", "0.10156", "0.3315", "0.10986", "0.3005", "0.10986" ], [ "0.3938", "0.0998", "0.4282", "0.0998", "0.4282", "0.111", "0.3938", "0.111" ], [ "0.531", "0.1027", "0.781", "0.1027", "0.781", "0.1164", "0.531", "0.1164" ], [ "0.11755", "0.1259", "0.4902", "0.1259", "0.4902", "0.1395", "0.11755", "0.1395" ], [ "0.513", "0.1271", "0.885", "0.1271", "0.885", "0.1401", "0.513", "0.1401" ], [ "0.11755", "0.1401", "0.2032", "0.1401", "0.2032", "0.152", "0.11755", "0.152" ], [ "0.531", "0.1401", "0.885", "0.1401", "0.885", "0.1538", "0.531", "0.1538" ], [ "0.531", "0.1538", "0.8833", "0.1538", "0.8833", "0.1644", "0.531", "0.1644" ], [ "0.531", "0.1663", "0.885", "0.1663", "0.885", "0.1799", "0.531", "0.1799" ], [ "0.1184", "0.1799", "0.4902", "0.1787", "0.4902", "0.1936", "0.1184", "0.1948" ], [ "0.531", "0.1799", "0.885", "0.1799", "0.885", "0.1936", "0.531", "0.1936" ], [ "0.11755", "0.1965", "0.4878", "0.1965", "0.4878", "0.2114", "0.11755", "0.2114" ], [ "0.531", "0.1924", "0.8833", "0.1924", "0.8833", "0.206", "0.531", "0.206" ], [ "0.531", "0.206", "0.885", "0.206", "0.885", "0.2197", "0.531", "0.2197" ], [ "0.1184", "0.212", "0.4895", "0.2131", "0.4895", "0.228", "0.1184", "0.2268" ], [ "0.529", "0.2185", "0.8867", "0.2173", "0.8867", "0.2322", "0.529", "0.2328" ], [ "0.12006", "0.2299", "0.4895", "0.2299", "0.4895", "0.2434", "0.12006", "0.2434" ], [ "0.531", "0.2322", "0.7793", "0.2322", "0.7793", "0.2452", "0.531", "0.2452" ], [ "0.1184", "0.2452", "0.4895", "0.2452", "0.4895", "0.2588", "0.1184", "0.2588" ], [ "0.5146", "0.2546", "0.8833", "0.2546", "0.8833", "0.2683", "0.5146", "0.2683" ], [ "0.11755", "0.2625", "0.4878", "0.2625", "0.4878", "0.2756", "0.11755", "0.2756" ], [ "0.529", "0.2683", "0.8833", "0.2666", "0.8833", "0.2808", "0.529", "0.2832" ], [ "0.1184", "0.2778", "0.4878", "0.2778", "0.4878", "0.2915", "0.1184", "0.2915" ], [ "0.531", "0.2832", "0.678", "0.282", "0.678", "0.2935", "0.531", "0.2944" ], [ "0.679", "0.2808", "0.885", "0.2798", "0.885", "0.2944", "0.679", "0.2957" ], [ "0.6753", "0.2874", "0.684", "0.2874", "0.684", "0.2903", "0.6753", "0.2903" ], [ "0.12006", "0.2935", "0.4878", "0.2935", "0.4878", "0.3071", "0.12006", "0.3071" ], [ "0.534", "0.2964", "0.882", "0.2964", "0.882", "0.3071", "0.534", "0.3071" ], [ "0.1184", "0.3088", "0.4895", "0.3088", "0.4895", "0.3237", "0.1184", "0.3237" ], [ "0.529", "0.3071", "0.882", "0.3071", "0.882", "0.3206", "0.529", "0.3206" ], [ "0.1184", "0.326", "0.4895", "0.326", "0.4895", "0.339", "0.1184", "0.339" ], [ "0.531", "0.3206", "0.8867", "0.3206", "0.8867", "0.3342", "0.531", "0.3342" ], [ "0.529", "0.3333", "0.885", "0.3333", "0.885", "0.3467", "0.529", "0.3467" ], [ "0.12006", "0.3433", "0.4878", "0.3433", "0.4878", "0.354", "0.12006", "0.354" ], [ "0.531", "0.3467", "0.771", "0.3467", "0.771", "0.3599", "0.531", "0.3599" ], [ "0.1184", "0.3594", "0.4895", "0.3594", "0.4895", "0.3723", "0.1184", "0.3723" ], [ "0.11755", "0.3723", "0.2988", "0.3735", "0.2988", "0.3884", "0.11755", "0.3872" ], [ "0.5146", "0.3706", "0.8833", "0.3706", "0.8833", "0.3843", "0.5146", "0.3843" ], [ "0.534", "0.3855", "0.8833", "0.3855", "0.8833", "0.3955", "0.534", "0.3955" ], [ "0.1427", "0.401", "0.2477", "0.4026", "0.2477", "0.4175", "0.1427", "0.415" ], [ "0.531", "0.3967", "0.8867", "0.3967", "0.8867", "0.4104", "0.531", "0.4104" ], [ "0.1184", "0.4026", "0.1453", "0.4026", "0.1453", "0.4146", "0.1184", "0.4146" ], [ "0.1427", "0.4062", "0.1486", "0.4062", "0.1486", "0.411", "0.1427", "0.411" ], [ "0.532", "0.4104", "0.885", "0.4104", "0.885", "0.4233", "0.532", "0.4233" ], [ "0.532", "0.4229", "0.885", "0.4229", "0.885", "0.4358", "0.532", "0.4358" ], [ "0.1184", "0.4287", "0.4895", "0.4287", "0.4895", "0.4424", "0.1184", "0.4424" ], [ "0.531", "0.4358", "0.8867", "0.4358", "0.8867", "0.4495", "0.531", "0.4495" ], [ "0.1184", "0.4448", "0.4902", "0.4448", "0.4902", "0.4578", "0.1184", "0.4578" ], [ "0.529", "0.4478", "0.885", "0.4495", "0.885", "0.4631", "0.529", "0.4614" ], [ "0.11755", "0.4614", "0.4878", "0.4614", "0.4878", "0.4744", "0.11755", "0.4744" ], [ "0.531", "0.462", "0.885", "0.462", "0.885", "0.4756", "0.531", "0.4756" ], [ "0.1184", "0.4768", "0.4895", "0.4768", "0.4895", "0.4905", "0.1184", "0.4905" ], [ "0.529", "0.4756", "0.8213", "0.4756", "0.8213", "0.4893", "0.529", "0.4893" ], [ "0.1184", "0.4922", "0.4878", "0.4922", "0.4878", "0.506", "0.1184", "0.506" ], [ "0.511", "0.4976", "0.885", "0.4988", "0.885", "0.513", "0.511", "0.5127" ], [ "0.1184", "0.509", "0.4895", "0.509", "0.4895", "0.5225", "0.1184", "0.5225" ], [ "0.532", "0.513", "0.586", "0.513", "0.586", "0.524", "0.532", "0.524" ], [ "0.5767", "0.511", "0.885", "0.513", "0.885", "0.527", "0.5767", "0.525" ], [ "0.12006", "0.525", "0.4895", "0.525", "0.4895", "0.538", "0.12006", "0.538" ], [ "0.529", "0.525", "0.885", "0.525", "0.885", "0.538", "0.529", "0.538" ], [ "0.11755", "0.5405", "0.4895", "0.539", "0.4895", "0.554", "0.11755", "0.5547" ], [ "0.531", "0.538", "0.885", "0.538", "0.885", "0.552", "0.531", "0.552" ], [ "0.531", "0.5503", "0.8867", "0.5503", "0.8867", "0.5654", "0.531", "0.5654" ], [ "0.1184", "0.557", "0.4895", "0.557", "0.4895", "0.571", "0.1184", "0.571" ], [ "0.529", "0.5635", "0.7173", "0.5635", "0.7173", "0.5767", "0.529", "0.5767" ], [ "0.1184", "0.5737", "0.4895", "0.5737", "0.4895", "0.5874", "0.1184", "0.5874" ], [ "0.12006", "0.5913", "0.4895", "0.5913", "0.4895", "0.6016", "0.12006", "0.6016" ], [ "0.5146", "0.5884", "0.8867", "0.5884", "0.8867", "0.6016", "0.5146", "0.6016" ], [ "0.11755", "0.605", "0.287", "0.6055", "0.287", "0.6206", "0.11755", "0.619" ], [ "0.531", "0.6016", "0.885", "0.6016", "0.885", "0.615", "0.531", "0.615" ], [ "0.529", "0.6133", "0.8833", "0.6133", "0.8833", "0.627", "0.529", "0.627" ], [ "0.1184", "0.632", "0.3198", "0.632", "0.3198", "0.6465", "0.1184", "0.6465" ], [ "0.529", "0.627", "0.8867", "0.629", "0.8867", "0.6426", "0.529", "0.64" ], [ "0.529", "0.64", "0.817", "0.6396", "0.817", "0.654", "0.529", "0.655" ], [ "0.1184", "0.659", "0.4895", "0.659", "0.4895", "0.673", "0.1184", "0.673" ], [ "0.511", "0.662", "0.885", "0.6646", "0.885", "0.6787", "0.511", "0.677" ], [ "0.11755", "0.676", "0.1763", "0.676", "0.1763", "0.6904", "0.11755", "0.6904" ], [ "0.531", "0.6777", "0.8833", "0.6777", "0.8833", "0.6914", "0.531", "0.6914" ], [ "0.532", "0.6924", "0.8833", "0.6924", "0.8833", "0.703", "0.532", "0.703" ], [ "0.531", "0.7036", "0.8867", "0.7036", "0.8867", "0.7173", "0.531", "0.7173" ], [ "0.11755", "0.7163", "0.215", "0.7188", "0.215", "0.733", "0.11755", "0.731" ], [ "0.532", "0.7163", "0.8867", "0.7163", "0.8867", "0.731", "0.532", "0.731" ], [ "0.531", "0.7285", "0.8716", "0.7285", "0.8716", "0.742", "0.531", "0.742" ], [ "0.1184", "0.7407", "0.4895", "0.7407", "0.4895", "0.7544", "0.1184", "0.7544" ], [ "0.1368", "0.7544", "0.4902", "0.7544", "0.4902", "0.767", "0.1368", "0.767" ], [ "0.511", "0.753", "0.885", "0.7544", "0.885", "0.7686", "0.511", "0.767" ], [ "0.1377", "0.767", "0.4902", "0.767", "0.4902", "0.781", "0.1377", "0.781" ], [ "0.531", "0.767", "0.8867", "0.767", "0.8867", "0.781", "0.531", "0.781" ], [ "0.1368", "0.78", "0.4902", "0.78", "0.4902", "0.7944", "0.1368", "0.7944" ], [ "0.531", "0.781", "0.885", "0.781", "0.885", "0.7944", "0.531", "0.7944" ], [ "0.1351", "0.792", "0.4895", "0.7915", "0.4895", "0.8057", "0.1351", "0.807" ], [ "0.529", "0.792", "0.8833", "0.792", "0.8833", "0.8057", "0.529", "0.8057" ], [ "0.1368", "0.8057", "0.4902", "0.8057", "0.4902", "0.8193", "0.1368", "0.8193" ], [ "0.532", "0.808", "0.8833", "0.808", "0.8833", "0.8184", "0.532", "0.8184" ], [ "0.1368", "0.8193", "0.2603", "0.821", "0.2603", "0.832", "0.1368", "0.8306" ], [ "0.531", "0.8193", "0.885", "0.8193", "0.885", "0.833", "0.531", "0.833" ], [ "0.531", "0.832", "0.8867", "0.832", "0.8867", "0.8457", "0.531", "0.8457" ], [ "0.11755", "0.8413", "0.4895", "0.843", "0.4895", "0.857", "0.11755", "0.8555" ], [ "0.531", "0.8467", "0.5903", "0.8467", "0.5903", "0.858", "0.531", "0.858" ], [ "0.1368", "0.858", "0.4878", "0.858", "0.4878", "0.868", "0.1368", "0.868" ], [ "0.1351", "0.869", "0.4878", "0.869", "0.4878", "0.883", "0.1351", "0.883" ], [ "0.5146", "0.8706", "0.8833", "0.8706", "0.8833", "0.8813", "0.5146", "0.8813" ], [ "0.1368", "0.883", "0.4895", "0.883", "0.4895", "0.8965", "0.1368", "0.8965" ], [ "0.534", "0.8843", "0.8833", "0.8843", "0.8833", "0.8945", "0.534", "0.8945" ], [ "0.1368", "0.898", "0.4878", "0.898", "0.4878", "0.9077", "0.1368", "0.9077" ], [ "0.531", "0.898", "0.8833", "0.898", "0.8833", "0.9077", "0.531", "0.9077" ] ]

[ "Model Choice", "SQuAD NER F1", "Language Technologies, Volume 1 (Long Papers),", "pages 1585-1594, New Orleans, Louisiana. Associ-", "SQuAD Test F1", "BERT-large (ours)", "64.04", "77.55", "BERT-large (Lewis et al., 2019)", "56.40", "64.50", " ation for Computational Linguistics.", "Table 4: A comparison of top results using the BERT-", "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and", "large model.", "Kristina Toutanova. 2019. BERT: Pre-training of", "deep bidirectional transformers for language under", "standing.In Proceedings of the 2019 Conference", "BERT-large, although using the original BERT-", "of the North American Chapter of the Association", "large gives similar performance of 62.69 on the", "for Computational Linguistics: Human Language", "Technologies, Volume 1 (Long and Short Papers),", "SQuAD dev set. We report numbers on the sam-", "pages 4171-4186, Minneapolis, Minnesota. Associ-", "ple of SQuAD questions which are named entities,", "ation for Computational Linguistics.", "which we refer to as SQuAD-NER. The subset cor-", "Bhuwan Dhingra, Danish Danish, and Dheeraj Ra", "responding to the SQuAD development dataset has", "jagopal.2018.Simple and effective semi-supervised", "4,338 samples, and may differ slightly from (Lewis", "question answering.", "In Proceedings of the 2018", "", "et al., 2019) due to differences in NER preprocess", "Conference of theNorth American Chapter of the", "ing. We also trained a fully-supervised model on", "Association for Computational Linguistics: Human", "the SQuAD training dataset with varying amounts", "Language Technologies, Volume 2 (Short Papers),", "pages 582-587, New Orleans, Louisiana. Associa-", "ofdataandfoundourunsupervisedperformance", "tion for Computational Linguistics.", "equals the supervised performance trained on about", "3,000 labeled examples.", "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red", "field,Michael Collins, Ankur Parikh, Chris Al-", "Conclusion", "berti, Danielle Epstein, Illia Polosukhin, Jacob De-", "4", "", "vlin, Kenton Lee, Kristina Toutanova, Llion Jones,", "Matthew Kelcey, Ming-Wei Chang, Andrew M.Dai,", "In this paper we introduce a retrieval-based ap-", "Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019.", "proach to unsupervised extractive question answer-", "Natural questions: A benchmark for question an-", "ing. A simple template-based approach achieves", "swering research. Transactions of the Association", "state-of-the-art results for unsupervised methods", "for Computational Linguistics, 7:453-466.", "on the SQuAD dataset of 64.04 F1, and 77.55 F1", "Patrick Lewis, Ludovic Denoyer, and Sebastian Riedel.", "when the answer is a named entity. We analyze the", "2019.", "Unsupervised question answering by cloze", "effect of several components in our template-based", "translation. In Proceedings of the 57th Annual Meet-", "approaches through ablation studies. We aim to", "ing of the Association for Computational Linguistics,", "pages 4896-4910, Florence, Italy. Association for", "experiment with other datasets and other domains,", " Computational Linguistics.", "incorporate our synthetic data in a semi-supervised", "settingandtestthefeasibility of ourframeworkin", "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-", "a multi-lingual setting.", "dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,", "Luke Zettlemoyer, and Veselin Stoyanov. 2019", "5Acknowledgements", "Roberta:A robustly optimized bert pretraining ap-", "proach. arXiv preprint arXiv:1907.11692.", "We thank Xiaofei Ma for fruitful discussions on the", "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and", "project.", "Percy Liang. 2016. SQuAD: 100,000+ questions for", "machine comprehension of textIn Proceedings of", "the 2016 Conference on Empirical Methods in Natu-", "References", "ral Language Processing, pages 2383-2392, Austin,", "Texas. Association for Computational Linguistics.", "Chris Alberti, Daniel Andor, Emily Pitler, Jacob De-", "vlin, and Michael Collins. 2019. Synthetic QA cor-", "Liang Wang, Sujian Li, Wei Zhao, Kewei Shen, Meng", "pora generation with roundtrip consistency. In Pro-", "Sun, Ruoyu Jia, and Jingming Liu. 2018. Multi-", "ceedings of the 57th Annual Meeting of the Asso-", "perspective context aggregation for semi-supervised", "ciation for Computational Linguistics, pages 6168-", "cloze-style reading comprehension. In Proceedings", "6173, Florence, Italy. Association for Computa-", "of the 27th International Conference on Computa-", "tional Linguistics.", "tional Linguistics, pages 857-867, Santa Fe, New", "Mexico, USA. Association for Computational Lin-", "Yu-An Chung, Hung-Yi Lee, and James Glass. 2018.", "guistics.", "Supervised and unsupervised transfer learning for", "question answering.In Proceedings of the 2018", "Thomas Wolf, Lysandre Debut,Victor Sanh,Julien", "Conference of the North American Chapter of the", "Chaumond,ClementDelangue,AnthonyMoi,Pier-", "Association for Computational Linguistics:Human", "ricCistac,TimRault, R'emiLouf,MorganFuntow-" ]

{ "arxiv_id": "2004.11892", "language": "en", "page_idx": 4 }

[ [ "0.1772", "0.1218", "0.8228", "0.1225", "0.8228", "0.1401", "0.1772", "0.1395" ], [ "0.335", "0.1427", "0.664", "0.1433", "0.664", "0.161", "0.335", "0.1604" ], [ "0.1168", "0.1724", "0.886", "0.1724", "0.886", "0.1901", "0.1168", "0.1901" ], [ "0.156", "0.1925", "0.841", "0.1925", "0.841", "0.207", "0.156", "0.207" ], [ "0.1177", "0.2102", "0.8813", "0.2122", "0.8813", "0.2267", "0.1177", "0.2247" ], [ "0.2834", "0.2247", "0.716", "0.226", "0.716", "0.2405", "0.2834", "0.2393" ], [ "0.3513", "0.2405", "0.648", "0.2405", "0.648", "0.254", "0.3513", "0.254" ], [ "0.02287", "0.2583", "0.05963", "0.2583", "0.05963", "0.406", "0.02287", "0.406" ], [ "0.1462", "0.2734", "0.407", "0.2734", "0.407", "0.2874", "0.1462", "0.2874" ], [ "0.4307", "0.2727", "0.705", "0.2727", "0.705", "0.286", "0.4307", "0.286" ], [ "0.74", "0.2734", "0.853", "0.2747", "0.853", "0.289", "0.74", "0.2886" ], [ "0.3179", "0.2886", "0.3733", "0.2886", "0.3733", "0.2998", "0.3179", "0.2998" ], [ "0.4429", "0.2922", "0.451", "0.2935", "0.4495", "0.2974", "0.4412", "0.2966" ], [ "0.6045", "0.2886", "0.661", "0.2886", "0.661", "0.2986", "0.6045", "0.2986" ], [ "0.282", "0.2998", "0.407", "0.3005", "0.407", "0.3125", "0.282", "0.3113" ], [ "0.5557", "0.3005", "0.726", "0.3005", "0.726", "0.3113", "0.5557", "0.3113" ], [ "0.2688", "0.3132", "0.4192", "0.3132", "0.4192", "0.324", "0.2688", "0.324" ], [ "0.54", "0.3113", "0.7256", "0.3113", "0.7256", "0.3245", "0.54", "0.3245" ], [ "0.5933", "0.3467", "0.6772", "0.3467", "0.6772", "0.3833", "0.5933", "0.3833" ], [ "0.3179", "0.3547", "0.3757", "0.348", "0.3816", "0.3726", "0.3228", "0.3801" ], [ "0.5986", "0.3655", "0.603", "0.3655", "0.603", "0.3706", "0.5986", "0.3706" ], [ "0.02124", "0.3977", "0.05637", "0.3965", "0.05963", "0.4841", "0.0237", "0.4849" ], [ "0.04166", "0.3945", "0.04575", "0.3945", "0.04575", "0.3997", "0.04166", "0.3997" ], [ "0.2607", "0.4048", "0.4224", "0.4048", "0.4224", "0.4153", "0.2607", "0.4153" ], [ "0.581", "0.406", "0.681", "0.406", "0.681", "0.4172", "0.581", "0.4172" ], [ "0.4558", "0.4092", "0.524", "0.4104", "0.524", "0.4216", "0.4558", "0.4204" ], [ "0.1356", "0.4092", "0.2573", "0.4092", "0.2573", "0.4192", "0.1356", "0.4192" ], [ "0.5933", "0.4185", "0.6323", "0.4185", "0.6323", "0.4287", "0.5933", "0.4287" ], [ "0.66", "0.4172", "0.7173", "0.4172", "0.7173", "0.428", "0.66", "0.428" ], [ "0.4126", "0.4204", "0.5605", "0.4224", "0.5605", "0.4343", "0.4126", "0.432" ], [ "0.1765", "0.4216", "0.2239", "0.4216", "0.2239", "0.4287", "0.1765", "0.4287" ], [ "0.2484", "0.4224", "0.304", "0.4224", "0.304", "0.433", "0.2484", "0.433" ], [ "0.3586", "0.4224", "0.3823", "0.4224", "0.3823", "0.4255", "0.3586", "0.4255" ], [ "0.5947", "0.4287", "0.614", "0.4287", "0.614", "0.4363", "0.5947", "0.4363" ], [ "0.7363", "0.4268", "0.8545", "0.4268", "0.8545", "0.4375", "0.7363", "0.4375" ], [ "0.0866", "0.4539", "0.911", "0.4539", "0.911", "0.4685", "0.0866", "0.4685" ], [ "0.1765", "0.43", "0.2091", "0.43", "0.2091", "0.4375", "0.1765", "0.4375" ], [ "0.0866", "0.4685", "0.3743", "0.4685", "0.3743", "0.4817", "0.0866", "0.4817" ], [ "0.0237", "0.4944", "0.05637", "0.4944", "0.05637", "0.58", "0.0237", "0.58" ], [ "0.2507", "0.502", "0.3162", "0.502", "0.3162", "0.5166", "0.2507", "0.5166" ], [ "0.535", "0.502", "0.8945", "0.5024", "0.8945", "0.5166", "0.535", "0.515" ], [ "0.5376", "0.5176", "0.8945", "0.5176", "0.8945", "0.528", "0.5376", "0.528" ], [ "0.10297", "0.526", "0.4631", "0.526", "0.4631", "0.539", "0.10297", "0.539" ], [ "0.535", "0.529", "0.8823", "0.529", "0.8823", "0.5425", "0.535", "0.5425" ], [ "0.10297", "0.539", "0.4617", "0.539", "0.4617", "0.553", "0.10297", "0.553" ], [ "0.1013", "0.552", "0.4631", "0.552", "0.4631", "0.565", "0.1013", "0.565" ], [ "0.1013", "0.563", "0.4631", "0.563", "0.4631", "0.577", "0.1013", "0.577" ], [ "0.6626", "0.5645", "0.7695", "0.5645", "0.7695", "0.579", "0.6626", "0.579" ], [ "0.10297", "0.5776", "0.46", "0.5776", "0.46", "0.5884", "0.10297", "0.5884" ], [ "0.10297", "0.5903", "0.4592", "0.5903", "0.4592", "0.601", "0.10297", "0.601" ], [ "0.5215", "0.5854", "0.9126", "0.5854", "0.9126", "0.5986", "0.5215", "0.5986" ], [ "0.02287", "0.5913", "0.04984", "0.5854", "0.058", "0.606", "0.03105", "0.6123" ], [ "0.1013", "0.6016", "0.4631", "0.6016", "0.4631", "0.6157", "0.1013", "0.6157" ], [ "0.5195", "0.5986", "0.911", "0.5986", "0.911", "0.6123", "0.5195", "0.6123" ], [ "0.03595", "0.607", "0.0433", "0.6123", "0.03595", "0.617", "0.02942", "0.6123" ], [ "0.1013", "0.613", "0.46", "0.613", "0.46", "0.627", "0.1013", "0.627" ], [ "0.5195", "0.6123", "0.914", "0.6123", "0.914", "0.6255", "0.5195", "0.6255" ], [ "0.02777", "0.627", "0.05392", "0.627", "0.05392", "0.648", "0.02777", "0.648" ], [ "0.02777", "0.6143", "0.05392", "0.6143", "0.05392", "0.632", "0.02777", "0.632" ], [ "0.10297", "0.627", "0.4631", "0.627", "0.4631", "0.64", "0.10297", "0.64" ], [ "0.5195", "0.627", "0.91", "0.627", "0.91", "0.64", "0.5195", "0.64" ], [ "0.0237", "0.637", "0.058", "0.637", "0.058", "0.711", "0.0237", "0.711" ], [ "0.10297", "0.6396", "0.4617", "0.6396", "0.4617", "0.654", "0.10297", "0.654" ], [ "0.5195", "0.64", "0.91", "0.64", "0.91", "0.654", "0.5195", "0.654" ], [ "0.1013", "0.6514", "0.4617", "0.6514", "0.4617", "0.6655", "0.1013", "0.6655" ], [ "0.5186", "0.653", "0.9126", "0.654", "0.9126", "0.6685", "0.5186", "0.6675" ], [ "0.1013", "0.664", "0.4631", "0.664", "0.4631", "0.678", "0.1013", "0.678" ], [ "0.5215", "0.6675", "0.9126", "0.6675", "0.9126", "0.681", "0.5215", "0.681" ], [ "0.10297", "0.6787", "0.4592", "0.6787", "0.4592", "0.6895", "0.10297", "0.6895" ], [ "0.5186", "0.681", "0.9126", "0.6797", "0.9126", "0.6943", "0.5186", "0.696" ], [ "0.10297", "0.6914", "0.46", "0.6914", "0.46", "0.702", "0.10297", "0.702" ], [ "0.5195", "0.696", "0.914", "0.696", "0.914", "0.709", "0.5195", "0.709" ], [ "0.1013", "0.702", "0.4631", "0.7026", "0.4631", "0.7163", "0.1013", "0.7153" ], [ "0.1013", "0.714", "0.4631", "0.7153", "0.4631", "0.729", "0.1013", "0.728" ], [ "0.5195", "0.709", "0.7534", "0.709", "0.7534", "0.7227", "0.5195", "0.7227" ], [ "0.10455", "0.729", "0.4592", "0.729", "0.4592", "0.7393", "0.10455", "0.7393" ], [ "0.536", "0.7236", "0.914", "0.7236", "0.914", "0.7373", "0.536", "0.7373" ], [ "0.10455", "0.741", "0.4592", "0.741", "0.4592", "0.752", "0.10455", "0.752" ], [ "0.5215", "0.7393", "0.91", "0.7393", "0.91", "0.75", "0.5215", "0.75" ], [ "0.1013", "0.752", "0.4617", "0.752", "0.4617", "0.765", "0.1013", "0.765" ], [ "0.5215", "0.752", "0.9126", "0.752", "0.9126", "0.765", "0.5215", "0.765" ], [ "0.1013", "0.7646", "0.4617", "0.7646", "0.4617", "0.778", "0.1013", "0.778" ], [ "0.5215", "0.765", "0.911", "0.765", "0.911", "0.779", "0.5215", "0.779" ], [ "0.10297", "0.779", "0.4617", "0.779", "0.4617", "0.789", "0.10297", "0.789" ], [ "0.5215", "0.779", "0.9126", "0.779", "0.9126", "0.7925", "0.5215", "0.7925" ], [ "0.1013", "0.789", "0.46", "0.789", "0.46", "0.803", "0.1013", "0.803" ], [ "0.5215", "0.7925", "0.9126", "0.7925", "0.9126", "0.806", "0.5215", "0.806" ], [ "0.10297", "0.8047", "0.4592", "0.8047", "0.4592", "0.815", "0.10297", "0.815" ], [ "0.5195", "0.806", "0.9126", "0.806", "0.9126", "0.8193", "0.5195", "0.8193" ], [ "0.1013", "0.815", "0.4631", "0.815", "0.4631", "0.829", "0.1013", "0.829" ], [ "0.5215", "0.821", "0.911", "0.821", "0.911", "0.834", "0.5215", "0.834" ], [ "0.10297", "0.83", "0.4617", "0.83", "0.4617", "0.8403", "0.10297", "0.8403" ], [ "0.5215", "0.834", "0.9126", "0.834", "0.9126", "0.8477", "0.5215", "0.8477" ], [ "0.10297", "0.8423", "0.46", "0.8423", "0.46", "0.853", "0.10297", "0.853" ], [ "0.5195", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5195", "0.8613" ], [ "0.10706", "0.8643", "0.246", "0.8643", "0.246", "0.8774", "0.10706", "0.8774" ], [ "0.5195", "0.8623", "0.9126", "0.8623", "0.9126", "0.8755", "0.5195", "0.8755" ], [ "0.08417", "0.877", "0.1993", "0.875", "0.1993", "0.888", "0.08417", "0.89" ], [ "0.5195", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5195", "0.8896" ] ]

[ "Efficient Deweather Mixture-of-Experts with Uncertainty-aware", "Feature-wise Linear Modulation", "Rongyu Zhang'?, Yulin Luo?, Jiaming Liu?, Huanrui Yang3, Zhen Dong , Denis Gudovskiy4", "Tomoyuki Okuno, Yohei Nakata,Kurt Keutzer3, Yuan Dul*, Shanghang Zhang?", "INanjing University 2National Key Laboratory for Multimedia Information Processing, School of Computer Science,", "Peking University 3University of California, Berkeley 4Panasonic", "yuandu@nju.edu.cn, shanghang@pku.edu.cn", "", "Upstream Task: Adverse Weather Removal", "Downstream Task: Classify./Segment./Detect", "2D Object Detection", "Difficulty:", "", "Difficulty:", "1.Efficient deployment", "will images optimized forhuman", "2.Diverse feature calibration", "perception also good for machines?", "", "", "", "[']", "", "Transformer-based MoEModel", "Recognition Model", "Clean Image", "AdverseWeatherImage", "Attention", "Convolutional", "RouterFeedforward network", "Task-specifi", "Tokens/imag", "N", "layer", "Instance Segmentation", "Figure 1: Deweather pipeline: upstream adverse weather removal with Mixture-of-Experts (MoE) and downstream instance", "head/tail", "segmentation and 2D object detection tasks.", "140199", "Abstract", "ventional MoE counterpart.Experiments on the downstream", "segmentationandclassificationtasksfurtherdemonstratethe", "The Mixture-of-Experts (MoE) approach has demonstrated", "generalizability of MoFME to real open-world applications", "outstanding scalability in multi-task learning including low-", "level upstream tasks such as concurrent removal of multi-", "ple adverse weather effects. However, the conventional MoE", "Introduction", "architecture withparallelFeed Forward Network (FFN)ex-", "perts leads to significant parameter and computational over-", "There is a growing interest in low-level upstream tasks such", "2", "heads that hinder its efficient deployment. In addition, the", "as adverse weather removal (deweather) (Valanarasu et al.", "2", "naive MoE linear router is suboptimal in assigning task-", "2022). It intends to eliminate the impact of weather-induced", "2", "3", "specific features to multiple experts which limits its further", "noise on decision-critical downstream tasks such as detec-", "C.AIXIE", "scalability. In this work, we propose an efficient MoE archi-", "tion and segmentation (Zamir et al. 2022). Previous meth-", "tecture with weight sharing across the experts. Inspired by", "ods (Ren et al. 2019; Chen et al. 2021) approach each type", "the idea of linear feature modulation (FM), our architecture", "of weather effect independently,yet multiple effects can ap-", "implicitly instantiates multiple experts via learnableactiva", "pear simultaneously in the real world. Moreover, such meth-", "tion modulations on a single shared expert block. The pro-", "ods mainly focus on the deweathering performance metrics", "posed Feature Modulated Expert (FME) serves as a building", "block for the novel Mixture-of-Feature-Modulation-Experts", "rather than an efficient deployment.", "(MoFME)architecture, which can scaleup the numberof ex-", "One promising way to address several weather effects", "perts withlow overhead.Wefurther proposean Uncertainty", "concurrentlyistheconditionalcomputationparadigm(Ben-", "aware Router (UaR) to assign task-specific features to dif-", "gio 2013), where a model can selectively activate certain", "ferent FM modules with well-calibrated weights. This en-", "parts of architecture, i.e. the task-specific experts, depend-", "ablesMoFMEtoeffectivelylearndiverseexpertfunctions", "ing on the input. In particular, the sparse Mixture-of-Experts", "for multiple tasks. The conducted experiments on the multi-", "(MoE) (Riquelme et al. 2021) with parallel Feed Forward", "deweathertask showthat our MoFMEoutperformsthe base", "Network (FFN) experts rely on a router to activate a sub-", "lines in the image restoration quality by 0.1-0.2 dB and", "set of FFNs for each weather-specific input image. Figure 1", "achievesSOTA-compatibleperformancewhilesavingmore", "shows a pipeline with an upstream MoE model to overcome", "than72%ofparametersand39%inferencetimeoverthecon-", "a number of weather effects. For example, Ye et al. (2022)", "*Corresponding author", "propose the DAN-Net method that estimates gated attention", "Copyright2024", "maps for inputs and uses them to properly dispatch images" ]

{ "arxiv_id": "2312.16610", "language": "en", "page_idx": 0 }

[ [ "0.1871", "0.125", "0.812", "0.125", "0.812", "0.1395", "0.1871", "0.1395" ], [ "0.4084", "0.159", "0.591", "0.159", "0.591", "0.1736", "0.4084", "0.1736" ], [ "0.4495", "0.178", "0.55", "0.178", "0.55", "0.1925", "0.4495", "0.1925" ], [ "0.3774", "0.1901", "0.6216", "0.1913", "0.6216", "0.2058", "0.3774", "0.2052" ], [ "0.0237", "0.26", "0.05963", "0.26", "0.05963", "0.3257", "0.0237", "0.3257" ], [ "0.2507", "0.2747", "0.3145", "0.2747", "0.3145", "0.289", "0.2507", "0.289" ], [ "0.5195", "0.2734", "0.91", "0.2727", "0.91", "0.2874", "0.5195", "0.2886" ], [ "0.5186", "0.2886", "0.911", "0.2874", "0.911", "0.3018", "0.5186", "0.303" ], [ "0.10297", "0.2966", "0.4631", "0.2966", "0.4631", "0.31", "0.10297", "0.31" ], [ "0.5195", "0.303", "0.9126", "0.303", "0.9126", "0.3164", "0.5195", "0.3164" ], [ "0.02533", "0.3132", "0.05963", "0.3145", "0.05637", "0.3726", "0.02287", "0.3718" ], [ "0.10455", "0.3113", "0.46", "0.3113", "0.46", "0.3213", "0.10455", "0.3213" ], [ "0.5195", "0.3164", "0.9126", "0.3164", "0.9126", "0.33", "0.5195", "0.33" ], [ "0.10297", "0.3225", "0.4617", "0.3225", "0.4617", "0.336", "0.10297", "0.336" ], [ "0.5195", "0.3289", "0.9126", "0.3289", "0.9126", "0.3435", "0.5195", "0.3435" ], [ "0.1013", "0.3352", "0.4617", "0.3352", "0.4617", "0.3484", "0.1013", "0.3484" ], [ "0.1013", "0.348", "0.4617", "0.348", "0.4617", "0.361", "0.1013", "0.361" ], [ "0.5195", "0.3447", "0.911", "0.3447", "0.911", "0.358", "0.5195", "0.358" ], [ "0.1013", "0.3599", "0.4631", "0.3599", "0.4631", "0.3738", "0.1013", "0.3738" ], [ "0.5215", "0.358", "0.911", "0.358", "0.911", "0.3718", "0.5215", "0.3718" ], [ "0.02287", "0.3655", "0.05392", "0.3623", "0.058", "0.3901", "0.02696", "0.3933" ], [ "0.1013", "0.3706", "0.4631", "0.3718", "0.4631", "0.3865", "0.1013", "0.385" ], [ "0.5195", "0.3706", "0.9126", "0.3706", "0.9126", "0.3838", "0.5195", "0.3838" ], [ "0.10297", "0.385", "0.4631", "0.385", "0.4631", "0.3984", "0.10297", "0.3984" ], [ "0.5215", "0.385", "0.9126", "0.385", "0.9126", "0.3984", "0.5215", "0.3984" ], [ "0.1013", "0.3984", "0.4617", "0.3977", "0.4617", "0.411", "0.1013", "0.4124" ], [ "0.5215", "0.3984", "0.9126", "0.3984", "0.9126", "0.4136", "0.5215", "0.4136" ], [ "0.02287", "0.4048", "0.05963", "0.4048", "0.05963", "0.455", "0.02287", "0.455" ], [ "0.10297", "0.4104", "0.4631", "0.4104", "0.4631", "0.4236", "0.10297", "0.4236" ], [ "0.5195", "0.4136", "0.914", "0.4136", "0.914", "0.4268", "0.5195", "0.4268" ], [ "0.10297", "0.4236", "0.4631", "0.4236", "0.4631", "0.4375", "0.10297", "0.4375" ], [ "0.5195", "0.428", "0.875", "0.428", "0.875", "0.4414", "0.5195", "0.4414" ], [ "0.1054", "0.4375", "0.2303", "0.4375", "0.2303", "0.4475", "0.1054", "0.4475" ], [ "0.255", "0.4375", "0.46", "0.4375", "0.46", "0.4475", "0.255", "0.4475" ], [ "0.536", "0.443", "0.9126", "0.443", "0.9126", "0.457", "0.536", "0.457" ], [ "0.10297", "0.4475", "0.4617", "0.4475", "0.4617", "0.461", "0.10297", "0.461" ], [ "0.02777", "0.443", "0.05963", "0.443", "0.05963", "0.486", "0.02777", "0.486" ], [ "0.10297", "0.4602", "0.4631", "0.4602", "0.4631", "0.4734", "0.10297", "0.4734" ], [ "0.5195", "0.4583", "0.911", "0.4583", "0.911", "0.4717", "0.5195", "0.4717" ], [ "0.10297", "0.473", "0.4631", "0.473", "0.4631", "0.486", "0.10297", "0.486" ], [ "0.5215", "0.4717", "0.9126", "0.4717", "0.9126", "0.486", "0.5215", "0.486" ], [ "0.10455", "0.4873", "0.4617", "0.4873", "0.4617", "0.4976", "0.10455", "0.4976" ], [ "0.5195", "0.4849", "0.91", "0.4849", "0.91", "0.4988", "0.5195", "0.4988" ], [ "0.02287", "0.4912", "0.06046", "0.4912", "0.06046", "0.5986", "0.02287", "0.5986" ], [ "0.1013", "0.4988", "0.3882", "0.4988", "0.3882", "0.512", "0.1013", "0.512" ], [ "0.5186", "0.4988", "0.9126", "0.4976", "0.9126", "0.512", "0.5186", "0.513" ], [ "0.5215", "0.513", "0.9126", "0.513", "0.9126", "0.5264", "0.5215", "0.5264" ], [ "0.2361", "0.531", "0.3513", "0.532", "0.3513", "0.547", "0.2361", "0.5454" ], [ "0.5215", "0.5264", "0.914", "0.5264", "0.914", "0.5405", "0.5215", "0.5405" ], [ "0.5215", "0.5405", "0.9126", "0.5405", "0.9126", "0.5537", "0.5215", "0.5537" ], [ "0.2141", "0.532", "0.2418", "0.532", "0.2418", "0.5444", "0.2141", "0.5444" ], [ "0.0866", "0.5503", "0.4795", "0.5503", "0.4795", "0.5645", "0.0866", "0.5645" ], [ "0.5195", "0.553", "0.9126", "0.5537", "0.9126", "0.5684", "0.5195", "0.5674" ], [ "0.08496", "0.563", "0.478", "0.5645", "0.478", "0.579", "0.08496", "0.5776" ], [ "0.5215", "0.5674", "0.9126", "0.5674", "0.9126", "0.5806", "0.5215", "0.5806" ], [ "0.0866", "0.579", "0.4795", "0.579", "0.4795", "0.5923", "0.0866", "0.5923" ], [ "0.5195", "0.5835", "0.9126", "0.5835", "0.9126", "0.5967", "0.5195", "0.5967" ], [ "0.08496", "0.5923", "0.4795", "0.5923", "0.4795", "0.606", "0.08496", "0.606" ], [ "0.02124", "0.584", "0.05963", "0.584", "0.05963", "0.711", "0.02124", "0.711" ], [ "0.5215", "0.5967", "0.9126", "0.5967", "0.9126", "0.61", "0.5215", "0.61" ], [ "0.08496", "0.607", "0.4795", "0.607", "0.4795", "0.6206", "0.08496", "0.6206" ], [ "0.5215", "0.61", "0.911", "0.61", "0.911", "0.6235", "0.5215", "0.6235" ], [ "0.08496", "0.6206", "0.4763", "0.6206", "0.4763", "0.634", "0.08496", "0.634" ], [ "0.5195", "0.6235", "0.9126", "0.6235", "0.9126", "0.637", "0.5195", "0.637" ], [ "0.08496", "0.634", "0.478", "0.634", "0.478", "0.648", "0.08496", "0.648" ], [ "0.5195", "0.638", "0.9126", "0.638", "0.9126", "0.6514", "0.5195", "0.6514" ], [ "0.08496", "0.648", "0.478", "0.6465", "0.478", "0.661", "0.08496", "0.662" ], [ "0.5195", "0.6514", "0.898", "0.6514", "0.898", "0.6655", "0.5195", "0.6655" ], [ "0.08496", "0.661", "0.478", "0.661", "0.478", "0.675", "0.08496", "0.675" ], [ "0.08417", "0.674", "0.478", "0.675", "0.478", "0.6895", "0.08417", "0.688" ], [ "0.5376", "0.6685", "0.9126", "0.6685", "0.9126", "0.6816", "0.5376", "0.6816" ], [ "0.5195", "0.6816", "0.9126", "0.6816", "0.9126", "0.696", "0.5195", "0.696" ], [ "0.08496", "0.6895", "0.478", "0.6895", "0.478", "0.7026", "0.08496", "0.7026" ], [ "0.5195", "0.696", "0.914", "0.696", "0.914", "0.709", "0.5195", "0.709" ], [ "0.0866", "0.7026", "0.4795", "0.7026", "0.4795", "0.7163", "0.0866", "0.7163" ], [ "0.5195", "0.7104", "0.911", "0.7104", "0.911", "0.7236", "0.5195", "0.7236" ], [ "0.08826", "0.7183", "0.478", "0.7183", "0.478", "0.729", "0.08826", "0.729" ], [ "0.5215", "0.7236", "0.9126", "0.7236", "0.9126", "0.7373", "0.5215", "0.7373" ], [ "0.08417", "0.73", "0.3945", "0.73", "0.3945", "0.7437", "0.08417", "0.7437" ], [ "0.5195", "0.738", "0.9126", "0.738", "0.9126", "0.752", "0.5195", "0.752" ], [ "0.10297", "0.744", "0.478", "0.744", "0.478", "0.7583", "0.10297", "0.7583" ], [ "0.5215", "0.752", "0.911", "0.752", "0.911", "0.765", "0.5215", "0.765" ], [ "0.0866", "0.7583", "0.478", "0.7583", "0.478", "0.7715", "0.0866", "0.7715" ], [ "0.5195", "0.765", "0.91", "0.765", "0.91", "0.779", "0.5195", "0.779" ], [ "0.0866", "0.773", "0.4763", "0.773", "0.4763", "0.786", "0.0866", "0.786" ], [ "0.08496", "0.786", "0.4756", "0.786", "0.4756", "0.8", "0.08496", "0.8" ], [ "0.5215", "0.781", "0.911", "0.781", "0.911", "0.7915", "0.5215", "0.7915" ], [ "0.5215", "0.7925", "0.911", "0.7925", "0.911", "0.806", "0.5215", "0.806" ], [ "0.08496", "0.8003", "0.4795", "0.8003", "0.4795", "0.8145", "0.08496", "0.8145" ], [ "0.5195", "0.8066", "0.914", "0.8066", "0.914", "0.821", "0.5195", "0.821" ], [ "0.08496", "0.8145", "0.478", "0.8145", "0.478", "0.8276", "0.08496", "0.8276" ], [ "0.5215", "0.821", "0.914", "0.821", "0.914", "0.834", "0.5215", "0.834" ], [ "0.08496", "0.8276", "0.478", "0.8276", "0.478", "0.8413", "0.08496", "0.8413" ], [ "0.5195", "0.834", "0.9126", "0.834", "0.9126", "0.8477", "0.5195", "0.8477" ], [ "0.08496", "0.8413", "0.478", "0.8413", "0.478", "0.855", "0.08496", "0.855" ], [ "0.5186", "0.8477", "0.911", "0.8467", "0.911", "0.8613", "0.5186", "0.8623" ], [ "0.0866", "0.863", "0.4795", "0.863", "0.4795", "0.877", "0.0866", "0.877" ], [ "0.5195", "0.8623", "0.914", "0.8623", "0.914", "0.8755", "0.5195", "0.8755" ], [ "0.0866", "0.877", "0.3718", "0.877", "0.3718", "0.89", "0.0866", "0.89" ], [ "0.5186", "0.8755", "0.746", "0.875", "0.746", "0.8896", "0.5186", "0.89" ] ]

[ "Inverse Reinforcement Learning with Natural Language Goals", "Li Zhou, Kevin Small", "Amazon Alexa", "[lizhouml, smakevin) @amazon.com", "00", "Abstract", "video game AI (Tucker, Gleave, and Russell 2018). Gen-", "erally, task goals are specified intrinsically by the environ-", "Humans generally use natural language (NL) to communicate", "ment and an agent is trained for each task. To generalize", "", "task requirementsto eachother.Ideally,NL should also beus-", "the learned policy to new goals, goal-conditioned imitation", "able for communicating goals to autonomous machines (e.g.", "learning (Ding et al. 2019) and reinforcement learning (RL)", "robots) to minimize friction in task specification. However,", "understanding and mapping NL goals to sequences of states", "algorithms (Kaelbling 1993; Schaul et al. 2015; Nair et al.", "and actions is challenging. Specifically, existing work along", "2018) have been proposed where the policy is explicitly con-", "6", "these lines has encountered difficulty in generalizing learned", "ditioned on a goal. Normally, the goals either share the same", "policies to new NL goals and environments. In this paper, we", "space with the states or can be easily mapped to the state", "propose a novel adversarial inverse reinforcement learning al-", "space. For example, in Ding et al. (2019), the goals and", "[OT'", "gorithm to learn a language-conditioned policy and reward", "states are both coordinates in the environment and the goals", "function. To improve generalization of the learned policy and", "are provided to the agent by specifying goal positions.", "rewardfunction,we", "variational goalgeneratortorela-", "Fu et al. (2019) and Bahdanau et al. (2019) represent", "bel trajectories and sample diverse goals during training. Our", "I'so]", "algorithm outperforms multiple baselines by a large margin", "two recent related works from the perspective of under-", "on a vision-based NL instruction following dataset (Room-2-", "standing NL goal specifications, where they propose learn-", "Room),demonstratingapromisingadvanceinenablingthe", "ing a language-conditioned reward function under the max-", "2690", "use of NL instructions in specifying agent goals.", "imum entropy inverse reinforcement learning (Ziebart et al.", "2008) and generative adversarial imitation learning frame-", "Introduction", "works (Ho and Ermon 2016). However, the NL goals are", "produced by a preset template-based grammar. Fu et al.", "1", "Humans use natural language (NL) to communicate with", "(2019)'s policy is optimized exactly in a grid environment", "each other. Accordingly, a desired method for conveying", "with known dynamics and when the policy is optimized", "goals or assigning tasks to an autonomous machine (e.g., a", "by sample-based reinforcement learning algorithms such as", "robot) is also by NL. For example, in household tasks, when", "8000:AX", "deep Q-learning (Mnih et al. 2015), the model performance", "asking an agent to pick up a toolbox from the garage, we", "drops significantly. This reveals a sample efficiency chal-", "do not assign the coordinates of the toolbox to the robot.", "lenge in this setting and implies that when the NL goals", "Instead, we state to the agent retrieve my toolbox from the", "and the environments are complicated, generalization to new", "garage. This requires the agent to understand the semantics", "goals and environments becomes a very difficult problem.", "of the natural language goals, associate states and actions", "with the natural language goals, and infer whether the nat-", "One potential method for addressing these shortcomings", "are goal relabeling techniques such as hindsight experience", "ural language goals are achieved or not - all of which are", "replay (HER) (Andrychowicz et al. 2017) and latent goal", "very challenging tasks. Furthermore, we want to be able to", "relabeling (Nair et al. 2018), which have been shown to im-", "learnpoliciesthatarerobusttolexicalvariationingoalsand", "prove sample efficiency in RL settings. However, when the", "generalize well to new goals and environments.", "goals are NL and are in a different space than the state space,", "Inverse reinforcement learning (IRL) (Abbeel and Ng", "goal relabeling cannot be applied directly as we cannot eas-", "2004; Fu, Luo, and Levine 2018), a form of imitation learn-", "ily relabel a state to a NL goal. Cideron et al. (2019) pro-", "ing (Osa et al. 2018), is the task of learning a reward func-", "tion and hence a policy based on expert demonstrations", "posestobuildsucharelabelingfunctionwithaSEQ2SEQ", "model that takes a trajectory as input and outputs a rela-", "Imitation learning has been successfully applied to a wide", "beled NL goal. However, this uses the ground-truth reward", "range of tasks including robot manipulation (Finn, Levine,", "function and their experiments are again based on simple", "and Abbeel 2016), autonomous driving (Kuefler et al. 2017),", "template-based NL goals. Moreover, as we will show in this", "human behavior forecasting (Rhinehart and Kitani 2017),", "paper, applying HER to IRL with NL goals doesn't signifi-", "Copyright 2021, Association for the Advancement of Artificial", "cantly improve performance, as the reward function does not", "Intelligence (www.aaai.org).All rights reserved.", "generalize well to relabeled goals." ]

{ "arxiv_id": "2008.06924", "language": "en", "page_idx": 0 }

[ [ "0.10455", "0.06946", "0.478", "0.06946", "0.478", "0.0833", "0.10455", "0.0833" ], [ "0.5195", "0.0682", "0.9126", "0.06946", "0.9126", "0.0833", "0.5195", "0.0821" ], [ "0.08826", "0.0852", "0.4756", "0.0852", "0.4756", "0.09534", "0.08826", "0.09534" ], [ "0.5195", "0.084", "0.898", "0.084", "0.898", "0.09784", "0.5195", "0.09784" ], [ "0.0866", "0.09784", "0.478", "0.09784", "0.478", "0.1111", "0.0866", "0.1111" ], [ "0.706", "0.1061", "0.834", "0.1061", "0.834", "0.1206", "0.706", "0.1206" ], [ "0.0866", "0.11237", "0.4795", "0.11237", "0.4795", "0.1256", "0.0866", "0.1256" ], [ "0.568", "0.11615", "0.6724", "0.11615", "0.6724", "0.1288", "0.568", "0.1288" ], [ "0.892", "0.11554", "0.9126", "0.11554", "0.9126", "0.1276", "0.892", "0.1276" ], [ "0.08496", "0.125", "0.478", "0.1237", "0.478", "0.1383", "0.08496", "0.1395" ], [ "0.6733", "0.1237", "0.865", "0.1237", "0.865", "0.1383", "0.6733", "0.1383" ], [ "0.0866", "0.1395", "0.478", "0.1395", "0.478", "0.1528", "0.0866", "0.1528" ], [ "0.5215", "0.1465", "0.9126", "0.1465", "0.9126", "0.161", "0.5215", "0.161" ], [ "0.08496", "0.1528", "0.4763", "0.1528", "0.4763", "0.1666", "0.08496", "0.1666" ], [ "0.5195", "0.161", "0.911", "0.161", "0.911", "0.1749", "0.5195", "0.1749" ], [ "0.08905", "0.1686", "0.4763", "0.1686", "0.4763", "0.1787", "0.08905", "0.1787" ], [ "0.5195", "0.1749", "0.9126", "0.1749", "0.9126", "0.1881", "0.5195", "0.1881" ], [ "0.08496", "0.1812", "0.478", "0.1812", "0.478", "0.1945", "0.08496", "0.1945" ], [ "0.5195", "0.1881", "0.652", "0.1881", "0.652", "0.202", "0.5195", "0.202" ], [ "0.0866", "0.1945", "0.4763", "0.1945", "0.4763", "0.2084", "0.0866", "0.2084" ], [ "0.08826", "0.2102", "0.478", "0.2102", "0.478", "0.2203", "0.08826", "0.2203" ], [ "0.5215", "0.2153", "0.556", "0.2153", "0.556", "0.2279", "0.5215", "0.2279" ], [ "0.5522", "0.214", "0.911", "0.214", "0.911", "0.2291", "0.5522", "0.2291" ], [ "0.0866", "0.2229", "0.4795", "0.2229", "0.4795", "0.2361", "0.0866", "0.2361" ], [ "0.08496", "0.2361", "0.478", "0.2361", "0.478", "0.25", "0.08496", "0.25" ], [ "0.5596", "0.2311", "0.689", "0.2311", "0.689", "0.2443", "0.5596", "0.2443" ], [ "0.0866", "0.252", "0.4763", "0.252", "0.4763", "0.262", "0.0866", "0.262" ], [ "0.5195", "0.2474", "0.914", "0.2474", "0.914", "0.2615", "0.5195", "0.2615" ], [ "0.08496", "0.2632", "0.478", "0.2632", "0.478", "0.2766", "0.08496", "0.2766" ], [ "0.5195", "0.262", "0.9126", "0.262", "0.9126", "0.276", "0.5195", "0.276" ], [ "0.0866", "0.2778", "0.4795", "0.2778", "0.4795", "0.2917", "0.0866", "0.2917" ], [ "0.5186", "0.2747", "0.9126", "0.276", "0.9126", "0.2903", "0.5186", "0.289" ], [ "0.0866", "0.2917", "0.3906", "0.2917", "0.3906", "0.305", "0.0866", "0.305" ], [ "0.5195", "0.2903", "0.9126", "0.2903", "0.9126", "0.3037", "0.5195", "0.3037" ], [ "0.5215", "0.3062", "0.911", "0.3062", "0.911", "0.3164", "0.5215", "0.3164" ], [ "0.1732", "0.3193", "0.3882", "0.3213", "0.3882", "0.336", "0.1732", "0.334" ], [ "0.5215", "0.318", "0.91", "0.318", "0.91", "0.332", "0.5215", "0.332" ], [ "0.5186", "0.332", "0.8286", "0.3308", "0.8286", "0.3452", "0.5186", "0.3467" ], [ "0.0866", "0.3435", "0.4756", "0.3435", "0.4756", "0.3567", "0.0866", "0.3567" ], [ "0.08496", "0.3567", "0.4795", "0.3567", "0.4795", "0.3706", "0.08496", "0.3706" ], [ "0.585", "0.3528", "0.8447", "0.3516", "0.8447", "0.3694", "0.585", "0.3706" ], [ "0.08417", "0.3694", "0.4795", "0.3718", "0.4795", "0.385", "0.08417", "0.3833" ], [ "0.6006", "0.3718", "0.8203", "0.3718", "0.8203", "0.3865", "0.6006", "0.3865" ], [ "0.08496", "0.385", "0.4756", "0.385", "0.4756", "0.3984", "0.08496", "0.3984" ], [ "0.08496", "0.3984", "0.4795", "0.3984", "0.4795", "0.4124", "0.08496", "0.4124" ], [ "0.5215", "0.3945", "0.911", "0.3945", "0.911", "0.4092", "0.5215", "0.4092" ], [ "0.0866", "0.414", "0.478", "0.414", "0.478", "0.4248", "0.0866", "0.4248" ], [ "0.5215", "0.4092", "0.911", "0.4104", "0.911", "0.4236", "0.5215", "0.4224" ], [ "0.0866", "0.4268", "0.478", "0.4268", "0.478", "0.44", "0.0866", "0.44" ], [ "0.5215", "0.4236", "0.8965", "0.4236", "0.8965", "0.4375", "0.5215", "0.4375" ], [ "0.8945", "0.428", "0.9062", "0.428", "0.9062", "0.432", "0.8945", "0.432" ], [ "0.0866", "0.4426", "0.4756", "0.4426", "0.4756", "0.4526", "0.0866", "0.4526" ], [ "0.523", "0.4375", "0.914", "0.4375", "0.914", "0.4507", "0.523", "0.4507" ], [ "0.08496", "0.455", "0.4795", "0.455", "0.4795", "0.4685", "0.08496", "0.4685" ], [ "0.5195", "0.4507", "0.914", "0.4507", "0.914", "0.4653", "0.5195", "0.4653" ], [ "0.08496", "0.4666", "0.4795", "0.4673", "0.4795", "0.4817", "0.08496", "0.481" ], [ "0.5195", "0.464", "0.911", "0.464", "0.911", "0.4778", "0.5195", "0.4778" ], [ "0.08496", "0.4817", "0.4763", "0.4817", "0.4763", "0.4956", "0.08496", "0.4956" ], [ "0.5195", "0.4778", "0.91", "0.4778", "0.91", "0.4912", "0.5195", "0.4912" ], [ "0.0866", "0.4976", "0.4763", "0.4976", "0.4763", "0.5083", "0.0866", "0.5083" ], [ "0.5195", "0.4924", "0.9126", "0.4924", "0.9126", "0.506", "0.5195", "0.506" ], [ "0.08417", "0.5103", "0.478", "0.509", "0.478", "0.523", "0.08417", "0.5234" ], [ "0.5186", "0.506", "0.914", "0.505", "0.914", "0.5195", "0.5186", "0.521" ], [ "0.08417", "0.523", "0.4756", "0.5234", "0.4756", "0.537", "0.08417", "0.536" ], [ "0.5215", "0.5215", "0.9126", "0.5215", "0.9126", "0.532", "0.5215", "0.532" ], [ "0.08417", "0.537", "0.478", "0.536", "0.478", "0.55", "0.08417", "0.5503" ], [ "0.5215", "0.5327", "0.914", "0.5327", "0.914", "0.547", "0.5215", "0.547" ], [ "0.08417", "0.5503", "0.478", "0.55", "0.478", "0.5645", "0.08417", "0.565" ], [ "0.5195", "0.5474", "0.7686", "0.547", "0.7686", "0.561", "0.5195", "0.5625" ], [ "0.0866", "0.5664", "0.4756", "0.5664", "0.4756", "0.577", "0.0866", "0.577" ], [ "0.5513", "0.5693", "0.88", "0.5693", "0.88", "0.584", "0.5513", "0.584" ], [ "0.08496", "0.579", "0.4763", "0.579", "0.4763", "0.5923", "0.08496", "0.5923" ], [ "0.0866", "0.5947", "0.4763", "0.5947", "0.4763", "0.605", "0.0866", "0.605" ], [ "0.5195", "0.5933", "0.91", "0.5933", "0.91", "0.607", "0.5195", "0.607" ], [ "0.08496", "0.607", "0.4795", "0.607", "0.4795", "0.6206", "0.08496", "0.6206" ], [ "0.5215", "0.608", "0.911", "0.608", "0.911", "0.622", "0.5215", "0.622" ], [ "0.08496", "0.6206", "0.4795", "0.6206", "0.4795", "0.634", "0.08496", "0.634" ], [ "0.5215", "0.622", "0.9126", "0.622", "0.9126", "0.6353", "0.5215", "0.6353" ], [ "0.08496", "0.634", "0.478", "0.634", "0.478", "0.648", "0.08496", "0.648" ], [ "0.5195", "0.6353", "0.911", "0.6353", "0.911", "0.6484", "0.5195", "0.6484" ], [ "0.08417", "0.6484", "0.3694", "0.6484", "0.3694", "0.662", "0.08417", "0.662" ], [ "0.5215", "0.651", "0.91", "0.651", "0.91", "0.661", "0.5215", "0.661" ], [ "0.5215", "0.664", "0.911", "0.664", "0.911", "0.675", "0.5215", "0.675" ], [ "0.232", "0.6753", "0.3308", "0.678", "0.3308", "0.6934", "0.232", "0.6914" ], [ "0.5195", "0.677", "0.9126", "0.677", "0.9126", "0.69", "0.5195", "0.69" ], [ "0.5215", "0.6924", "0.911", "0.6924", "0.911", "0.7026", "0.5215", "0.7026" ], [ "0.08496", "0.7007", "0.3816", "0.7007", "0.3816", "0.7153", "0.08496", "0.7153" ], [ "0.5215", "0.7056", "0.9126", "0.7056", "0.9126", "0.7163", "0.5215", "0.7163" ], [ "0.1274", "0.7163", "0.3774", "0.7163", "0.3774", "0.731", "0.1274", "0.731" ], [ "0.523", "0.7183", "0.9126", "0.7183", "0.9126", "0.732", "0.523", "0.732" ], [ "0.08826", "0.7363", "0.4756", "0.7363", "0.4756", "0.75", "0.08826", "0.75" ], [ "0.5215", "0.732", "0.9126", "0.732", "0.9126", "0.7456", "0.5215", "0.7456" ], [ "0.08496", "0.7505", "0.478", "0.75", "0.478", "0.7646", "0.08496", "0.765" ], [ "0.5215", "0.7476", "0.911", "0.7476", "0.911", "0.7583", "0.5215", "0.7583" ], [ "0.0866", "0.765", "0.4763", "0.765", "0.4763", "0.779", "0.0866", "0.779" ], [ "0.5215", "0.7603", "0.911", "0.7603", "0.911", "0.7734", "0.5215", "0.7734" ], [ "0.08496", "0.779", "0.4763", "0.779", "0.4763", "0.7935", "0.08496", "0.7935" ], [ "0.5195", "0.7734", "0.8823", "0.7734", "0.8823", "0.787", "0.5195", "0.787" ], [ "0.08496", "0.7925", "0.2712", "0.7925", "0.2712", "0.803", "0.08496", "0.803" ], [ "0.2712", "0.7925", "0.4795", "0.7925", "0.4795", "0.806", "0.2712", "0.806" ], [ "0.5215", "0.8003", "0.55", "0.8003", "0.55", "0.813", "0.5215", "0.813" ], [ "0.556", "0.799", "0.8965", "0.8", "0.8965", "0.8145", "0.556", "0.813" ], [ "0.0866", "0.806", "0.478", "0.806", "0.478", "0.8193", "0.0866", "0.8193" ], [ "0.5605", "0.815", "0.8486", "0.815", "0.8486", "0.829", "0.5605", "0.829" ], [ "0.08826", "0.8213", "0.4756", "0.8213", "0.4756", "0.832", "0.08826", "0.832" ], [ "0.08496", "0.832", "0.4763", "0.834", "0.4763", "0.8486", "0.08496", "0.8467" ], [ "0.5215", "0.832", "0.91", "0.834", "0.91", "0.8477", "0.5215", "0.845" ], [ "0.0866", "0.8477", "0.4763", "0.8477", "0.4763", "0.8613", "0.0866", "0.8613" ], [ "0.5215", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5215", "0.8613" ], [ "0.0866", "0.8613", "0.478", "0.8613", "0.478", "0.875", "0.0866", "0.875" ], [ "0.5195", "0.8623", "0.911", "0.8623", "0.911", "0.8755", "0.5195", "0.8755" ], [ "0.08496", "0.8755", "0.4795", "0.8755", "0.4795", "0.8896", "0.08496", "0.8896" ], [ "0.5195", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5195", "0.8896" ] ]

[ "In this work, we propose a sample efficient IRL algorithm", "(GAN) (Goodfellow et al. 2014). The GAN generator is the", "withnaturallanguage(NL)goals.Tothebestof ourknowl-", "policy π to learn and the GAN discriminator is defined as", "edge, our work is the first IRL algorithm that works with real", "exp(fe,s(s, a, s'))", "human generated NL goals (as opposed to template-based", "Do,g(s,a, s) =", "(1)", "language goals (Fu, Luo, and Levine 2018; Bahdanau et al.", "exp(fo,s(s, a, s)) +π(as)", "2019) in a real world vision-based environment. Our con-", "where foe,s(s,a, s) = ge(s,a) +he(s') -he(s), ge(s,a)", "tributions include: (1) proposing a NL goal-conditioned ad-", "is the reward function to learn, and h(s') - h(s) is the", "versarialinversereinforcementlearningalgorithm,(2)spec-", "reward shaping term. The policy and the discriminator are", "ifying a variational goal generator that efficiently sample di-", "updated alternately.", "verse NL goals given a trajectory, (3) utilizing goal relabel-", "ingand sampling strategiestoimprovethegeneralizationof", "3.2", "Inverse Reinforcement Learning with Natural", "both the policy and the reward function to new NL goals and", "new environments, (4) proposing a self-supervised learning", "Language Goals", "mechanismtofurtherimprovethegeneralizationinnewen-", "Our AIRL-based solution builds on language-conditioned", "vironments. Through these innovations, we show that our", "reward learning (LC-RL) (Fu et al. 2019) by adding", "algorithm significantly outperform strong baselines on the", "template-free NL capabilities and continuous state space", "Room-2-Room dataset (Anderson et al. 2018).", "support. In this problem setting, the policy and the reward", "functionareconditioned onaNLgoalG,sowefirstextend", "2 Problem Formulation", "the discriminator in Equation (1) to have fe,(s, a, s', G) ", "ge(s, a, G) + hs(s',G) - h(s, G) such that", "A task is defined as a pair (E,G), where E is an environ-", "ment that the agent can interact with and G is a NL goal", "go(s, a, G) = MLP([e°; Att(s, G); e\"))", "that the agent has to fulfill. G = {wi, w2,.., wn) consists", "hs(s, G) = MLP([e’; Att(s, G))", "of N words. For example, in our experiments, E is a re", "alistic 3D indoor environment and G is a human-generated", "where MLP() is a multilayer Perceptron, es and ea", "navigationinstruction.Thetruerewardfunctionofeachtask", "are the embeddings of state s and action a, respec-", "is unknown, but there exists a set of expert demonstrations", "tively, and Att(s, G) is an attention function. Att(s, G)", "-", "consisting of a task (E, G)and a trajectory T.The trajec", "aiewwhere ew is the word embedding of wi in", "tory T= {si,a1, S2, a2,..., ST, aT] consists of a sequence", "G, αi = (Linear(es)·e\") /(Z, Linear(es)-ew) and", "of T states perceived by the experts and the actions taken", "Linear() is a single-layer Perceptron. We use soft actor-", "by the experts given the states. For example, in our experi-", "critic (SAC) (Haarnoja et al. 2018), one of the state-of-the-", "ments the statesarefirst-personviewsin a3D indoorenvi-", "art off-policy RL algorithms, to optimize policy given the", "ronment and the actions are movements towards a direction", "reward function ge(s, a, G). SAC includes a policy network", "in the environment.While our proposed methods can be ap", "topredictan actiongivena stateand agoal,andaQ-network", "pliedtoboth continuousanddiscreteactionspaces,wefocus", "that estimate the Q-value of an action given a state and a", "on discrete action spaces in this paper. The objective of the", "goal. We define the policy network as", "algorithmistolearnapolicythatimitates expertdemonstra-", "Tw(s, a, G) = Softmax(ea - MLP([e'; Att(s, G)D))", "tions such that given a new NL goal in an existing or new", "environment,the agent can performa sequence of actions", "and we define the Q-network qu(s, a, G) as the same net", "to achieve that goal. In this paper, we use G to represent", "work architecture as ge. Compared with on-policy algo-", "an arbitrary goal, G to represent the set of goals in expert", "rithms such as TRPO (Schulman et al. 2015), SAC uti-", "demonstrations, and G, to represent the i-th goal in G. The", "lizes a replay buffer to re-use sampled trajectories. The re-", " same notation style applies to E and too.", "play buffer is beneficial to training both the discrimina-", "tor and the policy. To update the discriminator, we sam-", "3Method", "ple negative (s, a, s',G) examples from the replay buffer", "and sample positive (s,a,s,G)examples from the expert", "3.1 Preliminary: Adversarial Inverse", "demonstrations.To updatethe policy,we sample a batchof", "Reinforcement Learning (AIRL)", "(s, a, s', G) from the replay buffer and use ge to estimate", "AIRL (Fu, Luo, and Levine 2018) is based on maximum en-", "their rewards; then we update the Q- and policy network", "tropy inverse reinforcement learning (MaxEntIRL) (Ziebart", "using these reward-augmented samples.We modify SAC", "et al. 2008). In MaxEntIRL, the probability of a trajectory", "slightly to support discrete action spaces. Appendix B con-", "is defined as pe() = z exp(fe(→)), where fe is the reward", "tains details about model architecture and optimization.", "function to learn and Z =", ", exp(fe()) is the partition", "3.3", "A Variational Goal Generator for Improved", "function. MaxEntIRL learns the reward function from ex-", "Generalization and Sample Efficiency", "pertdemonstrationsbymaximizingthelikelihoodof trajec", "tories in the expert demonstrations: maxg E-~(log Pe(-)", "While SAC hasbetter sampleefficiencythanmanyon-", "Maximizing this likelihood is challenging because the par", "policy RL algorithms, as we will show in our experiments,", "tition function Z is hard to estimate. AIRL maximizes", "in environments with complicated NL goals and high dimen-", "this likelihood by using a generative adversarial network", "sional state spaces (e.g., vision-based instruction following)," ]

{ "arxiv_id": "2008.06924", "language": "en", "page_idx": 1 }

[ [ "0.0866", "0.06946", "0.4795", "0.06946", "0.4795", "0.0833", "0.0866", "0.0833" ], [ "0.5215", "0.07135", "0.9087", "0.07135", "0.9087", "0.0821", "0.5215", "0.0821" ], [ "0.08496", "0.084", "0.3022", "0.0821", "0.3022", "0.09534", "0.08496", "0.09784" ], [ "0.5195", "0.084", "0.914", "0.084", "0.914", "0.09784", "0.5195", "0.09784" ], [ "0.0866", "0.1016", "0.4795", "0.1016", "0.4795", "0.11554", "0.0866", "0.11554" ], [ "0.5195", "0.09784", "0.841", "0.09784", "0.841", "0.1111", "0.5195", "0.1111" ], [ "0.1111", "0.11426", "0.4795", "0.11426", "0.4795", "0.1276", "0.1111", "0.1276" ], [ "0.539", "0.113", "0.911", "0.113", "0.911", "0.1237", "0.539", "0.1237" ], [ "0.1111", "0.127", "0.3635", "0.127", "0.3635", "0.1401", "0.1111", "0.1401" ], [ "0.5215", "0.1256", "0.9126", "0.1256", "0.9126", "0.1395", "0.5215", "0.1395" ], [ "0.09235", "0.1401", "0.2091", "0.1401", "0.2091", "0.154", "0.09235", "0.154" ], [ "0.5215", "0.1395", "0.9126", "0.1395", "0.9126", "0.1528", "0.5215", "0.1528" ], [ "0.08905", "0.1528", "0.2876", "0.1515", "0.2876", "0.1666", "0.08905", "0.1674" ], [ "0.5215", "0.154", "0.914", "0.154", "0.914", "0.1674", "0.5215", "0.1674" ], [ "0.1316", "0.1654", "0.259", "0.1654", "0.259", "0.1787", "0.1316", "0.1787" ], [ "0.5195", "0.1674", "0.9126", "0.1674", "0.9126", "0.1812", "0.5195", "0.1812" ], [ "0.1528", "0.1799", "0.3594", "0.1787", "0.3594", "0.1901", "0.1528", "0.1913" ], [ "0.08826", "0.1704", "0.1013", "0.1642", "0.10864", "0.1736", "0.0948", "0.1799" ], [ "0.09235", "0.1812", "0.10706", "0.1812", "0.10706", "0.1913", "0.09235", "0.1913" ], [ "0.5195", "0.1812", "0.911", "0.1812", "0.911", "0.1945", "0.5195", "0.1945" ], [ "0.0907", "0.1901", "0.10864", "0.1901", "0.10864", "0.2052", "0.0907", "0.2052" ], [ "0.1528", "0.1901", "0.3162", "0.1913", "0.3162", "0.2058", "0.1528", "0.2052" ], [ "0.5186", "0.1932", "0.9126", "0.1945", "0.9126", "0.209", "0.5186", "0.2084" ], [ "0.0907", "0.2039", "0.10864", "0.2039", "0.10864", "0.2172", "0.0907", "0.2172" ], [ "0.134", "0.2052", "0.1814", "0.2052", "0.1814", "0.2166", "0.134", "0.2166" ], [ "0.0989", "0.2134", "0.1111", "0.2229", "0.10046", "0.2317", "0.0866", "0.2216" ], [ "0.1274", "0.214", "0.411", "0.2134", "0.411", "0.2311", "0.1274", "0.2317" ], [ "0.5195", "0.209", "0.9126", "0.209", "0.9126", "0.2229", "0.5195", "0.2229" ], [ "0.09235", "0.2299", "0.1584", "0.2299", "0.1584", "0.2412", "0.09235", "0.2412" ], [ "0.5215", "0.2247", "0.621", "0.2247", "0.621", "0.2361", "0.5215", "0.2361" ], [ "0.08905", "0.2412", "0.2524", "0.2412", "0.2524", "0.2551", "0.08905", "0.2551" ], [ "0.6167", "0.2412", "0.739", "0.2437", "0.739", "0.262", "0.6167", "0.26" ], [ "0.08826", "0.2551", "0.10864", "0.2551", "0.10864", "0.2676", "0.08826", "0.2676" ], [ "0.134", "0.2551", "0.2754", "0.2551", "0.2754", "0.2683", "0.134", "0.2683" ], [ "0.613", "0.2583", "0.8164", "0.26", "0.8164", "0.2778", "0.613", "0.276" ], [ "0.0866", "0.2676", "0.1095", "0.2676", "0.1095", "0.281", "0.0866", "0.281" ], [ "0.134", "0.2683", "0.2361", "0.2683", "0.2361", "0.2798", "0.134", "0.2798" ], [ "0.0866", "0.2798", "0.1095", "0.2798", "0.1095", "0.2935", "0.0866", "0.2935" ], [ "0.152", "0.2778", "0.3342", "0.276", "0.3342", "0.2935", "0.152", "0.2954" ], [ "0.523", "0.286", "0.614", "0.286", "0.614", "0.2974", "0.523", "0.2974" ], [ "0.6177", "0.2842", "0.9126", "0.286", "0.9126", "0.3005", "0.6177", "0.2986" ], [ "0.0866", "0.2917", "0.1095", "0.2917", "0.1095", "0.305", "0.0866", "0.305" ], [ "0.1356", "0.2942", "0.1584", "0.2942", "0.1584", "0.303", "0.1356", "0.303" ], [ "0.156", "0.3037", "0.3718", "0.3037", "0.3718", "0.3176", "0.156", "0.3176" ], [ "0.5215", "0.3018", "0.9087", "0.3018", "0.9087", "0.3125", "0.5215", "0.3125" ], [ "0.5195", "0.3145", "0.9126", "0.3145", "0.9126", "0.3276", "0.5195", "0.3276" ], [ "0.08826", "0.305", "0.1095", "0.305", "0.1095", "0.3176", "0.08826", "0.3176" ], [ "0.0866", "0.3176", "0.1095", "0.3176", "0.1095", "0.33", "0.0866", "0.33" ], [ "0.1528", "0.315", "0.33", "0.3164", "0.33", "0.3308", "0.1528", "0.33" ], [ "0.0866", "0.33", "0.1095", "0.33", "0.1095", "0.3423", "0.0866", "0.3423" ], [ "0.134", "0.33", "0.1748", "0.33", "0.1748", "0.3416", "0.134", "0.3416" ], [ "0.5186", "0.3276", "0.9126", "0.3289", "0.9126", "0.3423", "0.5186", "0.3416" ], [ "0.08496", "0.3423", "0.1095", "0.3423", "0.1095", "0.3547", "0.08496", "0.3547" ], [ "0.1316", "0.3416", "0.3506", "0.3403", "0.3506", "0.3547", "0.1316", "0.356" ], [ "0.5195", "0.3423", "0.9126", "0.3423", "0.9126", "0.356", "0.5195", "0.356" ], [ "0.0866", "0.3547", "0.1095", "0.3547", "0.1095", "0.3674", "0.0866", "0.3674" ], [ "0.1323", "0.3547", "0.4307", "0.3547", "0.4307", "0.3687", "0.1323", "0.3687" ], [ "0.5195", "0.356", "0.9126", "0.356", "0.9126", "0.3694", "0.5195", "0.3694" ], [ "0.0866", "0.3674", "0.1095", "0.3674", "0.1095", "0.3801", "0.0866", "0.3801" ], [ "0.134", "0.3674", "0.3757", "0.3674", "0.3757", "0.3818", "0.134", "0.3818" ], [ "0.5186", "0.3694", "0.9126", "0.3687", "0.9126", "0.3833", "0.5186", "0.3838" ], [ "0.08496", "0.3801", "0.10864", "0.3801", "0.10864", "0.3926", "0.08496", "0.3926" ], [ "0.134", "0.379", "0.2507", "0.3806", "0.2507", "0.3945", "0.134", "0.3926" ], [ "0.5195", "0.3833", "0.914", "0.3833", "0.914", "0.3965", "0.5195", "0.3965" ], [ "0.08496", "0.3926", "0.1095", "0.3926", "0.1095", "0.4048", "0.08496", "0.4048" ], [ "0.134", "0.3933", "0.2345", "0.3933", "0.2345", "0.4048", "0.134", "0.4048" ], [ "0.152", "0.4028", "0.4714", "0.4016", "0.4714", "0.4192", "0.152", "0.4204" ], [ "0.5195", "0.3977", "0.714", "0.3977", "0.714", "0.4124", "0.5195", "0.4124" ], [ "0.08496", "0.4048", "0.10864", "0.4048", "0.10864", "0.4185", "0.08496", "0.4185" ], [ "0.08496", "0.4172", "0.1095", "0.4172", "0.1095", "0.4312", "0.08496", "0.4312" ], [ "0.1323", "0.4192", "0.1602", "0.4192", "0.1602", "0.4312", "0.1323", "0.4312" ], [ "0.1544", "0.4287", "0.3823", "0.43", "0.3823", "0.4443", "0.1544", "0.443" ], [ "0.6465", "0.4255", "0.785", "0.4255", "0.785", "0.44", "0.6465", "0.44" ], [ "0.08417", "0.4414", "0.1111", "0.4414", "0.1111", "0.4583", "0.08417", "0.4583" ], [ "0.0866", "0.43", "0.1095", "0.43", "0.1095", "0.4443", "0.0866", "0.4443" ], [ "0.1528", "0.4414", "0.4722", "0.44", "0.4722", "0.4583", "0.1528", "0.459" ], [ "0.5195", "0.443", "0.9126", "0.4458", "0.9126", "0.4602", "0.5195", "0.4583" ], [ "0.08417", "0.4539", "0.1111", "0.4539", "0.1111", "0.4702", "0.08417", "0.4702" ], [ "0.134", "0.4558", "0.1748", "0.4558", "0.1748", "0.4673", "0.134", "0.4673" ], [ "0.1316", "0.4653", "0.406", "0.4653", "0.406", "0.483", "0.1316", "0.483" ], [ "0.5195", "0.4602", "0.914", "0.4602", "0.914", "0.4734", "0.5195", "0.4734" ], [ "0.08496", "0.4673", "0.1095", "0.4673", "0.1095", "0.481", "0.08496", "0.481" ], [ "0.398", "0.4685", "0.4495", "0.4685", "0.4495", "0.4797", "0.398", "0.4797" ], [ "0.08496", "0.481", "0.10864", "0.481", "0.10864", "0.4944", "0.08496", "0.4944" ], [ "0.134", "0.481", "0.478", "0.481", "0.478", "0.4944", "0.134", "0.4944" ], [ "0.5215", "0.476", "0.911", "0.476", "0.911", "0.486", "0.5215", "0.486" ], [ "0.5195", "0.486", "0.9126", "0.4873", "0.9126", "0.5005", "0.5195", "0.5" ], [ "0.1095", "0.4937", "0.3545", "0.4937", "0.3545", "0.507", "0.1095", "0.507" ], [ "0.1316", "0.505", "0.3381", "0.506", "0.3381", "0.521", "0.1316", "0.5195" ], [ "0.5195", "0.502", "0.911", "0.502", "0.911", "0.515", "0.5195", "0.515" ], [ "0.3381", "0.507", "0.4429", "0.507", "0.4429", "0.5195", "0.3381", "0.5195" ], [ "0.08496", "0.506", "0.10864", "0.506", "0.10864", "0.518", "0.08496", "0.518" ], [ "0.08496", "0.5195", "0.1095", "0.5195", "0.1095", "0.532", "0.08496", "0.532" ], [ "0.3677", "0.521", "0.4763", "0.518", "0.4763", "0.532", "0.3677", "0.534" ], [ "0.5195", "0.515", "0.914", "0.515", "0.914", "0.529", "0.5195", "0.529" ], [ "0.134", "0.5215", "0.3708", "0.5215", "0.3708", "0.532", "0.134", "0.532" ], [ "0.1316", "0.531", "0.3896", "0.532", "0.3896", "0.547", "0.1316", "0.5454" ], [ "0.5195", "0.528", "0.911", "0.529", "0.911", "0.5425", "0.5195", "0.5415" ], [ "0.5195", "0.5435", "0.9126", "0.5435", "0.9126", "0.5566", "0.5195", "0.5566" ], [ "0.08496", "0.5327", "0.1095", "0.5327", "0.1095", "0.5454", "0.08496", "0.5454" ], [ "0.08496", "0.5454", "0.1748", "0.5454", "0.1748", "0.5566", "0.08496", "0.5566" ], [ "0.5195", "0.5566", "0.914", "0.5566", "0.914", "0.571", "0.5195", "0.571" ], [ "0.5195", "0.571", "0.914", "0.571", "0.914", "0.584", "0.5195", "0.584" ], [ "0.08826", "0.589", "0.4763", "0.589", "0.4763", "0.603", "0.08826", "0.603" ], [ "0.5195", "0.5854", "0.9126", "0.5854", "0.9126", "0.5986", "0.5195", "0.5986" ], [ "0.0866", "0.605", "0.4763", "0.605", "0.4763", "0.6157", "0.0866", "0.6157" ], [ "0.5215", "0.5996", "0.911", "0.5996", "0.911", "0.61", "0.5215", "0.61" ], [ "0.08496", "0.617", "0.4795", "0.617", "0.4795", "0.631", "0.08496", "0.631" ], [ "0.5215", "0.613", "0.911", "0.613", "0.911", "0.6235", "0.5215", "0.6235" ], [ "0.5195", "0.625", "0.914", "0.6255", "0.914", "0.64", "0.5195", "0.6396" ], [ "0.0866", "0.6333", "0.4756", "0.6333", "0.4756", "0.643", "0.0866", "0.643" ], [ "0.08826", "0.6465", "0.4739", "0.6465", "0.4739", "0.657", "0.08826", "0.657" ], [ "0.523", "0.6416", "0.9126", "0.6416", "0.9126", "0.6514", "0.523", "0.6514" ], [ "0.08496", "0.659", "0.4795", "0.659", "0.4795", "0.6724", "0.08496", "0.6724" ], [ "0.5215", "0.655", "0.9126", "0.655", "0.9126", "0.6655", "0.5215", "0.6655" ], [ "0.08496", "0.6724", "0.343", "0.6724", "0.343", "0.686", "0.08496", "0.686" ], [ "0.5215", "0.6694", "0.911", "0.6694", "0.911", "0.6797", "0.5215", "0.6797" ], [ "0.5195", "0.6816", "0.911", "0.6816", "0.911", "0.696", "0.5195", "0.696" ], [ "0.08496", "0.7007", "0.3782", "0.7007", "0.3782", "0.7153", "0.08496", "0.7153" ], [ "0.5215", "0.6963", "0.911", "0.6963", "0.911", "0.707", "0.5215", "0.707" ], [ "0.5195", "0.709", "0.9126", "0.709", "0.9126", "0.7227", "0.5195", "0.7227" ], [ "0.1259", "0.7163", "0.2378", "0.7163", "0.2378", "0.73", "0.1259", "0.73" ], [ "0.5195", "0.7227", "0.9126", "0.7227", "0.9126", "0.7363", "0.5195", "0.7363" ], [ "0.0866", "0.7373", "0.478", "0.7373", "0.478", "0.7505", "0.0866", "0.7505" ], [ "0.5215", "0.738", "0.9126", "0.738", "0.9126", "0.7485", "0.5215", "0.7485" ], [ "0.0866", "0.7524", "0.4756", "0.7524", "0.4756", "0.763", "0.0866", "0.763" ], [ "0.5215", "0.7505", "0.9126", "0.7505", "0.9126", "0.7646", "0.5215", "0.7646" ], [ "0.08496", "0.765", "0.4795", "0.765", "0.4795", "0.779", "0.08496", "0.779" ], [ "0.5215", "0.7646", "0.911", "0.7646", "0.911", "0.778", "0.5215", "0.778" ], [ "0.08417", "0.778", "0.4795", "0.779", "0.4795", "0.7925", "0.08417", "0.7915" ], [ "0.5215", "0.779", "0.9126", "0.779", "0.9126", "0.7925", "0.5215", "0.7925" ], [ "0.08496", "0.7925", "0.478", "0.7925", "0.478", "0.806", "0.08496", "0.806" ], [ "0.5195", "0.7915", "0.9043", "0.7915", "0.9043", "0.806", "0.5195", "0.806" ], [ "0.0866", "0.808", "0.478", "0.808", "0.478", "0.8184", "0.0866", "0.8184" ], [ "0.536", "0.806", "0.911", "0.806", "0.911", "0.8193", "0.536", "0.8193" ], [ "0.0866", "0.8213", "0.478", "0.8213", "0.478", "0.832", "0.0866", "0.832" ], [ "0.5215", "0.8213", "0.911", "0.8213", "0.911", "0.832", "0.5215", "0.832" ], [ "0.08496", "0.834", "0.4763", "0.834", "0.4763", "0.8477", "0.08496", "0.8477" ], [ "0.5195", "0.834", "0.9126", "0.834", "0.9126", "0.8477", "0.5195", "0.8477" ], [ "0.0866", "0.8496", "0.478", "0.8496", "0.478", "0.86", "0.0866", "0.86" ], [ "0.5195", "0.8477", "0.9126", "0.8477", "0.9126", "0.8613", "0.5195", "0.8613" ], [ "0.0866", "0.863", "0.4763", "0.863", "0.4763", "0.8735", "0.0866", "0.8735" ], [ "0.5195", "0.8623", "0.914", "0.8623", "0.914", "0.8755", "0.5195", "0.8755" ], [ "0.08496", "0.8755", "0.4795", "0.8755", "0.4795", "0.8896", "0.08496", "0.8896" ], [ "0.5195", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5195", "0.8896" ] ]

[ "Algorithm 1 Inverse Reinforcement Learning with Natural", "room.However, there may be no stairs in a new environ", "Language Goals (LangGoallRL)", "ment. As a result, we can not just sample goals from expert", "Input: GoalGenerator: the variational goal generator from section", "demonstrations to train in the new environments.", "3.3; D: expert demonstrations; R = @: replay buffer; b: batch", "Thehigh-levelidea of ourmethodisto first learn apolicy", "size; N: number of expert relabeling goals.", "Tp that is goal-agnostic. That is, the policy selects an action", "1: D=0,g=0", "purely based on the state without conditioning on a goal.", "2:for each(Ei,Gi,T)EDdo", "This model gives us a prior of how the agent will act given a", "for n = 1, 2, .., N do", "state. We use behavior cloning to train this policy on expert", "ExpertGoalRelabeling (EGR)", "3.", "4:", "demonstrations. More details are available in Appendix B.", "5:", "Gi,n ~ GoalGenerator(T.)", "We then sample trajectories in the new environments using", "6:", "end for", "1", "add (Ex,Gi, Ti) and [(Et, Gt,n,Ti)N=I to D", "this policy and sample goals of these trajectories using the", "8:end for", "goal generator:", "9: while not converged do", "Inew ~ Tp(Enew)", "10:", "r1, r2 ~ Uniform(0, 1)", "Gnew ~ GoalGenerator(+new)", "11:", "if r1 < 0.5 then", "12:", "Sample a goal (E;,G,) ~ D", "(Enew.Gnew", "Tnew) are treated as expert demonstrations", "13:", "else", " Hindsight Goal Sampling (HGS)", "in the new environments.Wethen fine-tunethe discrimina", "tor and the policy in new environments using Algorithm 1", "14:", "15:", "Sample a goal (E,, G,) ~ g", "16:", "endif", "(w/o expert goal relabeling). The self-supervised learning", "17:", "Sample a trajectory T, ~π(Ej, G,)", "signals come from both the generated expert demonstrations", "18:", "Sample a relabeled goal G', ~ GoalGenerator(T))", "for the new environments, and the hindsight goal relabeling", "19:", "Add (Ej, Gj, G'j, Ti) to replay buffer R", "and sampling proposed in Section 3.3. The generated expert", "20:", "Add (E,, G)) to g", "demonstrations can be seen as seeds to bootstrap hindsight", "21:", "if r2<0.5then", "Sample a batch P+ = ((s,ak, st+, Gk))%=1~ D", "goal relabeling and sampling.", "22", "23:", "else", " Hindsight Goal Relabeling (HGR)", "4 Experiments", "25:", "24:", "Sample a batch P+ = ((sk,ak, $++1,G))=1 ~ R", "While LangGoallRL works on any natural language (NL)", "26:", "endif", "Sample a batch P_ =[(sk,ak, st+1,Gk)%", "goal based IRL problems, our experiments focus on a", "27:", "-1~R", "28:", "Update discriminator parameters with P+ and P_ as pos-", "challengingvision-based instruction followingproblem.", "We evaluate our modelontheRoom-2-Room(R2R)", "itive and negative examples, respectively.", "SampleabatchQ={(s,a,s", "dataset (Anderson et al. 2018), a visually-grounded NL nav-", "1,Gk)h=1 ~ R", "29:", "30:", "=ge(sk,ak,Gk)", "igation task in realistic 3D indoor environments. The dataset", "ExpandeachentryofQwithrewardr", "Optimize using Soft Actor-Critic with Q", "contains 7,189 routes sampled from 90 real world indoor", "environments. A route is a sequence of viewpoints in the", "31:", "32: end while", "indoor environments with the agent's first-person camera", "views. Each route is annotated by humans with 3 navigation", "1). The difference between HER and HGR is that the for-", "instructions. The average length of navigation instructions", "mer is goal relabelingforthepolicy whilethe latter isgoal", "is 29words.The dataset is split intotrain(61 environments", "relabeling for the discriminator. HER is most efficient when", "and 14.025 instructions),seen validation(61 environments", "same as train set, and 1,020 instructions), unseen validation", "rewardsaresparse:however,inour setting,therewardspro-", "vided by the discriminator are not sparse and we do not ob", "(11newenvironmentsand2.349instructions),andtest(18", "serve a boost of performance after applying HER. This is", "new environments and 4.173 instructions).Wedon'tuse the", "discussed in greater detail in Section 4.", "test setfor evaluation becausetheground-truthroutes of the", "test set are not released. Along with the dataset, a simula-", "3.4 Self-Supervised Learning in New", "tor is provided to allow the embodied agent to interact with", "the environments. The observation of the agent is the first-", "Environments", "person camera view, which is a panoramic image. The ac-", "In this section, we specifically examine the scenario where", "tion of the agent is the nearby viewpoints that the agent can", "thelearned policyis deployed toa new environment.For ex-", "move to. For details about the dataset, please see Appendix", "ample, after training an embodied agent to perform tasks in a", "A. The R2R dataset has become a very popular testbed for", "set of buildings, we may deploy this agent to a new building", "language grounding in visual context (Tan, Yu, and Bansal", "with different floor plans. We assume that we have access", "2019; Wang et al. 2019; Fried et al. 2018; Zhu et al. 2020).", "to these new environments,but we do not have any expert", "We evaluate the model performance based on the trajec-", "demonstrations noranyNLgoals innew environments.Note", "tory successrate.Eachnavigation instruction inthe R2R", "that we can not directly apply NL goals from existing envi-", "dataset is labeled with a goal position. Following Anderson", "ronmentstonewenvironments,becausegoalsaretiedtothe", "et al. (2018), the agent is considered successfully reaching", "environments.Forexample,in instruction-following tasks.", "the goal if the navigation error (the distance between the", "an example goal is go downstairs and walk pass the living", "stop position and the goal position) is less than 3 meters." ]

{ "arxiv_id": "2008.06924", "language": "en", "page_idx": 3 }

[ [ "0.1897", "0.09143", "0.806", "0.09265", "0.806", "0.10986", "0.1897", "0.10925" ], [ "0.3123", "0.10986", "0.686", "0.111", "0.686", "0.13", "0.3123", "0.1289" ], [ "0.1948", "0.1538", "0.3064", "0.1561", "0.3064", "0.1733", "0.1948", "0.1716" ], [ "0.451", "0.158", "0.5483", "0.158", "0.5483", "0.1692", "0.451", "0.1692" ], [ "0.69", "0.1567", "0.817", "0.1567", "0.817", "0.1716", "0.69", "0.1716" ], [ "0.1663", "0.1733", "0.3342", "0.1733", "0.3342", "0.187", "0.1663", "0.187" ], [ "0.3855", "0.1729", "0.6094", "0.1733", "0.6094", "0.1882", "0.3855", "0.187" ], [ "0.6826", "0.1733", "0.818", "0.1733", "0.818", "0.187", "0.6826", "0.187" ], [ "0.1335", "0.1907", "0.3687", "0.1907", "0.3687", "0.2037", "0.1335", "0.2037" ], [ "0.4072", "0.1907", "0.592", "0.1895", "0.592", "0.2037", "0.4072", "0.2048" ], [ "0.615", "0.1907", "0.882", "0.1895", "0.882", "0.2037", "0.615", "0.2048" ], [ "0.2627", "0.253", "0.3418", "0.253", "0.3418", "0.2673", "0.2627", "0.2673" ], [ "0.5415", "0.253", "0.665", "0.2546", "0.665", "0.2654", "0.5415", "0.2632" ], [ "0.5425", "0.2683", "0.586", "0.2683", "0.586", "0.279", "0.5425", "0.279" ], [ "0.613", "0.2756", "0.748", "0.2756", "0.748", "0.2861", "0.613", "0.2861" ], [ "0.1453", "0.2832", "0.4592", "0.2832", "0.4592", "0.2964", "0.1453", "0.2964" ], [ "0.525", "0.2893", "0.5366", "0.2935", "0.531", "0.3018", "0.519", "0.2976" ], [ "0.546", "0.2852", "0.601", "0.2852", "0.601", "0.2922", "0.546", "0.2922" ], [ "0.621", "0.2903", "0.864", "0.2903", "0.864", "0.3005", "0.621", "0.3005" ], [ "0.1427", "0.2964", "0.461", "0.2957", "0.461", "0.31", "0.1427", "0.3113" ], [ "0.02016", "0.3027", "0.06046", "0.3027", "0.06046", "0.4353", "0.02016", "0.4353" ], [ "0.544", "0.3018", "0.8228", "0.3018", "0.8228", "0.3088", "0.544", "0.3088" ], [ "0.1453", "0.3113", "0.461", "0.3113", "0.461", "0.3247", "0.1453", "0.3247" ], [ "0.5474", "0.3164", "0.5576", "0.3164", "0.5576", "0.3206", "0.5474", "0.3206" ], [ "0.5635", "0.3237", "0.571", "0.3237", "0.571", "0.328", "0.5635", "0.328" ], [ "0.596", "0.3196", "0.7734", "0.3196", "0.7734", "0.33", "0.596", "0.33" ], [ "0.5483", "0.332", "0.6035", "0.332", "0.6035", "0.3362", "0.5483", "0.3362" ], [ "0.1453", "0.3247", "0.4592", "0.3247", "0.4592", "0.339", "0.1453", "0.339" ], [ "0.5225", "0.3333", "0.5366", "0.3333", "0.5366", "0.3416", "0.5225", "0.3416" ], [ "0.1453", "0.339", "0.4575", "0.339", "0.4575", "0.3528", "0.1453", "0.3528" ], [ "0.555", "0.3362", "0.8486", "0.3362", "0.8486", "0.3467", "0.555", "0.3467" ], [ "0.5474", "0.3467", "0.825", "0.3467", "0.825", "0.354", "0.5474", "0.354" ], [ "0.147", "0.3552", "0.4575", "0.3552", "0.4575", "0.3652", "0.147", "0.3652" ], [ "0.544", "0.3528", "0.8135", "0.3528", "0.8135", "0.3635", "0.544", "0.3635" ], [ "0.1453", "0.3677", "0.461", "0.3677", "0.461", "0.3813", "0.1453", "0.3813" ], [ "0.1453", "0.3818", "0.4592", "0.3818", "0.4592", "0.3955", "0.1453", "0.3955" ], [ "0.51", "0.3801", "0.8833", "0.3813", "0.8833", "0.3955", "0.51", "0.3943" ], [ "0.1453", "0.3967", "0.4575", "0.3967", "0.4575", "0.4104", "0.1453", "0.4104" ], [ "0.513", "0.398", "0.881", "0.398", "0.881", "0.408", "0.513", "0.408" ], [ "0.1453", "0.4104", "0.4575", "0.4104", "0.4575", "0.4233", "0.1453", "0.4233" ], [ "0.511", "0.4104", "0.785", "0.4104", "0.785", "0.4233", "0.511", "0.4233" ], [ "0.1436", "0.4246", "0.4575", "0.4246", "0.4575", "0.4382", "0.1436", "0.4382" ], [ "0.02016", "0.437", "0.0638", "0.4358", "0.068", "0.5166", "0.02519", "0.517" ], [ "0.1436", "0.4395", "0.4575", "0.4395", "0.4575", "0.4531", "0.1436", "0.4531" ], [ "0.147", "0.4548", "0.4575", "0.4548", "0.4575", "0.4656", "0.147", "0.4656" ], [ "0.511", "0.4495", "0.882", "0.4495", "0.882", "0.4644", "0.511", "0.4644" ], [ "0.1453", "0.4673", "0.4592", "0.4673", "0.4592", "0.481", "0.1453", "0.481" ], [ "0.511", "0.466", "0.882", "0.466", "0.882", "0.4797", "0.511", "0.4797" ], [ "0.1436", "0.4822", "0.4568", "0.481", "0.4568", "0.4946", "0.1436", "0.4954" ], [ "0.511", "0.4827", "0.831", "0.4827", "0.831", "0.4963", "0.511", "0.4963" ], [ "0.1453", "0.4954", "0.4592", "0.4954", "0.4592", "0.509", "0.1453", "0.509" ], [ "0.529", "0.4976", "0.882", "0.4988", "0.882", "0.513", "0.529", "0.5127" ], [ "0.1436", "0.5103", "0.4592", "0.5083", "0.4592", "0.5215", "0.1436", "0.524" ], [ "0.511", "0.5156", "0.882", "0.5156", "0.882", "0.5293", "0.511", "0.5293" ], [ "0.02351", "0.5215", "0.06046", "0.5215", "0.06046", "0.589", "0.02351", "0.589" ], [ "0.1453", "0.524", "0.4592", "0.524", "0.4592", "0.5376", "0.1453", "0.5376" ], [ "0.51", "0.531", "0.882", "0.53", "0.882", "0.5444", "0.51", "0.546" ], [ "0.1453", "0.538", "0.4592", "0.538", "0.4592", "0.552", "0.1453", "0.552" ], [ "0.511", "0.5474", "0.882", "0.5474", "0.882", "0.561", "0.511", "0.561" ], [ "0.1453", "0.554", "0.2317", "0.554", "0.2317", "0.5654", "0.1453", "0.5654" ], [ "0.511", "0.5635", "0.882", "0.5635", "0.882", "0.5767", "0.511", "0.5767" ], [ "0.02016", "0.5767", "0.06213", "0.5767", "0.06213", "0.7285", "0.02016", "0.7285" ], [ "0.1453", "0.583", "0.2603", "0.583", "0.2603", "0.5957", "0.1453", "0.5957" ], [ "0.511", "0.579", "0.882", "0.579", "0.882", "0.5933", "0.511", "0.5933" ], [ "0.1351", "0.586", "0.1427", "0.586", "0.1427", "0.5903", "0.1351", "0.5903" ], [ "0.51", "0.594", "0.882", "0.5957", "0.882", "0.61", "0.51", "0.6094" ], [ "0.1184", "0.5845", "0.1318", "0.5845", "0.1318", "0.593", "0.1184", "0.593" ], [ "0.11755", "0.608", "0.4902", "0.608", "0.4902", "0.6216", "0.11755", "0.6216" ], [ "0.511", "0.6123", "0.882", "0.6123", "0.882", "0.627", "0.511", "0.627" ], [ "0.11755", "0.6235", "0.4878", "0.6235", "0.4878", "0.638", "0.11755", "0.638" ], [ "0.51", "0.627", "0.882", "0.6274", "0.882", "0.6426", "0.51", "0.641" ], [ "0.11584", "0.638", "0.4878", "0.638", "0.4878", "0.652", "0.11584", "0.652" ], [ "0.511", "0.6436", "0.881", "0.6445", "0.881", "0.6577", "0.511", "0.6567" ], [ "0.11584", "0.6567", "0.4878", "0.6567", "0.4878", "0.6704", "0.11584", "0.6704" ], [ "0.51", "0.66", "0.882", "0.66", "0.882", "0.6733", "0.51", "0.6733" ], [ "0.11584", "0.673", "0.4878", "0.673", "0.4878", "0.686", "0.11584", "0.686" ], [ "0.509", "0.676", "0.882", "0.675", "0.882", "0.6895", "0.509", "0.6904" ], [ "0.1142", "0.6885", "0.486", "0.6885", "0.486", "0.7017", "0.1142", "0.7017" ], [ "0.51", "0.6924", "0.882", "0.6924", "0.882", "0.706", "0.51", "0.706" ], [ "0.1142", "0.7036", "0.4878", "0.703", "0.4878", "0.7173", "0.1142", "0.7188" ], [ "0.51", "0.707", "0.6885", "0.707", "0.6885", "0.72", "0.51", "0.72" ], [ "0.11584", "0.7188", "0.4878", "0.7197", "0.4878", "0.734", "0.11584", "0.733" ], [ "0.529", "0.7246", "0.882", "0.7246", "0.882", "0.7393", "0.529", "0.7393" ], [ "0.11584", "0.7354", "0.4878", "0.7363", "0.4878", "0.7505", "0.11584", "0.75" ], [ "0.511", "0.741", "0.882", "0.741", "0.882", "0.755", "0.511", "0.755" ], [ "0.11584", "0.7505", "0.486", "0.752", "0.486", "0.7666", "0.11584", "0.7656" ], [ "0.511", "0.7573", "0.882", "0.7573", "0.882", "0.771", "0.511", "0.771" ], [ "0.11584", "0.767", "0.4878", "0.767", "0.4878", "0.781", "0.11584", "0.781" ], [ "0.51", "0.7715", "0.882", "0.7725", "0.882", "0.7876", "0.51", "0.786" ], [ "0.11755", "0.785", "0.4878", "0.785", "0.4878", "0.799", "0.11755", "0.799" ], [ "0.51", "0.789", "0.8833", "0.789", "0.8833", "0.8027", "0.51", "0.8027" ], [ "0.1184", "0.802", "0.476", "0.802", "0.476", "0.8125", "0.1184", "0.8125" ], [ "0.51", "0.8047", "0.882", "0.8047", "0.882", "0.8184", "0.51", "0.8184" ], [ "0.1368", "0.818", "0.4895", "0.818", "0.4895", "0.8306", "0.1368", "0.8306" ], [ "0.511", "0.822", "0.8833", "0.822", "0.8833", "0.835", "0.511", "0.835" ], [ "0.11755", "0.8345", "0.4895", "0.8345", "0.4895", "0.847", "0.11755", "0.847" ], [ "0.51", "0.8374", "0.885", "0.8374", "0.885", "0.8516", "0.51", "0.8516" ], [ "0.11584", "0.8496", "0.4902", "0.8496", "0.4902", "0.8633", "0.11584", "0.8633" ], [ "0.511", "0.8525", "0.882", "0.8525", "0.882", "0.8677", "0.511", "0.8677" ], [ "0.11755", "0.865", "0.4878", "0.865", "0.4878", "0.879", "0.11755", "0.879" ], [ "0.51", "0.869", "0.882", "0.868", "0.882", "0.883", "0.51", "0.8843" ], [ "0.11584", "0.882", "0.486", "0.8813", "0.486", "0.8955", "0.11584", "0.8965" ], [ "0.532", "0.891", "0.8833", "0.8936", "0.8833", "0.9077", "0.532", "0.9062" ], [ "0.147", "0.905", "0.288", "0.905", "0.288", "0.9185", "0.147", "0.9185" ], [ "0.511", "0.905", "0.8154", "0.905", "0.8154", "0.9185", "0.511", "0.9185" ] ]

[ "PreWoMe: Exploiting Presuppositions as Working Memory", "for Long Form Question Answering", "Wookje Han", "JinsolPark", "Kyungjae Lee*", "Columbia University", "Carnegie Mellon University", "LG AIResearch", "wookje.han@columbia.edu", "jinsolp@cs.cmu.edu", "kyungjae.lee@lgresearch.ai", "Abstract", "Misleading Questions", "Question", "warbetweenEthiopia and Italy?", "Information-seeking questions in long-form", "(a)", "DesirableAn", "wars between Ethiopia and Italy.The first war (1895-1896)", "question answering (LFQA) often prove mis-", " ", "waswonbyEthiopia,andthesecond(1935-1937)waswonbylItaly", "leading due to ambiguity or false presuppo-", "ou", "th", "the Nobel Peace Prizefor DNA structure?", "DesirableAnt", "sition in the question. While many existing", "(b)", "approaches handle misleading questions, they", "NobodywontheNobelpeaceprizeforDNAstructure.Instead,James", "Watson,FrancisCrick,andMauriceWilkinswontheNobelPrizein", "aretailored to limited questions,which are in-", "Physiology or Medicine for their discovery of the DNA structure.", "sufficient in a real-world setting with unpre-", "dictable input characteristics. In this work, we", "Figure l: Two types of misleading questions: (a) hav-", "propose Pre WoMe, a unified approach capable", "ing ambiguity,and (b)falsepresupposition.Highlights", "of handling any type of information-seeking", "show essential cues to avoid misleading.", "question. The key idea of Pre WoMe involves", "[T's]", "extracting presuppositions in the question and", "exploitingthemasworkingmemorytogener-", "won the Nobel peace prize for DNA structure\"),", "ate feedback and action about the question. Our", "which should be corrected before answering the", "experiment shows that PreWoMe is effective", "question (Yu et al., 2022; Kim et al., 2022)", "not only in tackling misleading questions but", "Desirable answers to misleading questions re-", "also in handling normal ones, thereby demon-", "quire discerning and resolving misleadings. A pre-", "", "strating the effectiveness of leveraging presup-", "vious work by Amplayo et al. (2022) focuses on", "positions, feedback, and action for real-world", "ambiguous questions, by introducing query refine-", "QA settings.", "ment prompts that encourage LMs to consider mul-", "1901A", "Introduction", "tiple facets of the question. Another work (Kim", "", "et al., 2021) tackles questions containing FP, by ex-", "1", "Answering information-seeking long-form ques-", "tracting and verifying presuppositions. While such", "tions has recently shown significant progress (Fan", "works assume only one specific type of misleading", "et al., 2019: Krishna et al., 2021; Xu et al., 2023)", "factor, in a real scenario, the users ask various types", "However, users' questions in the real-world may", "of questions, which can be misleading or normal", "often mislead language models (LMs) to output", "(i.e, non-misleading).I This makes prior single-", "misinformation. For example, users ask questions", "type-tailored approaches hard to be deployed in", "with false presuppositions (FP) (Kim et al., 2022;", "real-world QA settings.", "Yu et al., 2022) which can induce hallucinations if", "In this paper, we emphasize the role of presuppo-", "LMs believe the presuppositions are true (Maynez", "sitions in handling any type of information-seeking", "et al., 2020). Also, users often ask questions that", "questions in a unified way. Presuppositions are cru-", "do not have a single and clear answer, i.e., an am-", "cial building blocks of questions and play an essen", "biguous question, which is difficult for LMs to", " tial role in understanding the meaning of the ques-", "identify (Min et al..2020: Stelmakh et al., 2022)", "tion (Ge, 2011; Duzi and Cihalova, 2015). This", "Figure 1 shows examples of two types of mis-", "suggests that exploiting presuppositions is promis-", "leading questions: (a) an ambiguous question that", "ing in handling misleading questions, while gen-", "requires considering prior knowledge (\"two differ-", "erally applicable to any type of question. Based", "ent wars between Italy and Ethiopia\")to answer", "on this, we propose a new approach, PreWoMe", "properly, and (b) a question with FP (\"somebody", "'Throughout this paper, we use the term misleading ques-", "*Corresponding author", "tions for questions that are ambiguous or have FP." ]

{ "arxiv_id": "2310.16147", "language": "en", "page_idx": 0 }

[ [ "0.1184", "0.0861", "0.4878", "0.0861", "0.4878", "0.0998", "0.1184", "0.0998" ], [ "0.663", "0.08905", "0.686", "0.08905", "0.686", "0.0974", "0.663", "0.0974" ], [ "0.763", "0.09143", "0.785", "0.09143", "0.785", "0.0956", "0.763", "0.0956" ], [ "0.81", "0.09143", "0.827", "0.09143", "0.827", "0.0956", "0.81", "0.0956" ], [ "0.8555", "0.09143", "0.867", "0.09143", "0.867", "0.0956", "0.8555", "0.0956" ], [ "0.11755", "0.1027", "0.486", "0.1027", "0.486", "0.1164", "0.11755", "0.1164" ], [ "0.657", "0.111", "0.686", "0.111", "0.686", "0.1194", "0.657", "0.1194" ], [ "0.11584", "0.11816", "0.4382", "0.11816", "0.4382", "0.1318", "0.11584", "0.1318" ], [ "0.6885", "0.1217", "0.8447", "0.12054", "0.8447", "0.1318", "0.6885", "0.133" ], [ "0.1351", "0.133", "0.4878", "0.1354", "0.4878", "0.1484", "0.1351", "0.1467" ], [ "0.6245", "0.1372", "0.721", "0.1372", "0.721", "0.152", "0.6245", "0.152" ], [ "0.8296", "0.1426", "0.848", "0.1426", "0.848", "0.1509", "0.8296", "0.1509" ], [ "0.11755", "0.1509", "0.4878", "0.1509", "0.4878", "0.1644", "0.11755", "0.1644" ], [ "0.11584", "0.1663", "0.486", "0.165", "0.486", "0.1799", "0.11584", "0.1812" ], [ "0.692", "0.1758", "0.8657", "0.1758", "0.8657", "0.1859", "0.692", "0.1859" ], [ "0.11755", "0.1829", "0.4878", "0.1829", "0.4878", "0.1965", "0.11755", "0.1965" ], [ "0.11755", "0.1965", "0.1823", "0.199", "0.1814", "0.2131", "0.11584", "0.2114" ], [ "0.644", "0.1912", "0.702", "0.1912", "0.702", "0.2073", "0.644", "0.2073" ], [ "0.1184", "0.2185", "0.486", "0.2185", "0.486", "0.2322", "0.1184", "0.2322" ], [ "0.659", "0.2214", "0.686", "0.2214", "0.686", "0.2299", "0.659", "0.2299" ], [ "0.1351", "0.234", "0.486", "0.2328", "0.486", "0.2477", "0.1351", "0.2488" ], [ "0.693", "0.2328", "0.8154", "0.2322", "0.8154", "0.2434", "0.693", "0.2446" ], [ "0.6675", "0.2434", "0.679", "0.2434", "0.679", "0.2546", "0.6675", "0.2546" ], [ "0.1335", "0.2505", "0.486", "0.2505", "0.486", "0.2642", "0.1335", "0.2642" ], [ "0.1377", "0.2683", "0.486", "0.2683", "0.486", "0.279", "0.1377", "0.279" ], [ "0.511", "0.2666", "0.882", "0.2666", "0.882", "0.2798", "0.511", "0.2798" ], [ "0.1351", "0.2832", "0.4626", "0.2832", "0.4626", "0.2964", "0.1351", "0.2964" ], [ "0.51", "0.2808", "0.882", "0.2798", "0.882", "0.2944", "0.51", "0.2957" ], [ "0.12177", "0.3005", "0.486", "0.3005", "0.486", "0.3154", "0.12177", "0.3154" ], [ "0.511", "0.2957", "0.81", "0.2957", "0.81", "0.3088", "0.511", "0.3088" ], [ "0.1351", "0.3176", "0.4878", "0.3176", "0.4878", "0.3308", "0.1351", "0.3308" ], [ "0.1351", "0.3342", "0.4895", "0.3342", "0.4895", "0.3474", "0.1351", "0.3474" ], [ "0.511", "0.3342", "0.881", "0.3342", "0.881", "0.3474", "0.511", "0.3474" ], [ "0.1351", "0.3499", "0.4895", "0.3499", "0.4895", "0.3635", "0.1351", "0.3635" ], [ "0.513", "0.3516", "0.882", "0.3516", "0.882", "0.3652", "0.513", "0.3652" ], [ "0.1351", "0.3652", "0.4517", "0.3652", "0.4517", "0.379", "0.1351", "0.379" ], [ "0.511", "0.3677", "0.882", "0.3665", "0.882", "0.3813", "0.511", "0.3818" ], [ "0.12177", "0.383", "0.4878", "0.383", "0.4878", "0.398", "0.12177", "0.398" ], [ "0.511", "0.3843", "0.882", "0.3843", "0.882", "0.398", "0.511", "0.398" ], [ "0.1351", "0.3997", "0.4878", "0.3997", "0.4878", "0.4133", "0.1351", "0.4133" ], [ "0.51", "0.3997", "0.882", "0.3997", "0.882", "0.4133", "0.51", "0.4133" ], [ "0.1368", "0.4163", "0.2107", "0.4163", "0.2107", "0.4275", "0.1368", "0.4275" ], [ "0.513", "0.415", "0.8833", "0.415", "0.8833", "0.4287", "0.513", "0.4287" ], [ "0.511", "0.4312", "0.882", "0.4316", "0.882", "0.4465", "0.511", "0.4453" ], [ "0.1142", "0.4407", "0.2551", "0.4424", "0.2551", "0.4602", "0.1142", "0.4578" ], [ "0.513", "0.449", "0.882", "0.449", "0.882", "0.462", "0.513", "0.462" ], [ "0.11755", "0.4697", "0.4878", "0.4697", "0.4878", "0.484", "0.11755", "0.484" ], [ "0.513", "0.4656", "0.8833", "0.4656", "0.8833", "0.4785", "0.513", "0.4785" ], [ "0.11584", "0.484", "0.4895", "0.485", "0.4895", "0.4995", "0.11584", "0.4988" ], [ "0.51", "0.481", "0.882", "0.481", "0.882", "0.4946", "0.51", "0.4946" ], [ "0.11584", "0.502", "0.486", "0.502", "0.486", "0.5156", "0.11584", "0.5156" ], [ "0.511", "0.4963", "0.8833", "0.4963", "0.8833", "0.511", "0.511", "0.511" ], [ "0.11755", "0.5186", "0.486", "0.5186", "0.486", "0.532", "0.11755", "0.532" ], [ "0.511", "0.513", "0.8833", "0.513", "0.8833", "0.527", "0.511", "0.527" ], [ "0.11755", "0.5337", "0.4878", "0.5337", "0.4878", "0.5474", "0.11755", "0.5474" ], [ "0.51", "0.5293", "0.881", "0.5293", "0.881", "0.542", "0.51", "0.542" ], [ "0.51", "0.5435", "0.882", "0.5444", "0.882", "0.5586", "0.51", "0.558" ], [ "0.11755", "0.55", "0.4878", "0.55", "0.4878", "0.5635", "0.11755", "0.5635" ], [ "0.11755", "0.5664", "0.4902", "0.5664", "0.4902", "0.58", "0.11755", "0.58" ], [ "0.511", "0.561", "0.636", "0.561", "0.636", "0.5747", "0.511", "0.5747" ], [ "0.11584", "0.5806", "0.4878", "0.582", "0.4878", "0.5967", "0.11584", "0.5957" ], [ "0.513", "0.5884", "0.6323", "0.5884", "0.6323", "0.6025", "0.513", "0.6025" ], [ "0.11755", "0.5986", "0.403", "0.5986", "0.403", "0.6123", "0.11755", "0.6123" ], [ "0.1184", "0.615", "0.4014", "0.615", "0.4014", "0.629", "0.1184", "0.629" ], [ "0.4055", "0.615", "0.4878", "0.615", "0.4878", "0.627", "0.4055", "0.627" ], [ "0.509", "0.6133", "0.882", "0.615", "0.882", "0.629", "0.509", "0.627" ], [ "0.11584", "0.63", "0.4878", "0.63", "0.4878", "0.6436", "0.11584", "0.6436" ], [ "0.509", "0.6313", "0.881", "0.63", "0.881", "0.6445", "0.509", "0.6455" ], [ "0.11755", "0.648", "0.4902", "0.648", "0.4902", "0.661", "0.11755", "0.661" ], [ "0.51", "0.648", "0.882", "0.6455", "0.882", "0.659", "0.51", "0.661" ], [ "0.11755", "0.663", "0.4937", "0.663", "0.4937", "0.677", "0.11755", "0.677" ], [ "0.51", "0.663", "0.839", "0.663", "0.839", "0.6777", "0.51", "0.6777" ], [ "0.11755", "0.6787", "0.4895", "0.6787", "0.4895", "0.6934", "0.11755", "0.6934" ], [ "0.11755", "0.695", "0.4895", "0.695", "0.4895", "0.709", "0.11755", "0.709" ], [ "0.511", "0.6904", "0.76", "0.6904", "0.76", "0.7036", "0.511", "0.7036" ], [ "0.11584", "0.7114", "0.4937", "0.7104", "0.4937", "0.7246", "0.11584", "0.7256" ], [ "0.509", "0.7104", "0.882", "0.7114", "0.882", "0.7256", "0.509", "0.7246" ], [ "0.11584", "0.728", "0.4895", "0.728", "0.4895", "0.741", "0.11584", "0.741" ], [ "0.511", "0.728", "0.882", "0.728", "0.882", "0.741", "0.511", "0.741" ], [ "0.11755", "0.7446", "0.4895", "0.7446", "0.4895", "0.7583", "0.11755", "0.7583" ], [ "0.51", "0.742", "0.881", "0.7437", "0.881", "0.7583", "0.51", "0.7573" ], [ "0.11755", "0.7603", "0.4878", "0.7603", "0.4878", "0.774", "0.11755", "0.774" ], [ "0.511", "0.7583", "0.881", "0.7583", "0.881", "0.7725", "0.511", "0.7725" ], [ "0.11584", "0.7754", "0.4895", "0.7754", "0.4895", "0.789", "0.11584", "0.789" ], [ "0.51", "0.775", "0.881", "0.775", "0.881", "0.788", "0.51", "0.788" ], [ "0.11755", "0.792", "0.4307", "0.792", "0.4307", "0.8057", "0.11755", "0.8057" ], [ "0.513", "0.7915", "0.882", "0.7915", "0.882", "0.8057", "0.513", "0.8057" ], [ "0.2795", "0.807", "0.4895", "0.8057", "0.4895", "0.821", "0.2795", "0.822" ], [ "0.511", "0.8057", "0.882", "0.808", "0.882", "0.8223", "0.511", "0.821" ], [ "0.509", "0.8223", "0.882", "0.822", "0.882", "0.839", "0.509", "0.8403" ], [ "0.11755", "0.808", "0.2896", "0.808", "0.2896", "0.822", "0.11755", "0.822" ], [ "0.1184", "0.826", "0.4878", "0.826", "0.4878", "0.836", "0.1184", "0.836" ], [ "0.11755", "0.8403", "0.4878", "0.8403", "0.4878", "0.854", "0.11755", "0.854" ], [ "0.529", "0.839", "0.882", "0.8403", "0.882", "0.855", "0.529", "0.854" ], [ "0.11584", "0.855", "0.4895", "0.8555", "0.4895", "0.8706", "0.11584", "0.869" ], [ "0.51", "0.857", "0.882", "0.8555", "0.882", "0.8706", "0.51", "0.8716" ], [ "0.11584", "0.8726", "0.486", "0.8726", "0.486", "0.8857", "0.11584", "0.8857" ], [ "0.511", "0.8726", "0.8833", "0.8716", "0.8833", "0.8857", "0.511", "0.887" ], [ "0.11755", "0.888", "0.4878", "0.888", "0.4878", "0.903", "0.11755", "0.903" ], [ "0.511", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.511", "0.903" ], [ "0.11584", "0.905", "0.2568", "0.905", "0.2568", "0.9185", "0.11584", "0.9185" ], [ "0.511", "0.905", "0.882", "0.905", "0.882", "0.9185", "0.511", "0.9185" ] ]

[ "(Presupposition as Working Memory), that han-", "9", "\"Ther", "rlea", "nn", "dles any type of information-seeking question in a", "LLM", "unified way without any parameter updates.", "(a)PresuppositionExtraction", "The main idea of PreWoMe is to extract pre-", "(P1.P2...P.):", "PA", "suppositions and use them as working memory to", "generate analysis and directions for answering the", "(b)Feedback,Action Generation", "question. Our contributions can be summarized as", "follows:", "(F,A)", ". We analyze the performance of recent large lan", "LLM", "guage models (LLMs), GPT-4 and GPT-3.5, on", "(c)AnswerGeneration", "a", "misleading questions. To the best of our knowl-", "edge,wearethefirsttoexploretheperformance", "Figure 2: Overall pipeline of Pre WoMe. q is the given", "of LLMs this large on misleading questions.", "question, (P1, P2, .., ) are the model-generated pre-", "• We propose Pre WoMe, a new approach designed", "suppositions. (F, A) is feedback and action.", "for real-world QA settings that 1) is capable of", "handling any type of information-seeking ques-", "Self-correcting ApproachesWhile LLMs occa-", "tion, and 2) does not entail any parameter up-", "sionally struggle to generate accurate answers with", "dates, making it easily adaptable to LLMs.", "out hallucination, recent works suggest that LLMs", ". We propose using presuppositions as a working", "can enhance accuracy by self-correcting their re-", "memory when answering information-seeking", "sponses through iterative prompting (Madaan et al.,", "questions.", "2023; Shinn et al., 2023; Saunders et al., 2022;", "Press et al., 2022; Weng et al., 2023). Such ap", "2 Background", "proaches are based on the idea that verifying cor-", "PresuppositionPresuppositions are conditions", "rectness is easier than directly generating an accu-", "that are believed to be true by speakers in a dis-", "rate answer. Inspired by this idea, we hypothesize", "course (Yu et al., 2022). Thus, in order for an utter-", "that 1) LLMs are capable of generating and veri-", "ance to be appropriate, the presuppositions should", "fying presuppositions implicit in a given question,", "be true. For example, if someone said I care for my", "and 2) guiding LLMs to generate answers after the", "verification will yield better results than generating", "sister, then it can be assumed that the speaker has a", "sister. If not, the utterance would be inappropriate.", "answers directly.", "In this paper, we suggest using presuppositions to", "3PreWoMe", "solve an information-seeking QA task.", "Chain-of-thought Prompting (CoT)", "Weiet al.", "We introduce PreWoMe, a new approach that an-", "(2022) explores how thinking with intermediate", "swers any type of long-form information-seeking", "steps greatly improves LM's performance. Chain-", "questioninaunifiedway.PreWoMefollowsthree", "of-thought Prompting effectively enhances the rea-", " intermediate steps, as illustrated in Figure 2.", "soning abilities of LMs, particularly in mathemat-", "ical and logical reasoning tasks. This technique", "3.1 PresuppositionExtraction", "involves guiding the model through a series of in-", "To exploit presuppositions as working memory, Pre-", "terconnected thought processes. Inspired by this", "WoMe first generates presuppositions contained in", "concept, we propose a novel approach that lever-", "the question q. Motivated by the few-shot learning", "ages presupposition, feedback, and action, serving", "capability of LLMs (Brown et al., 2020), we feed g", "as intermediate steps, which boosts the robustness", "with N (question,presupposition)pairs as few-shot", "of LMs in handling open-ended questions.", "examples within a designed prompt (See Appendix", "Working memory (Miller", "A). The LLM generates multiple presuppositions", "(pi, P2, , Par) in q using this input (Figure 2(a).", "Working Memory", "et al..2017)isthe immediateinformation buffer", "that is accessed while performing conscious tasks", "In addition, we append a manual presupposition", "(Li et al., 2022). We make presuppositions func-", "pa: \"There is a clear and single answer to the", "tion as working memory by extracting and feeding", "question\", as shown in Figure 2, which is a com-", "them into the model along with the question as", "mon presupposition of all normal questions. If pA", "intermediate steps.", "is recognized as False in the next step, the question" ]

{ "arxiv_id": "2310.16147", "language": "en", "page_idx": 1 }

[ [ "0.1587", "0.08435", "0.2024", "0.08435", "0.2024", "0.0956", "0.1587", "0.0956" ], [ "0.2805", "0.08435", "0.3225", "0.08435", "0.3225", "0.0956", "0.2805", "0.0956" ], [ "0.3997", "0.08435", "0.4492", "0.08435", "0.4492", "0.0956", "0.3997", "0.0956" ], [ "0.529", "0.0849", "0.881", "0.0849", "0.881", "0.0986", "0.529", "0.0986" ], [ "0.168", "0.0956", "0.1973", "0.0986", "0.1948", "0.10925", "0.1646", "0.1057" ], [ "0.288", "0.0968", "0.315", "0.0968", "0.315", "0.1051", "0.288", "0.1051" ], [ "0.3804", "0.0974", "0.4014", "0.1027", "0.3972", "0.111", "0.3762", "0.1057" ], [ "0.4492", "0.0998", "0.4727", "0.0998", "0.4727", "0.1081", "0.4492", "0.1081" ], [ "0.511", "0.10156", "0.882", "0.10156", "0.882", "0.1152", "0.511", "0.1152" ], [ "0.3284", "0.1027", "0.345", "0.1027", "0.345", "0.1069", "0.3284", "0.1069" ], [ "0.2083", "0.1069", "0.2258", "0.1069", "0.2258", "0.10986", "0.2083", "0.10986" ], [ "0.4148", "0.1069", "0.4358", "0.1069", "0.4358", "0.114", "0.4148", "0.114" ], [ "0.511", "0.11816", "0.881", "0.11816", "0.881", "0.1318", "0.511", "0.1318" ], [ "0.2957", "0.1277", "0.309", "0.1277", "0.309", "0.1312", "0.2957", "0.1312" ], [ "0.3315", "0.1277", "0.3477", "0.1277", "0.3477", "0.1318", "0.3315", "0.1318" ], [ "0.4517", "0.1318", "0.465", "0.1318", "0.465", "0.1354", "0.4517", "0.1354" ], [ "0.511", "0.1342", "0.882", "0.1342", "0.882", "0.1478", "0.511", "0.1478" ], [ "0.513", "0.1497", "0.882", "0.1497", "0.882", "0.1633", "0.513", "0.1633" ], [ "0.1351", "0.1567", "0.1561", "0.1567", "0.1561", "0.1644", "0.1351", "0.1644" ], [ "0.1721", "0.1609", "0.1897", "0.1609", "0.1897", "0.165", "0.1721", "0.165" ], [ "0.2551", "0.1561", "0.2747", "0.1561", "0.2747", "0.1633", "0.2551", "0.1633" ], [ "0.4165", "0.1633", "0.434", "0.1633", "0.434", "0.1675", "0.4165", "0.1675" ], [ "0.511", "0.1663", "0.882", "0.1663", "0.882", "0.1799", "0.511", "0.1799" ], [ "0.3342", "0.1829", "0.3435", "0.1829", "0.3435", "0.187", "0.3342", "0.187" ], [ "0.513", "0.1829", "0.882", "0.1829", "0.882", "0.1965", "0.513", "0.1965" ], [ "0.3855", "0.1841", "0.3914", "0.1841", "0.3914", "0.1895", "0.3855", "0.1895" ], [ "0.1411", "0.1841", "0.1486", "0.1841", "0.1486", "0.187", "0.1411", "0.187" ], [ "0.4207", "0.1841", "0.4282", "0.1841", "0.4282", "0.187", "0.4207", "0.187" ], [ "0.2241", "0.1948", "0.2678", "0.1948", "0.2678", "0.199", "0.2241", "0.199" ], [ "0.3418", "0.1924", "0.392", "0.1924", "0.392", "0.1996", "0.3418", "0.1996" ], [ "0.51", "0.1965", "0.854", "0.1978", "0.854", "0.212", "0.51", "0.2114" ], [ "0.1184", "0.2173", "0.4844", "0.2173", "0.4844", "0.228", "0.1184", "0.228" ], [ "0.513", "0.2214", "0.882", "0.2214", "0.882", "0.2363", "0.513", "0.2363" ], [ "0.1142", "0.2299", "0.4878", "0.2286", "0.4878", "0.2434", "0.1142", "0.2446" ], [ "0.511", "0.2382", "0.8833", "0.2382", "0.8833", "0.253", "0.511", "0.253" ], [ "0.11584", "0.2446", "0.2393", "0.2446", "0.2393", "0.2583", "0.11584", "0.2583" ], [ "0.509", "0.253", "0.882", "0.2542", "0.882", "0.2683", "0.509", "0.2673" ], [ "0.378", "0.2673", "0.4727", "0.282", "0.4668", "0.3005", "0.372", "0.2861" ], [ "0.2377", "0.2756", "0.2822", "0.2766", "0.2822", "0.2852", "0.2377", "0.284" ], [ "0.511", "0.2708", "0.8833", "0.2708", "0.8833", "0.284", "0.511", "0.284" ], [ "0.189", "0.282", "0.2334", "0.282", "0.2334", "0.2922", "0.189", "0.2922" ], [ "0.241", "0.2852", "0.2805", "0.2852", "0.2805", "0.2922", "0.241", "0.2922" ], [ "0.3047", "0.2861", "0.315", "0.2861", "0.315", "0.2903", "0.3047", "0.2903" ], [ "0.3535", "0.284", "0.388", "0.284", "0.388", "0.2922", "0.3535", "0.2922" ], [ "0.511", "0.2861", "0.8833", "0.2861", "0.8833", "0.2998", "0.511", "0.2998" ], [ "0.1377", "0.2964", "0.2301", "0.2964", "0.2301", "0.308", "0.1377", "0.308" ], [ "0.2451", "0.2964", "0.2805", "0.2964", "0.2805", "0.308", "0.2451", "0.308" ], [ "0.294", "0.2976", "0.4055", "0.2976", "0.4055", "0.308", "0.294", "0.308" ], [ "0.4048", "0.2964", "0.4785", "0.2957", "0.4785", "0.3071", "0.4048", "0.308" ], [ "0.511", "0.3018", "0.882", "0.3018", "0.882", "0.3154", "0.511", "0.3154" ], [ "0.2418", "0.31", "0.2805", "0.31", "0.2805", "0.3218", "0.2418", "0.3218" ], [ "0.294", "0.31", "0.4048", "0.31", "0.4048", "0.3218", "0.294", "0.3218" ], [ "0.4106", "0.31", "0.4768", "0.31", "0.4768", "0.3218", "0.4106", "0.3218" ], [ "0.1814", "0.3142", "0.1914", "0.3142", "0.1914", "0.3184", "0.1814", "0.3184" ], [ "0.1377", "0.3113", "0.2284", "0.3113", "0.2284", "0.3218", "0.1377", "0.3218" ], [ "0.51", "0.3176", "0.881", "0.3176", "0.881", "0.3308", "0.51", "0.3308" ], [ "0.189", "0.3247", "0.2301", "0.3247", "0.2301", "0.335", "0.189", "0.335" ], [ "0.288", "0.3247", "0.4055", "0.3247", "0.4055", "0.335", "0.288", "0.335" ], [ "0.4106", "0.3237", "0.4802", "0.3237", "0.4802", "0.335", "0.4106", "0.335" ], [ "0.2468", "0.326", "0.2805", "0.326", "0.2805", "0.3342", "0.2468", "0.3342" ], [ "0.1604", "0.3289", "0.168", "0.3289", "0.168", "0.332", "0.1604", "0.332" ], [ "0.511", "0.3333", "0.882", "0.3342", "0.882", "0.3486", "0.511", "0.3474" ], [ "0.11755", "0.3552", "0.4902", "0.3552", "0.4902", "0.3682", "0.11755", "0.3682" ], [ "0.511", "0.3499", "0.881", "0.3499", "0.881", "0.3635", "0.511", "0.3635" ], [ "0.11755", "0.3694", "0.4902", "0.3694", "0.4902", "0.383", "0.11755", "0.383" ], [ "0.511", "0.3652", "0.882", "0.3665", "0.882", "0.3813", "0.511", "0.3801" ], [ "0.11584", "0.3843", "0.4895", "0.3843", "0.4895", "0.398", "0.11584", "0.398" ], [ "0.511", "0.383", "0.882", "0.383", "0.882", "0.3967", "0.511", "0.3967" ], [ "0.11584", "0.3984", "0.3745", "0.3984", "0.3745", "0.412", "0.11584", "0.412" ], [ "0.513", "0.3997", "0.882", "0.3997", "0.882", "0.4133", "0.513", "0.4133" ], [ "0.513", "0.415", "0.7466", "0.415", "0.7466", "0.4287", "0.513", "0.4287" ], [ "0.11755", "0.4382", "0.4895", "0.4382", "0.4895", "0.4531", "0.11755", "0.4531" ], [ "0.511", "0.4382", "0.882", "0.4382", "0.882", "0.452", "0.511", "0.452" ], [ "0.11755", "0.4548", "0.4895", "0.4548", "0.4895", "0.4697", "0.11755", "0.4697" ], [ "0.51", "0.4548", "0.8833", "0.4548", "0.8833", "0.4685", "0.51", "0.4685" ], [ "0.1142", "0.4697", "0.4492", "0.4702", "0.4492", "0.485", "0.1142", "0.484" ], [ "0.511", "0.4702", "0.8833", "0.4702", "0.8833", "0.484", "0.511", "0.484" ], [ "0.511", "0.4863", "0.8833", "0.4868", "0.8833", "0.5005", "0.511", "0.4995" ], [ "0.11755", "0.4976", "0.293", "0.4976", "0.293", "0.511", "0.11755", "0.511" ], [ "0.51", "0.502", "0.882", "0.503", "0.882", "0.5166", "0.51", "0.5156" ], [ "0.11755", "0.5186", "0.4878", "0.5186", "0.4878", "0.532", "0.11755", "0.532" ], [ "0.511", "0.517", "0.882", "0.5186", "0.882", "0.533", "0.511", "0.532" ], [ "0.11584", "0.533", "0.4878", "0.5337", "0.4878", "0.549", "0.11584", "0.5474" ], [ "0.513", "0.535", "0.882", "0.535", "0.882", "0.549", "0.513", "0.549" ], [ "0.11755", "0.5503", "0.4895", "0.5503", "0.4895", "0.564", "0.11755", "0.564" ], [ "0.513", "0.552", "0.881", "0.552", "0.881", "0.5654", "0.513", "0.5654" ], [ "0.11584", "0.5664", "0.309", "0.5664", "0.309", "0.58", "0.11584", "0.58" ], [ "0.511", "0.5664", "0.881", "0.5664", "0.881", "0.58", "0.511", "0.58" ], [ "0.11755", "0.582", "0.4878", "0.582", "0.4878", "0.5967", "0.11755", "0.5967" ], [ "0.513", "0.585", "0.881", "0.585", "0.881", "0.5957", "0.513", "0.5957" ], [ "0.11584", "0.597", "0.4895", "0.5986", "0.4895", "0.6133", "0.11584", "0.6123" ], [ "0.509", "0.5986", "0.881", "0.5986", "0.881", "0.6123", "0.509", "0.6123" ], [ "0.11584", "0.615", "0.4878", "0.615", "0.4878", "0.629", "0.11584", "0.629" ], [ "0.5073", "0.6143", "0.8447", "0.615", "0.8447", "0.63", "0.5073", "0.629" ], [ "0.11584", "0.6313", "0.4878", "0.6313", "0.4878", "0.6445", "0.11584", "0.6445" ], [ "0.11755", "0.648", "0.4895", "0.648", "0.4895", "0.661", "0.11755", "0.661" ], [ "0.1142", "0.662", "0.4902", "0.663", "0.4902", "0.6777", "0.1142", "0.677" ], [ "0.517", "0.6714", "0.523", "0.6714", "0.523", "0.677", "0.517", "0.677" ], [ "0.54", "0.6675", "0.6406", "0.6675", "0.6406", "0.6816", "0.54", "0.6816" ], [ "0.11584", "0.6787", "0.4895", "0.6787", "0.4895", "0.6924", "0.11584", "0.6924" ], [ "0.11755", "0.6953", "0.4895", "0.6953", "0.4895", "0.709", "0.11755", "0.709" ], [ "0.11755", "0.712", "0.2896", "0.712", "0.2896", "0.7256", "0.11755", "0.7256" ], [ "0.513", "0.7114", "0.881", "0.7114", "0.881", "0.7256", "0.513", "0.7256" ], [ "0.11755", "0.727", "0.4937", "0.727", "0.4937", "0.7407", "0.11755", "0.7407" ], [ "0.511", "0.728", "0.881", "0.727", "0.881", "0.741", "0.511", "0.742" ], [ "0.11755", "0.7437", "0.4902", "0.7437", "0.4902", "0.7573", "0.11755", "0.7573" ], [ "0.511", "0.7437", "0.882", "0.742", "0.882", "0.7573", "0.511", "0.7583" ], [ "0.11755", "0.7603", "0.4895", "0.7603", "0.4895", "0.774", "0.11755", "0.774" ], [ "0.51", "0.7603", "0.882", "0.7603", "0.882", "0.774", "0.51", "0.774" ], [ "0.11755", "0.7754", "0.4878", "0.7754", "0.4878", "0.789", "0.11755", "0.789" ], [ "0.51", "0.775", "0.881", "0.775", "0.881", "0.788", "0.51", "0.788" ], [ "0.11584", "0.792", "0.4878", "0.792", "0.4878", "0.8057", "0.11584", "0.8057" ], [ "0.51", "0.792", "0.882", "0.792", "0.882", "0.8057", "0.51", "0.8057" ], [ "0.11755", "0.807", "0.4895", "0.807", "0.4895", "0.821", "0.11755", "0.821" ], [ "0.511", "0.807", "0.882", "0.807", "0.882", "0.821", "0.511", "0.821" ], [ "0.1142", "0.8237", "0.4895", "0.8223", "0.4895", "0.8374", "0.1142", "0.8384" ], [ "0.51", "0.8247", "0.8833", "0.8247", "0.8833", "0.8384", "0.51", "0.8384" ], [ "0.11755", "0.8403", "0.4895", "0.8403", "0.4895", "0.854", "0.11755", "0.854" ], [ "0.509", "0.839", "0.882", "0.8403", "0.882", "0.855", "0.509", "0.854" ], [ "0.11755", "0.8555", "0.4895", "0.8555", "0.4895", "0.869", "0.11755", "0.869" ], [ "0.509", "0.857", "0.882", "0.857", "0.882", "0.8706", "0.509", "0.8706" ], [ "0.11584", "0.8726", "0.4895", "0.8726", "0.4895", "0.8857", "0.11584", "0.8857" ], [ "0.51", "0.8726", "0.882", "0.8726", "0.882", "0.8857", "0.51", "0.8857" ], [ "0.11584", "0.887", "0.4895", "0.888", "0.4895", "0.903", "0.11584", "0.902" ], [ "0.51", "0.888", "0.882", "0.888", "0.882", "0.903", "0.51", "0.903" ], [ "0.11584", "0.905", "0.2795", "0.905", "0.2795", "0.9185", "0.11584", "0.9185" ], [ "0.511", "0.905", "0.8833", "0.905", "0.8833", "0.9185", "0.511", "0.9185" ] ]

[ "(a) ASQA", "(b) (QA)", "(c) BioASQ", "We also report the final performance of Pre-", "16.7%", "13.3%", "23.3%", "20.0%", "WoMe and vanilla LLM when given correctly and", "23.3", "", "36.7%", "incorrectly classifying feedback (Table 3 Perfor", "3.3", "3.39", "00", "mace). Specifically, we report D-F1 for ASQA and", "Rouge-L for BIOASQ and (QA)? of Pre WoMe and", "43.3%", "86:79", "46.7", "3.39", "vanilla GPT-4. It can be seen that regardless of the", "CP", "question type, performance gains increase when", "O", "n", "", "OursWins", "OursLoses", "feedback correctly classifies the question type", "Figure 3:Human evaluation results for each dataset. We", "Importance of PresuppositionsDo presupposi-", "compare Pre WoMewith vanilla LLM (GPT-4) over 90", "tions really contribute to generating better feedback", "questions in total.", "and action? We explore this question by omitting", "Peformanincoret", "Predictior", "presupposition extraction step and generating feed-", "Normal", "Ambig.", "PE", "Correct", "back and action with GPT-4 just based on the ques-", "Normal 74.9%", "18.5%", "6.6%23.19/21.26", "19.03 /18.25", "tion. As shown in Table 2 (w/o. Presup), PreWoMe", "35.9%", "9.5% 139.70 /34.04", "43.65 /41.51", "", "Ambig.54.6%", "without presuppositions extraction shows poor per", "11.6%", "74.4%30.45/22.38", "21.93/24.02", "14.0%", "", "formance, highlighting the role of presuppositions", "Table 3: Prediction shows distribution of predicted ques-", "as working memory. One interesting observation is", "tion type of PreWoMe using generated feedback. Per-", "that the performance of Pre WoMe without presup", "formance indicates result of (Pre WoMe / vanilla LLM)", "position is worse than the vanilla LLM, indicating", "with correct and incorrect predictions.", "that misleading (F,A) becomes a noise, consistent", "with findings in subsection 4.1.", "more aspects of ambiguity and correcting FP when", "Importance of Feedback, ActionSimilarly, we", "facing misleading questions. The prompt used for", " also explore the importance of Feedback and Ac-", "human evaluation is provided in AppendixF.", "tion. We omit the feedback and action generation", "stepand justfeed thequestions withthepresuppo-", "4.3Ablation Studies", "sitions that are generated in the previous step into", "For ablation studies, we first conduct analysis on", "GPT-4. Table 2 (w/o. (F,A)) shows that using pre-", "the LLM's capability of handling each intermediate", "suppositions without feedback/action generation", "step, then verify the importance of each intermedi-", "step shows lower performance compared to the full", "ate step within PreWoMe.", "PreWoMe for misleading questions.This result", "Presupposition AnalysisTo analyze the LLM's", "demonstratestheimportance offeedback/action.", "capability of generating the correct presuppositions,", "and also that generating presuppositions itself is", "we randomly sample 150 questions and manually", " not sufficient to handle misleading questions.", "annotate whether the corresponding 395 generated", "presuppositions were indeed assumed by the ques-", "tion. The accuracy of presuppositions was ex-", "5", "Conclusion", "tremely high (391 correct out of 395 total ~ 99%),", "demonstrating GPT-4's capability of extracting ac-", "curate presuppositions.", "We propose Pre WoMe, a new approach that handles", "Feedback AnalysisAs mentioned in subsec-", "any type of question in a unified way by exploiting", "tion 3.2, feedback works as a classifier that de-", "presuppositions as working memory. Pre WoMe", "termines the type of the given question. We extract", "uses intermediate steps to guide the model to gen", "the question type predicted by the feedback using", "erate presuppositions along with useful feedback", "a rule-based approach and report the distribution in", "and action for itself. Our experiments show that", "Table 3 (Prediction). Feedback effectively detects", "PreWoMe boosts the performance of LLMs on dif-", "normal questions and questions with FP (74.9%", "ferent types of questions without parameter updates.", "and 74.4%). However, ambiguous questions have", "This is achieved even though the model is given", "low detect ratio, consistent with Min et al. (2020).", "no information about the type of incoming ques-", "This indicates that there is more room to optimize", "tion, demonstrating that Pre WoMe is effective for a", "for detecting ambiguity in the question, which we", "real-world setting and will be a significant stepping", "leave for future work.", "stone to future works tackling real-world QA tasks." ]

{ "arxiv_id": "2310.16147", "language": "en", "page_idx": 4 }

[ [ "0.1748", "0.0341", "0.4795", "0.0341", "0.4795", "0.04736", "0.1748", "0.04736" ], [ "0.1748", "0.101", "0.7183", "0.1016", "0.7183", "0.1193", "0.1748", "0.1187" ], [ "0.1765", "0.1256", "0.6357", "0.127", "0.6357", "0.1445", "0.1765", "0.1433" ], [ "0.183", "0.1692", "0.6763", "0.1704", "0.6763", "0.1849", "0.183", "0.1843" ], [ "0.1847", "0.1863", "0.2998", "0.1863", "0.2998", "0.1995", "0.1847", "0.1995" ], [ "0.1854", "0.1995", "0.5986", "0.1995", "0.5986", "0.2134", "0.1854", "0.2134" ], [ "0.455", "0.252", "0.546", "0.252", "0.546", "0.2632", "0.455", "0.2632" ], [ "0.02287", "0.2615", "0.05963", "0.2615", "0.05963", "0.4072", "0.02287", "0.4072" ], [ "0.2336", "0.2822", "0.767", "0.2822", "0.767", "0.2954", "0.2336", "0.2954" ], [ "0.232", "0.2954", "0.767", "0.2966", "0.767", "0.31", "0.232", "0.3093" ], [ "0.2345", "0.31", "0.767", "0.31", "0.767", "0.324", "0.2345", "0.324" ], [ "0.2336", "0.324", "0.7656", "0.324", "0.7656", "0.3372", "0.2336", "0.3372" ], [ "0.2336", "0.3372", "0.767", "0.3372", "0.767", "0.351", "0.2336", "0.351" ], [ "0.2336", "0.3516", "0.7656", "0.3516", "0.7656", "0.3655", "0.2336", "0.3655" ], [ "0.2336", "0.3655", "0.7656", "0.3655", "0.7656", "0.379", "0.2336", "0.379" ], [ "0.03513", "0.385", "0.04004", "0.385", "0.04004", "0.3926", "0.03513", "0.3926" ], [ "0.2345", "0.3801", "0.7646", "0.3801", "0.7646", "0.3901", "0.2345", "0.3901" ], [ "0.2336", "0.3926", "0.767", "0.3926", "0.767", "0.406", "0.2336", "0.406" ], [ "0.02287", "0.3997", "0.0621", "0.3997", "0.0621", "0.4924", "0.02287", "0.4924" ], [ "0.2336", "0.406", "0.7656", "0.406", "0.7656", "0.4192", "0.2336", "0.4192" ], [ "0.2345", "0.4204", "0.7646", "0.4204", "0.7646", "0.4312", "0.2345", "0.4312" ], [ "0.2336", "0.433", "0.7656", "0.433", "0.7656", "0.4463", "0.2336", "0.4463" ], [ "0.2336", "0.4463", "0.7656", "0.4463", "0.7656", "0.4602", "0.2336", "0.4602" ], [ "0.232", "0.459", "0.7656", "0.4602", "0.7656", "0.4749", "0.232", "0.4734" ], [ "0.2336", "0.4749", "0.7656", "0.4749", "0.7656", "0.488", "0.2336", "0.488" ], [ "0.02124", "0.4893", "0.05963", "0.4893", "0.05963", "0.712", "0.02124", "0.712" ], [ "0.2336", "0.488", "0.7656", "0.488", "0.7656", "0.502", "0.2336", "0.502" ], [ "0.232", "0.5005", "0.7656", "0.502", "0.7656", "0.5166", "0.232", "0.515" ], [ "0.2336", "0.515", "0.7656", "0.515", "0.7656", "0.529", "0.2336", "0.529" ], [ "0.232", "0.529", "0.513", "0.528", "0.513", "0.5425", "0.232", "0.5435" ], [ "0.1765", "0.5747", "0.1847", "0.5713", "0.1887", "0.579", "0.1805", "0.582" ], [ "0.2009", "0.571", "0.3367", "0.571", "0.3367", "0.5854", "0.2009", "0.5854" ], [ "0.1748", "0.604", "0.824", "0.604", "0.824", "0.617", "0.1748", "0.617" ], [ "0.1732", "0.617", "0.826", "0.617", "0.826", "0.632", "0.1732", "0.632" ], [ "0.1748", "0.631", "0.824", "0.631", "0.824", "0.6445", "0.1748", "0.6445" ], [ "0.1765", "0.6465", "0.8228", "0.6465", "0.8228", "0.657", "0.1765", "0.657" ], [ "0.1732", "0.659", "0.824", "0.659", "0.824", "0.6724", "0.1732", "0.6724" ], [ "0.1732", "0.672", "0.824", "0.6704", "0.824", "0.685", "0.1732", "0.686" ], [ "0.1748", "0.686", "0.824", "0.686", "0.824", "0.6997", "0.1748", "0.6997" ], [ "0.1724", "0.6987", "0.818", "0.6997", "0.818", "0.714", "0.1724", "0.7134" ], [ "0.1748", "0.72", "0.824", "0.72", "0.824", "0.7344", "0.1748", "0.7344" ], [ "0.1732", "0.735", "0.8267", "0.735", "0.8267", "0.7485", "0.1732", "0.7485" ], [ "0.1732", "0.7476", "0.826", "0.7485", "0.826", "0.762", "0.1732", "0.761" ], [ "0.1748", "0.762", "0.824", "0.762", "0.824", "0.776", "0.1748", "0.776" ], [ "0.1732", "0.776", "0.826", "0.7744", "0.826", "0.789", "0.1732", "0.7905" ], [ "0.1732", "0.7886", "0.824", "0.789", "0.824", "0.8037", "0.1732", "0.803" ], [ "0.1748", "0.8047", "0.8228", "0.8047", "0.8228", "0.815", "0.1748", "0.815" ], [ "0.1748", "0.8174", "0.824", "0.8174", "0.824", "0.8306", "0.1748", "0.8306" ], [ "0.1732", "0.8306", "0.824", "0.8306", "0.824", "0.8447", "0.1732", "0.8447" ], [ "0.1765", "0.845", "0.8228", "0.845", "0.8228", "0.856", "0.1765", "0.856" ], [ "0.1732", "0.857", "0.824", "0.858", "0.824", "0.8726", "0.1732", "0.872" ], [ "0.1748", "0.872", "0.824", "0.872", "0.824", "0.886", "0.1748", "0.886" ], [ "0.1765", "0.887", "0.4714", "0.887", "0.4714", "0.898", "0.1765", "0.898" ], [ "0.1978", "0.911", "0.5933", "0.911", "0.5933", "0.925", "0.1978", "0.925" ], [ "0.4927", "0.9487", "0.5034", "0.9487", "0.5034", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2023", "TREEFORMER: DENSE GRADIENT TREES FOR", "EFFICIENT ATTENTION COMPUTATION", "Lovish Madaan, Srinadh Bhojanapalli, Himanshu Jain & Prateek Jain", "GoogleResearch", "(lovishm,bsrinadh,himj,prajain)@google.com", "ABSTRACT", "", "Standard inference and training with transformer based architectures scale quadrat-", "ically with input sequence length.This is prohibitively large for a variety of ap-", "plications especially in web-page translation, query-answering etc. Consequently.", "several approaches have been developed recently to speedup attention computation", "by enforcing different attention structures such as sparsity (Zaheer et al., 2020),", "low-rank (Wang et al., 2020), approximating attention using kernels (Choromanski", "et al., 2021). In this work, we view attention computation as that of nearest neighbor", "", "retrieval,andusedecisiontreebasedhierarchicalnavigationtoreducetheretrieval", "cost per query token from linear in sequence length to nearly logarithmic. Based", "['S]", "on such hierarchical navigation, we design Treeformer which can use one of two", "efficientattentionlayers-TF-ATTENTIONandTC-ATTENTION.TF-ATTENTION", "computes the attention in a fine-grained style, while TC-ATTENTION is a coarse", "attention layer which also ensures that the gradients are \"dense\". To optimize", "such challenging discrete layers, we propose a two-level bootstrapped training", "method. Using extensive experiments on standard NLP benchmarks, especially for", "01060800A0", "long-sequences, we demonstrate that our TREEFORMER architecture can be almost", "as accurate as baseline Transformer while using 3Ox lesser FLOPs in the attention", "layer. Compared to Linformer, the accuracy can be as much as 12% higher while", "using similar FLOPs in the attention layer.", "1", "INTRODUCTION", "Self attention layer is the key component of Transformers (Vaswani et al., 2017), enabling them to", "achieve state of the art performance across tasks in Natural Language Processing (Devlin et al., 2019;", "Radford et al., 2019; Raffel et al., 2019) and Computer Vision (Dosovitskiy et al., 2021). Attention", "computationscalesquadratically(n)withtheinput sequencelength(n),makingitakeybottleneck", "in scaling Transformers to long inputs.This has resulted in a lot of alternate proposals to efficiently", "compute attention using different approximations. However, as shown in Tay et al. (2021), methods", "that offer dramatic speedups (Wang et al., 2020; Choromanski et al., 2021) suffer in accuracy, and", "methods that match accuracy of Transformer (Zaheer et al., 2020) don't offer significant speedups.", "In this paper we develop a novel efficient attention mechanism using input dependent sparsity structure", "in attention computation, motivated by the analysis showing the sparse nature of attention (Shi et al.,", "2021;Gupta et al.,2021).We propose to use decision trees to efficiently compute attention by only", "retrieving the top nearest neighboring keys for a given query. Decision trees are low cost, hierarchical", "non-linear models that have been popular in both supervised (Chen & Guestrin, 2016; Quinlan, 2014)", "and unsupervised (Duda et al., 1973) learning. Given a vector, each node of the decision tree decides", "whethertonavigatetotheleftortherightchilddependingonthesignofasimplelinearprojection", "We learn decision trees in each attention layer of a Transformer model so that a query pays attention", "only to keys that map to the same leaf nodes' of the decision trees. We refer to this as TF-ATTENTION.", "Suchsparsedecisiontreesthatretrieveonlya singleleafnodearetypicallyhardtotrainduetotheir", "discrete nature. Further, queries that pay strong attention to multiple keys result in bunching many", "keys together in a single leaf node creating imbalance. The key computational advantage of decision", "treesisrealized onlywhentheyarebalanced.", "'Leaf nodes are the bottom nodes of a decision tree with O children.", "1" ]

{ "arxiv_id": "2208.09015", "language": "en", "page_idx": 0 }

[ [ "0.1724", "0.03284", "0.4795", "0.0322", "0.4795", "0.04672", "0.1724", "0.04736" ], [ "0.1748", "0.1048", "0.824", "0.1048", "0.824", "0.1193", "0.1748", "0.1193" ], [ "0.1732", "0.1193", "0.826", "0.1193", "0.826", "0.1332", "0.1732", "0.1332" ], [ "0.1732", "0.1332", "0.826", "0.1332", "0.826", "0.1465", "0.1732", "0.1465" ], [ "0.1748", "0.1465", "0.824", "0.1465", "0.824", "0.1604", "0.1748", "0.1604" ], [ "0.1765", "0.1622", "0.824", "0.1622", "0.824", "0.1724", "0.1765", "0.1724" ], [ "0.1765", "0.1755", "0.8228", "0.1755", "0.8228", "0.1863", "0.1765", "0.1863" ], [ "0.1724", "0.1875", "0.826", "0.1881", "0.826", "0.2026", "0.1724", "0.202" ], [ "0.1732", "0.2084", "0.824", "0.2084", "0.824", "0.2229", "0.1732", "0.2229" ], [ "0.1748", "0.2229", "0.826", "0.2229", "0.826", "0.2361", "0.1748", "0.2361" ], [ "0.1748", "0.2361", "0.826", "0.2361", "0.826", "0.25", "0.1748", "0.25" ], [ "0.1732", "0.2507", "0.824", "0.2507", "0.824", "0.2644", "0.1732", "0.2644" ], [ "0.1732", "0.262", "0.824", "0.2632", "0.824", "0.2778", "0.1732", "0.2766" ], [ "0.1748", "0.2778", "0.824", "0.2778", "0.824", "0.2917", "0.1748", "0.2917" ], [ "0.1748", "0.2922", "0.3628", "0.2922", "0.3628", "0.3062", "0.1748", "0.3062" ], [ "0.1748", "0.3125", "0.826", "0.3125", "0.826", "0.3257", "0.1748", "0.3257" ], [ "0.1732", "0.3257", "0.824", "0.3257", "0.824", "0.339", "0.1732", "0.339" ], [ "0.1748", "0.3403", "0.824", "0.3403", "0.824", "0.3542", "0.1748", "0.3542" ], [ "0.1748", "0.3542", "0.8228", "0.3542", "0.8228", "0.3643", "0.1748", "0.3643" ], [ "0.1732", "0.3674", "0.826", "0.3674", "0.826", "0.3806", "0.1732", "0.3806" ], [ "0.1732", "0.3806", "0.4248", "0.3806", "0.4248", "0.3945", "0.1732", "0.3945" ], [ "0.1732", "0.4016", "0.826", "0.4016", "0.826", "0.4153", "0.1732", "0.4153" ], [ "0.1748", "0.4172", "0.824", "0.4172", "0.824", "0.428", "0.1748", "0.428" ], [ "0.1732", "0.43", "0.824", "0.43", "0.824", "0.443", "0.1732", "0.443" ], [ "0.1748", "0.443", "0.824", "0.443", "0.824", "0.457", "0.1748", "0.457" ], [ "0.1748", "0.457", "0.824", "0.457", "0.824", "0.4702", "0.1748", "0.4702" ], [ "0.1765", "0.473", "0.8228", "0.473", "0.8228", "0.483", "0.1765", "0.483" ], [ "0.1748", "0.4841", "0.8228", "0.4841", "0.8228", "0.4988", "0.1748", "0.4988" ], [ "0.1732", "0.4968", "0.826", "0.4976", "0.826", "0.512", "0.1732", "0.511" ], [ "0.1748", "0.511", "0.824", "0.511", "0.824", "0.5244", "0.1748", "0.5244" ], [ "0.1732", "0.526", "0.8228", "0.526", "0.8228", "0.539", "0.1732", "0.539" ], [ "0.1748", "0.539", "0.819", "0.539", "0.819", "0.553", "0.1748", "0.553" ], [ "0.1765", "0.5747", "0.1912", "0.5747", "0.1912", "0.5854", "0.1765", "0.5854" ], [ "0.2018", "0.5728", "0.636", "0.5737", "0.636", "0.5884", "0.2018", "0.587" ], [ "0.1748", "0.605", "0.2091", "0.605", "0.2091", "0.616", "0.1748", "0.616" ], [ "0.2091", "0.606", "0.2974", "0.606", "0.2974", "0.617", "0.2091", "0.617" ], [ "0.1724", "0.627", "0.826", "0.629", "0.826", "0.6465", "0.1724", "0.6445" ], [ "0.1732", "0.6416", "0.8267", "0.6416", "0.8267", "0.659", "0.1732", "0.659" ], [ "0.1748", "0.6577", "0.613", "0.6577", "0.613", "0.672", "0.1748", "0.672" ], [ "0.5156", "0.688", "0.625", "0.688", "0.625", "0.7026", "0.5156", "0.7026" ], [ "0.3022", "0.6963", "0.513", "0.6963", "0.513", "0.711", "0.3022", "0.711" ], [ "0.648", "0.696", "0.694", "0.69", "0.6978", "0.7056", "0.6504", "0.711" ], [ "0.8022", "0.6963", "0.826", "0.6963", "0.826", "0.711", "0.8022", "0.711" ], [ "0.564", "0.7056", "0.5933", "0.7056", "0.5933", "0.7227", "0.564", "0.7227" ], [ "0.5435", "0.726", "0.5596", "0.726", "0.5596", "0.7407", "0.5435", "0.7407" ], [ "0.1732", "0.7524", "0.824", "0.7524", "0.824", "0.7666", "0.1732", "0.7666" ], [ "0.1748", "0.7666", "0.824", "0.7666", "0.824", "0.78", "0.1748", "0.78" ], [ "0.1732", "0.779", "0.826", "0.78", "0.826", "0.7944", "0.1732", "0.7935" ], [ "0.1724", "0.7925", "0.55", "0.7935", "0.55", "0.808", "0.1724", "0.8066" ], [ "0.1748", "0.815", "0.8267", "0.815", "0.8267", "0.829", "0.1748", "0.829" ], [ "0.1748", "0.829", "0.824", "0.829", "0.824", "0.8423", "0.1748", "0.8423" ], [ "0.1748", "0.8423", "0.824", "0.8423", "0.824", "0.856", "0.1748", "0.856" ], [ "0.1765", "0.856", "0.824", "0.856", "0.824", "0.8706", "0.1765", "0.8706" ], [ "0.1732", "0.8687", "0.824", "0.8706", "0.824", "0.8853", "0.1732", "0.8833" ], [ "0.1748", "0.884", "0.824", "0.884", "0.824", "0.898", "0.1748", "0.898" ], [ "0.1732", "0.8965", "0.824", "0.898", "0.824", "0.912", "0.1732", "0.911" ], [ "0.1748", "0.911", "0.824", "0.911", "0.824", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2023", "To address these issues we propose another attention mechanism TC-ATTENTION, that in addition", "to keys in the leaf node, pays attention to keys at all the intermediate nodes traversed by the query.", "Wefurther combine the retrieved valuevectors with query independent, but trainable weights.This", "allows us to learn a decision tree that clusters keys in a hierarchical and balanced manner. In both", "thesemethodswetrainthedecisiontreeinanend-to-endmannerwiththerestofthecomponentsof", "theTransformer.Sincethe decision trees themselves areinexpensiveto evaluate,this allows us to", "reduce attention computation cost dramatically, requiring typically only a few dot products per query.", "In our goal to train these decision trees in an end to end manner with the rest of the Transformer", "components we faced some challenges. Naively using the decision trees to restrict attention computa-", "tion results in poor optimization with training getting stuck at high loss. To solve this we introduce a", "novel bootstrapping method that allows us to gradually restrict attention computation in layers using", "decision trees. while this keeps training expensive initially, it still dramatically reduces inference", "costs. We further use Dense Gradient Trees (Karthikeyan et al., 2021), allowing us to use simple", "gradient based optimization.", "We evaluate the proposed architecture on popular NLP models such as BERT (Devlin et al., 2019),", "and on the Long Range Arena benchmark (LRA) (Tay et al., 2021), developed specifically to compare", "efficient attention methods on long sequence length tasks. We show that TREEFORMER is able to", "match/improvetheperformanceoverotherpopularmodelswhileachievinglowercomputationalcost", "for e.g. -TREEFORMER has 8-9x lesser FLOPs than models like BigBird (Zaheer et al., 2020) and", "Performer (Choromanski et al., 2021).", "Our contributions in this paper are (i) We pose attention computation as a nearest neighbor", "retrievalproblem,andproposeanovelarchitectureTREEFoRMER-usingdensegradienttreesfor", "retrieving topkeys to pay attention to for a given query.We develop two novel attention mechanisms", "using decision trees - that, for a given query, pay attention to only keys in the matching leaf node", "(TF-ATTENTION) or to all the keys along the traversed path (TC-ATTENTION). (i) We develop a", "novelbootstrappingmethodtograduallysparsifyattentioncomputationtherebyshowingthatdecision", "trees can be trained as a part of neural networks using back propagation. (ii) We experimentally", "show the effectiveness of the proposed architecture using BERT models and on LRA (Tay et al.,", "2021). In particular, both TF-ATTENTION and TC-ATTENTION have matching accuracy to baseline", "Transformer architecture, but with up to 30× cheaper (in FLOPs) attention layer compared to standard", "Transformers and up to 9× cheaper attention layer compared to SOTA BigBird (Zaheer et al., 2020).", "2", "PRELIMINARIES:ATTENTIONANDDECISIONTREES", "2.1", "ATTENTION", "Let Q, K, V IRnxd be the input query, key, and yalue matrices, where n is the sequence length and", "d is the embedding dimension, and wQ, wK, WV Rdxd are their respective projection matrices.", "The standard attention in each Transformer layer can be defined as", "[QWQ(KWK)", "Attention(Q, K, V) = softmax", "VWV", "(1)", "Va", "A", "Note that for the self attention layer, all the three matrices Q, K, V are same as the input to the", "layer. Computing both attention A and its product with the projected value matrix VWV is an", "O(n?d + d2n) operation. In particular, assuming cpqr cost for multiplying two matrices of size p × q", "and q X r, the main cost of this layer is: c(2n?d + 2d2n).", "This computation becomes a major bottleneck in applications which have large sequence lengths n.", "To mitigate this, several efficient attention mechanisms have been proposed (Child et al., 2019; Wang", "et al., 2020; Choromanski et al., 2021; Kitaev et al., 2020; Sun et al., 2021; Wang et al., 2022). Child", "et al. (2019) proposed sparse attention (A) with different sparsity patterns that reduce attention cost", "from O(n2) to O(n1.5). Later Zaheer et al. (2020); Yun et al. (2020) proved universal approximation", "power of Transformers with sparse attention, improving the cost to O(ns), with constant sparsity s", "per query. Alternately Wang et al. (2020) considered a low rank approximation to attention, reducing", "computation cost to O(nk) for rank k approximation. Choromanski et al. (2021) considered a kernel", "2" ]

{ "arxiv_id": "2208.09015", "language": "en", "page_idx": 1 }

[ [ "0.1748", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1748", "0.04736" ], [ "0.1748", "0.1048", "0.822", "0.1048", "0.822", "0.1193", "0.1748", "0.1193" ], [ "0.1748", "0.1187", "0.824", "0.1187", "0.824", "0.132", "0.1748", "0.132" ], [ "0.1748", "0.1332", "0.824", "0.1332", "0.824", "0.1465", "0.1748", "0.1465" ], [ "0.1748", "0.1465", "0.824", "0.1465", "0.824", "0.1604", "0.1748", "0.1604" ], [ "0.1732", "0.159", "0.824", "0.159", "0.824", "0.1724", "0.1732", "0.1724" ], [ "0.1732", "0.1724", "0.648", "0.1724", "0.648", "0.1863", "0.1732", "0.1863" ], [ "0.2133", "0.2008", "0.7837", "0.2008", "0.7837", "0.214", "0.2133", "0.214" ], [ "0.3774", "0.2299", "0.4346", "0.2299", "0.4346", "0.2443", "0.3774", "0.2443" ], [ "0.4763", "0.2299", "0.5215", "0.2299", "0.5215", "0.2443", "0.4763", "0.2443" ], [ "0.5767", "0.2299", "0.634", "0.2299", "0.634", "0.2443", "0.5767", "0.2443" ], [ "0.2778", "0.2311", "0.3464", "0.2311", "0.3464", "0.2424", "0.2778", "0.2424" ], [ "0.1912", "0.2311", "0.2296", "0.2311", "0.2296", "0.2443", "0.1912", "0.2443" ], [ "0.689", "0.2311", "0.7935", "0.2311", "0.7935", "0.2437", "0.689", "0.2437" ], [ "0.1929", "0.2468", "0.3513", "0.2468", "0.3513", "0.2615", "0.1929", "0.2615" ], [ "0.5513", "0.2474", "0.661", "0.2474", "0.661", "0.2615", "0.5513", "0.2615" ], [ "0.675", "0.2474", "0.7715", "0.2456", "0.7715", "0.26", "0.675", "0.262" ], [ "0.46", "0.2488", "0.536", "0.2488", "0.536", "0.26", "0.46", "0.26" ], [ "0.367", "0.2488", "0.4436", "0.2488", "0.4436", "0.26", "0.367", "0.26" ], [ "0.767", "0.2488", "0.808", "0.2488", "0.808", "0.26", "0.767", "0.26" ], [ "0.1765", "0.2874", "0.2091", "0.2874", "0.2091", "0.2986", "0.1765", "0.2986" ], [ "0.2075", "0.286", "0.3325", "0.286", "0.3325", "0.2998", "0.2075", "0.2998" ], [ "0.1748", "0.3125", "0.8228", "0.3125", "0.8228", "0.327", "0.1748", "0.327" ], [ "0.1748", "0.327", "0.824", "0.327", "0.824", "0.3403", "0.1748", "0.3403" ], [ "0.1748", "0.3403", "0.824", "0.3403", "0.824", "0.3542", "0.1748", "0.3542" ], [ "0.1765", "0.3547", "0.8203", "0.3547", "0.8203", "0.3655", "0.1765", "0.3655" ], [ "0.1748", "0.3674", "0.826", "0.3674", "0.826", "0.3806", "0.1748", "0.3806" ], [ "0.1748", "0.3806", "0.8228", "0.3806", "0.8228", "0.3958", "0.1748", "0.3958" ], [ "0.1732", "0.3945", "0.8228", "0.3958", "0.8228", "0.4092", "0.1732", "0.4077" ], [ "0.1732", "0.4092", "0.824", "0.4092", "0.824", "0.4224", "0.1732", "0.4224" ], [ "0.1732", "0.4224", "0.8228", "0.4224", "0.8228", "0.4363", "0.1732", "0.4363" ], [ "0.1748", "0.4363", "0.824", "0.4363", "0.824", "0.4495", "0.1748", "0.4495" ], [ "0.1748", "0.4507", "0.8228", "0.4507", "0.8228", "0.464", "0.1748", "0.464" ], [ "0.1748", "0.4634", "0.824", "0.4634", "0.824", "0.4766", "0.1748", "0.4766" ], [ "0.1748", "0.4778", "0.4944", "0.4778", "0.4944", "0.4912", "0.1748", "0.4912" ], [ "0.1708", "0.4912", "0.824", "0.4956", "0.824", "0.5176", "0.1708", "0.513" ], [ "0.1748", "0.513", "0.8228", "0.513", "0.8228", "0.528", "0.1748", "0.528" ], [ "0.1748", "0.526", "0.826", "0.5264", "0.826", "0.5444", "0.1748", "0.5435" ], [ "0.1732", "0.5415", "0.5864", "0.5415", "0.5864", "0.5596", "0.1732", "0.5596" ], [ "0.3447", "0.5645", "0.6562", "0.5645", "0.6562", "0.582", "0.3447", "0.582" ], [ "0.8013", "0.563", "0.826", "0.563", "0.826", "0.5806", "0.8013", "0.5806" ], [ "0.1724", "0.586", "0.824", "0.584", "0.824", "0.6016", "0.1724", "0.604" ], [ "0.1748", "0.603", "0.8228", "0.603", "0.8228", "0.616", "0.1748", "0.616" ], [ "0.1748", "0.616", "0.824", "0.616", "0.824", "0.63", "0.1748", "0.63" ], [ "0.1748", "0.63", "0.4739", "0.63", "0.4739", "0.6445", "0.1748", "0.6445" ], [ "0.1732", "0.664", "0.517", "0.664", "0.517", "0.6787", "0.1732", "0.6787" ], [ "0.1732", "0.6943", "0.822", "0.6943", "0.822", "0.7085", "0.1732", "0.7085" ], [ "0.1748", "0.7104", "0.8228", "0.7104", "0.8228", "0.7236", "0.1748", "0.7236" ], [ "0.1748", "0.7236", "0.3594", "0.7236", "0.3594", "0.7373", "0.1748", "0.7373" ], [ "0.3447", "0.7505", "0.4453", "0.7505", "0.4453", "0.765", "0.3447", "0.765" ], [ "0.464", "0.7485", "0.6543", "0.7456", "0.6543", "0.763", "0.464", "0.7666" ], [ "0.455", "0.765", "0.4617", "0.765", "0.4617", "0.7695", "0.455", "0.7695" ], [ "0.1748", "0.779", "0.824", "0.779", "0.824", "0.7935", "0.1748", "0.7935" ], [ "0.1765", "0.7954", "0.822", "0.7954", "0.822", "0.806", "0.1765", "0.806" ], [ "0.1748", "0.808", "0.824", "0.808", "0.824", "0.8213", "0.1748", "0.8213" ], [ "0.1748", "0.8213", "0.826", "0.8213", "0.826", "0.836", "0.1748", "0.836" ], [ "0.1748", "0.8413", "0.824", "0.8413", "0.824", "0.856", "0.1748", "0.856" ], [ "0.1748", "0.856", "0.824", "0.856", "0.824", "0.869", "0.1748", "0.869" ], [ "0.1732", "0.869", "0.824", "0.8706", "0.824", "0.884", "0.1732", "0.8833" ], [ "0.1724", "0.884", "0.8228", "0.884", "0.8228", "0.898", "0.1724", "0.898" ], [ "0.1748", "0.898", "0.824", "0.898", "0.824", "0.911", "0.1748", "0.911" ], [ "0.1748", "0.911", "0.6177", "0.911", "0.6177", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2023", "based approach to approximate attention computation reducing the cost to O(nr), where r is the", "dimension of the random features used to compute the kernel. We refer readers to Tay et al. (2020)", "for a detailed survey of efficient attention methods. In this paper, we propose a decision tree based", "mechanism to speed up the attention computation to O(nh), where h is the height of the decision tree", "We show that with TREEFORMER even a h as small as 8 is sufficient for most applications, leading to", "the least amount of FLOPs required, while maintaining the performance.", "Table 1: Comparison of attention computation cost (eq.equation 1) for different models.", "Linformer", "BigBird", "Performer", "Transformer", "Model", "TREEFORMERTF-A", "Attention cost | O(n?d + n2)", "O(rnd +n(r + d))", "O(hnd +(2h+1)", "O(snd + sn)", "O(knd +kn)", "- 1)d)", "2.2", "DECISION TREES", "Decision trees are a classical technique that enables partitioning of the input space in several oblique", "blocks that can be navigated efficiently using the hierarchical structure. Decision trees and similar hier", "archical structures are highly popular in supervised learning (XgBoost (Chen & Guestrin, 2016), C4.5", "(Ouinlan.2014)).unsupervisedlearning(hierarchical clustering(Dudaetal.,1973)),andretrieva", "(KD tree (Bentley, 1975). Generally, hierarchical retrieval techniques are nominally data-dependent,", "e.g., PCA-tree (Sproull, 1991) and Cover Trees (Beygelzimer et al., 2006). Our TC-ATTENTION", "and TF-ATTENTION methods combine hierarchical retrieval methods with supervised tree learning", "techniques to enable learning of the hierarchical attention layer in end-to-end differentiable manner", "See Figure I for overall representation of the hierarchical navigation based attention layer. Each", "internal node of the tree has a classifier that determines which children a given input should navigate", "to. While the classifier makes hard decision, we use technique from straight-through estimators", "(Bengio et al., 2013), and in particular, Dense Gradient Trees from Karthikeyan et al. (2021) to train", "the tree parameters using gradient style methods.", "Construction Formally, let T(0) represent a binary tree with height h and 2h nodes. Let 0(.5)", "denote the tree parameters at (l, j]-th node where l E [O, 1, ::: , h] is the node level and j e", "[0, 1, . .: , 2' - i] is the node index at level l. Let decision at each internal node to send the point q", "to left or the right child is based on SIGN(f(e(1,); q), where,", "f(t,(q) = f(0(t,s);q) = (w(1.j),q) +b(,j),", "(2)", "and d(l,j) = [w(1,j); b(:)) consists of a oblique hyperplane and a bias. Let Pr(l,q) denote the", "node-index of T that was traversed by q at level l. That is, Pr(h, q) denote the leaf node to which q", "is mapped. Also, let St.,(K) denote the indices of vectors from the key matrix K that traverses node", "(l, j). That is, i e St,(K) iff j = Pr(l, ki).", "3TREEFORMERATTENTIONVARIANTS", "In this section, we describe the decision tree attention mechanism which replaces the standard dot", "product attention (Vaswani et al., 2017) in our TREEFORMER architecture. Recall that for each query", "qi, the attention is given by:", "A; = softmax(", "qTWe(WK)TKT)VWV.", "1", "That is, attention computation for qi is equivalent to applying a linear scan over all keys, and", "computing valueas theweightedaverageofvalue ofmost similarkeys to i.Recentworkshave", "shown that we do not need to attend to all tokens in a given input sequence (Shi et al., 2021), and", "demonstrate that top-k attention is a good approximation for standard attention (Gupta et al., 2021).", "We propose to leverage this observation by using a decision tree based hierarchical navigation instead", "of linear scan to find the top-most similar keys. Naturally, accuracy of the method would depend on", "accuracy of retrieval of most similarkeys by the hierarchical navigation method.So,wepropose", "to learn the decision tree parameters so as to explicitly optimize the loss function. We develop two", "approaches to compute the value for a query q; based on its traversal through the tree. In the following", "subsections we present our two different tree based attention layers.", "3" ]

{ "arxiv_id": "2208.09015", "language": "en", "page_idx": 2 }

[ [ "0.1732", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1732", "0.04736" ], [ "0.1732", "0.1029", "0.523", "0.1042", "0.523", "0.1187", "0.1732", "0.11743" ], [ "0.1732", "0.13", "0.826", "0.13", "0.826", "0.1445", "0.1732", "0.1445" ], [ "0.1732", "0.1445", "0.455", "0.1445", "0.455", "0.1578", "0.1732", "0.1578" ], [ "0.2524", "0.1692", "0.524", "0.1704", "0.524", "0.1881", "0.2524", "0.1875" ], [ "0.5376", "0.1635", "0.665", "0.1635", "0.665", "0.178", "0.5376", "0.178" ], [ "0.6787", "0.1718", "0.74", "0.1654", "0.7427", "0.1843", "0.681", "0.1901" ], [ "0.8022", "0.1724", "0.824", "0.1724", "0.824", "0.1875", "0.8022", "0.1875" ], [ "0.5923", "0.1831", "0.6167", "0.1831", "0.6167", "0.1963", "0.5923", "0.1963" ], [ "0.1724", "0.1995", "0.8267", "0.1989", "0.8267", "0.2166", "0.1724", "0.2172" ], [ "0.1765", "0.2172", "0.8228", "0.2172", "0.8228", "0.2317", "0.1765", "0.2317" ], [ "0.1765", "0.2343", "0.822", "0.2343", "0.822", "0.2443", "0.1765", "0.2443" ], [ "0.1748", "0.2456", "0.824", "0.2456", "0.824", "0.2588", "0.1748", "0.2588" ], [ "0.1748", "0.26", "0.824", "0.26", "0.824", "0.2734", "0.1748", "0.2734" ], [ "0.1748", "0.2734", "0.4004", "0.2734", "0.4004", "0.2874", "0.1748", "0.2874" ], [ "0.1732", "0.2922", "0.824", "0.2935", "0.824", "0.308", "0.1732", "0.307" ], [ "0.1732", "0.31", "0.824", "0.31", "0.824", "0.3245", "0.1732", "0.3245" ], [ "0.1732", "0.327", "0.824", "0.327", "0.824", "0.3403", "0.1732", "0.3403" ], [ "0.1748", "0.3416", "0.6006", "0.3416", "0.6006", "0.356", "0.1748", "0.356" ], [ "0.1748", "0.361", "0.826", "0.361", "0.826", "0.3757", "0.1748", "0.3757" ], [ "0.1748", "0.3757", "0.8228", "0.3757", "0.8228", "0.3894", "0.1748", "0.3894" ], [ "0.1748", "0.3894", "0.824", "0.3894", "0.824", "0.4028", "0.1748", "0.4028" ], [ "0.1732", "0.404", "0.6406", "0.4028", "0.6406", "0.4167", "0.1732", "0.4172" ], [ "0.1732", "0.4192", "0.777", "0.4192", "0.777", "0.4343", "0.1732", "0.4343" ], [ "0.1732", "0.443", "0.824", "0.4443", "0.824", "0.459", "0.1732", "0.4583" ], [ "0.1748", "0.459", "0.3513", "0.459", "0.3513", "0.473", "0.1748", "0.473" ], [ "0.1748", "0.4792", "0.8228", "0.4792", "0.8228", "0.4937", "0.1748", "0.4937" ], [ "0.1732", "0.4937", "0.826", "0.4937", "0.826", "0.507", "0.1732", "0.507" ], [ "0.1732", "0.509", "0.6587", "0.509", "0.6587", "0.5234", "0.1732", "0.5234" ], [ "0.6543", "0.512", "0.6626", "0.512", "0.6626", "0.515", "0.6543", "0.515" ], [ "0.66", "0.507", "0.8228", "0.509", "0.8228", "0.5234", "0.66", "0.5215" ], [ "0.1748", "0.5234", "0.824", "0.5234", "0.824", "0.537", "0.1748", "0.537" ], [ "0.1732", "0.5386", "0.826", "0.5386", "0.826", "0.552", "0.1732", "0.552" ], [ "0.1748", "0.552", "0.7534", "0.552", "0.7534", "0.565", "0.1748", "0.565" ], [ "0.1765", "0.5713", "0.824", "0.5713", "0.824", "0.586", "0.1765", "0.586" ], [ "0.1765", "0.5884", "0.4192", "0.5884", "0.4192", "0.5986", "0.1765", "0.5986" ], [ "0.4167", "0.5913", "0.423", "0.5913", "0.423", "0.5947", "0.4167", "0.5947" ], [ "0.4207", "0.5884", "0.822", "0.5884", "0.822", "0.5986", "0.4207", "0.5986" ], [ "0.1748", "0.601", "0.8228", "0.601", "0.8228", "0.6143", "0.1748", "0.6143" ], [ "0.1732", "0.6123", "0.822", "0.6123", "0.822", "0.6255", "0.1732", "0.6255" ], [ "0.1765", "0.629", "0.822", "0.629", "0.822", "0.6396", "0.1765", "0.6396" ], [ "0.1748", "0.64", "0.8228", "0.64", "0.8228", "0.654", "0.1748", "0.654" ], [ "0.1748", "0.656", "0.8228", "0.656", "0.8228", "0.6665", "0.1748", "0.6665" ], [ "0.1748", "0.6685", "0.824", "0.6685", "0.824", "0.6816", "0.1748", "0.6816" ], [ "0.1748", "0.683", "0.451", "0.683", "0.451", "0.6963", "0.1748", "0.6963" ], [ "0.1732", "0.7007", "0.824", "0.702", "0.824", "0.7163", "0.1732", "0.7153" ], [ "0.1732", "0.7173", "0.4387", "0.7173", "0.4387", "0.731", "0.1732", "0.731" ], [ "0.1732", "0.747", "0.4878", "0.747", "0.4878", "0.761", "0.1732", "0.761" ], [ "0.1732", "0.773", "0.824", "0.773", "0.824", "0.786", "0.1732", "0.786" ], [ "0.1748", "0.786", "0.826", "0.787", "0.826", "0.8003", "0.1748", "0.8" ], [ "0.1748", "0.8003", "0.824", "0.8003", "0.824", "0.8145", "0.1748", "0.8145" ], [ "0.1748", "0.815", "0.822", "0.8164", "0.822", "0.8276", "0.1748", "0.827" ], [ "0.1765", "0.83", "0.8228", "0.83", "0.8228", "0.8403", "0.1765", "0.8403" ], [ "0.1748", "0.8423", "0.826", "0.8423", "0.826", "0.856", "0.1748", "0.856" ], [ "0.1748", "0.856", "0.826", "0.856", "0.826", "0.869", "0.1748", "0.869" ], [ "0.1765", "0.872", "0.822", "0.872", "0.822", "0.882", "0.1765", "0.882" ], [ "0.1732", "0.882", "0.8228", "0.8833", "0.8228", "0.898", "0.1732", "0.8965" ], [ "0.1748", "0.898", "0.826", "0.898", "0.826", "0.911", "0.1748", "0.911" ], [ "0.1748", "0.911", "0.826", "0.911", "0.826", "0.925", "0.1748", "0.925" ], [ "0.4927", "0.9507", "0.5063", "0.9507", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2023", "3.1 TF-ATTENTION (Tree Fine-grained Attention)", "In this variant, we compute the value for a query q as a weighted sum of value vectors of keys that lie", "in the same leaf node as q. More formally,", "TF-ATTENTION(q, K,V;T) = softma)", "qTWQ(K,WK)T", ".Vswv", "(3)", "Vd", "where S = Sh,Pr(h.(WQ)Tg)(KWK) is the set of keys K that are mapped to same leaf node (i.e.", "Pr(h, (WQ)T g)) as (WQ)T q. Note that this is a sparse attention variant as we compute dot products", "onlybetweenthosequeriesandkeysbelongingtothesameleafnode,andtheattentionscorebetween", "a query and keys lying in different leaf nodes is O. Intuitively, this attention layer can be thought of as", "top-k attention where the tree structure is responsible for computing the k tokens which are the most", "important for a given query token.", "Computational Complexity Assuming a uniform distribution of the input sequence tokens across", "the leaf nodes of the decision tree, the amortized computational cost for TF-ATTENTION is O(d) +", "2cd2n = O(nkd) + 2cd?n, where k is generally a small factor assuming the height of the decision", "tree is of the order O(log()), and c is as defined in Section 2.1.", "If the decision tree parameters are such that all the queries and keys get clustered into a single leaf", "node of the tree, we get the original full attention in this case. Therefore the worst case time complexity", "of TF-ATTENTION is O(n?d + d?2n). Furthermore, this also illustrates that TF-ATTENTION can be", "as expressive as standard attention.We state this claim formally below.", "Lemma 1. There exists a tree T s.t. TF-ATTENTION(Q, K, V; T) = Attention(Q, K, V).", "Thelemma follows immediately from the above observation; for completeness, see Appendix for a", "proof of the above lemma.", "k-ary TF-ATTENTION We extend the binary tree construction of the TF-A variant to construct", "a k-ary tree attention model for height h, where each node at level l has a branching factor of bi,", "I e [0, 1, . . . , h - 1]. The total number of leaf nodes in this case are '", "-", "h-1 br.This is a general", "version of the TF-ATTENTION variant, where we can scale the number of ieaf nodes even for smaller", "heights, and thus reducing the amount of queries and keys in each leaf node. In all our experiments", "on k-ary TF-ATTENTION models, h = 2 and b = [100, 10], resulting in 1000 leaf nodes.", "Although the proposed TF-ATTENTION TREEFORMER variant works well in principle, there are a", "coupleofshortcomingsofthesame-", "", "-(i)Incasethedistributionofqueriesandkeysisskewedin", "even one of the layers of the model, the computational cost approaches the worst case time complexity", "of the standard attention model. So we might need to add explicit loss terms to encourage more", "uniformrepresentationacrossvariousleafnodes/subtrees.(ii)Itmightbepossiblethatforagiven", "query, some key tokens important to this query lie in a different leaf node of the decision tree to", "minimizethe overall loss.But due tothe sparse nature of the model, theattention score between the", "query and these keys will be O. Using multiple decision trees (forest structure) might help alleviate", "this but we leave it for future exploration.", "We address the above problems using a novel variant that computes the attention values in a coarse", "manner and hence has denser gradients.", "3.2 TC-ATTENTION (Tree Coarse Attention)", "In this variant, rather than computing attention as a weighted sum of values of keys in the leaf node.", "we take a coarse unweighted sum of the value vectors. That is, the query is only used to navigate to a", "leaf node, but with-in a leaf node, the value vector is fixed. So given a leaf node, the attention value is", "independentofthequeryitself.Notethatthisrequirementcompletelyflipsthetree'stendencytomap", "multiplekeystoasingleleafnode.Thatis,tooptimizethelossfunction.TF-ATTENTIoNwould", "prefer all the keys to map to one leaf node, TC-ATTENTION's incentives are completely reversed.", "That is, if TC-ATTENTION maps all the keys to the same leaf node then it would give an unweighted", "averageofthevaluevectorsofall thekeys,irrespectiveofquerywhichisaverageattention-naturally", "a weaker model and would have large loss. Instead, TC-ATTENTION would prefer to spread out keys", "as much as possible in leaf nodes so that it can compute more powerful query dependent attention.", "We present key distributions of learned TREEFORMER models in Figure 4 supporting this argument.", "4" ]

{ "arxiv_id": "2208.09015", "language": "en", "page_idx": 3 }

[ [ "0.1748", "0.03284", "0.4795", "0.03284", "0.4795", "0.04736", "0.1748", "0.04736" ], [ "0.652", "0.1477", "0.74", "0.1459", "0.74", "0.1559", "0.652", "0.1578" ], [ "0.2793", "0.1895", "0.2942", "0.1895", "0.2942", "0.1995", "0.2793", "0.1995" ], [ "0.2573", "0.2052", "0.273", "0.2052", "0.273", "0.2134", "0.2573", "0.2134" ], [ "0.3022", "0.207", "0.3105", "0.207", "0.3105", "0.2122", "0.3022", "0.2122" ], [ "0.4453", "0.2108", "0.4558", "0.2108", "0.4558", "0.214", "0.4453", "0.214" ], [ "0.728", "0.2084", "0.8203", "0.2058", "0.8203", "0.2166", "0.728", "0.2184" ], [ "0.25", "0.2203", "0.3228", "0.2203", "0.3228", "0.2317", "0.25", "0.2317" ], [ "0.2695", "0.238", "0.282", "0.238", "0.282", "0.2412", "0.2695", "0.2412" ], [ "0.246", "0.2551", "0.3179", "0.2551", "0.3179", "0.2827", "0.246", "0.2827" ], [ "0.3325", "0.2568", "0.3782", "0.2568", "0.3782", "0.2822", "0.3325", "0.2822" ], [ "0.546", "0.2917", "0.6006", "0.2917", "0.6006", "0.3018", "0.546", "0.3018" ], [ "0.1936", "0.26", "0.2378", "0.2568", "0.2402", "0.2766", "0.1953", "0.279" ], [ "0.6235", "0.2922", "0.669", "0.2922", "0.669", "0.2998", "0.6235", "0.2998" ], [ "0.6978", "0.2917", "0.761", "0.289", "0.761", "0.3005", "0.6978", "0.303" ], [ "0.597", "0.3145", "0.648", "0.3145", "0.648", "0.3213", "0.597", "0.3213" ], [ "0.4805", "0.2935", "0.5063", "0.2935", "0.5063", "0.3005", "0.4805", "0.3005" ], [ "0.1748", "0.334", "0.8228", "0.334", "0.8228", "0.348", "0.1748", "0.348" ], [ "0.1748", "0.3484", "0.814", "0.3484", "0.814", "0.3623", "0.1748", "0.3623" ], [ "0.1748", "0.387", "0.826", "0.387", "0.826", "0.401", "0.1748", "0.401" ], [ "0.1748", "0.4016", "0.824", "0.4016", "0.824", "0.4153", "0.1748", "0.4153" ], [ "0.1732", "0.4153", "0.4844", "0.4153", "0.4844", "0.4287", "0.1732", "0.4287" ], [ "0.1724", "0.4343", "0.6094", "0.435", "0.6094", "0.4495", "0.1724", "0.449" ], [ "0.451", "0.4602", "0.4558", "0.4602", "0.4558", "0.4673", "0.451", "0.4673" ], [ "0.5317", "0.4634", "0.5645", "0.4634", "0.5645", "0.483", "0.5317", "0.483" ], [ "0.357", "0.4697", "0.3938", "0.4697", "0.3938", "0.481", "0.357", "0.481" ], [ "0.5864", "0.464", "0.6357", "0.4622", "0.636", "0.4792", "0.5884", "0.481" ], [ "0.4028", "0.4734", "0.5", "0.4717", "0.5", "0.4893", "0.4028", "0.4912" ], [ "0.5063", "0.4841", "0.5884", "0.4841", "0.5884", "0.4956", "0.5063", "0.4956" ], [ "0.1724", "0.5005", "0.824", "0.5024", "0.824", "0.521", "0.1724", "0.518" ], [ "0.1732", "0.5176", "0.7754", "0.518", "0.7754", "0.5327", "0.1732", "0.532" ], [ "0.527", "0.547", "0.546", "0.547", "0.546", "0.56", "0.527", "0.56" ], [ "0.304", "0.561", "0.5312", "0.561", "0.5312", "0.5757", "0.304", "0.5757" ], [ "0.5186", "0.5596", "0.697", "0.563", "0.697", "0.5806", "0.5186", "0.577" ], [ "0.8022", "0.561", "0.826", "0.561", "0.826", "0.5757", "0.8022", "0.5757" ], [ "0.524", "0.5776", "0.5483", "0.5776", "0.5483", "0.5903", "0.524", "0.5903" ], [ "0.1748", "0.595", "0.824", "0.595", "0.824", "0.61", "0.1748", "0.61" ], [ "0.1732", "0.61", "0.4043", "0.6094", "0.4043", "0.6235", "0.1732", "0.625" ], [ "0.1724", "0.63", "0.824", "0.63", "0.824", "0.643", "0.1724", "0.643" ], [ "0.1732", "0.646", "0.826", "0.6445", "0.826", "0.6577", "0.1732", "0.659" ], [ "0.1748", "0.659", "0.826", "0.659", "0.826", "0.6724", "0.1748", "0.6724" ], [ "0.1732", "0.6724", "0.826", "0.672", "0.826", "0.686", "0.1732", "0.6875" ], [ "0.1748", "0.686", "0.824", "0.686", "0.824", "0.6997", "0.1748", "0.6997" ], [ "0.1732", "0.7007", "0.58", "0.7007", "0.58", "0.714", "0.1732", "0.714" ], [ "0.1732", "0.7197", "0.826", "0.72", "0.826", "0.735", "0.1732", "0.7344" ], [ "0.1748", "0.735", "0.824", "0.735", "0.824", "0.7485", "0.1748", "0.7485" ], [ "0.1748", "0.7485", "0.8228", "0.7485", "0.8228", "0.762", "0.1748", "0.762" ], [ "0.1732", "0.7603", "0.8267", "0.7603", "0.8267", "0.778", "0.1732", "0.778" ], [ "0.1732", "0.7744", "0.7524", "0.7744", "0.7524", "0.789", "0.1732", "0.789" ], [ "0.1748", "0.797", "0.8228", "0.797", "0.8228", "0.811", "0.1748", "0.811" ], [ "0.1732", "0.811", "0.824", "0.811", "0.824", "0.8257", "0.1732", "0.8257" ], [ "0.1748", "0.8257", "0.8228", "0.8257", "0.8228", "0.8403", "0.1748", "0.8403" ], [ "0.1724", "0.839", "0.8228", "0.8384", "0.8228", "0.853", "0.1724", "0.854" ], [ "0.1748", "0.8706", "0.4739", "0.8706", "0.4739", "0.884", "0.1748", "0.884" ], [ "0.1748", "0.8965", "0.826", "0.8965", "0.826", "0.911", "0.1748", "0.911" ], [ "0.1732", "0.911", "0.824", "0.911", "0.824", "0.925", "0.1732", "0.925" ], [ "0.4944", "0.9487", "0.5063", "0.9487", "0.5063", "0.961", "0.4944", "0.961" ] ]

[ "Published as a conference paper at ICLR 2023", "(0+++)", "O", "O", "", "e", "(etatem+0)w", "DO", "e", "", "", "(en+e)w", "Lne", "(eg+o)w", "+", "TC-A(g.K,V:7", "(oe)W", "Figure 1: The figure on the left represents the self attention layer in TREEFORMER models, and the", "tree self attention block is expanded for a representative view of the TC-ATTENTION mechanism.", "We further generalize this technique to ensure contributions from other nodes to the value vector.", "That is, we add contributions to a query's value vector from not only the leaf node but also from the", "internal nodes that were traversed by the query.", "Formally, let V, be the value vector associated with a node.Then,", "-", "Z", "Vi.jE", "VwV", "ISt.,(KWK)", "iEStj(KWK)", "ie., it is an average of all value vectors of keys such that l, j lies in their paths. Now, given v.,j for", "each node and the tree T (including it's parameters 0), we get the attention for a query q as:", "h", "TC-ATTENTION(q, K, V; T) =", "al*V(1,Pr(1,(WQ)Tq)", "(4)", "1=0", "where ar is the learnable attention weight at each level of the decision tree. See Figure 1 for a visual", "representation of TC-ATTENTION.", " Note that as TC-ATTENTION applies query independent averaging, it is easy to design counterex-", "amples where the standard attention layer is strictly more powerful than TC-ATTENTION. Now,", "we can make TC-ATTENTION as powerful as a standard attention layer by allowing traversal of", "multiple paths and Qu to be query dependent. However, our results indicate that TC-ATTENTION is", "sufficiently powerful, and might not require more modeling power in the attention layer. We leave", "further exploration of multiple traversal paths for future work.", "Computational Complexity The cost of TC-ATTENTION can be broken down into three parts:", "1. projection of query, key, and value tensors - note that we can embed projection matrices for both", "query and key in the parameters of the tree (with two different sets of tree parameters). So the cost is", "half of standard attention: cd?n 2. storing projected key vectors in the decision tree - O(2h+1 - 1)d),", "and 3. computing the path and attention layer transformation for all queries Q - O(ndh)", "Compared to the full attention mechanism - O(n? d + d? n), the computational complexity for TC", "ATTENTION is linear in sequence length n: O(ndh + (2h+1 - 1)d + d?n). In all our experiments,", "decision tree height h is generally ≤ 10. Also, note that compared to TF-ATTENTION, the computa", "tional complexity is guaranteed to be linear in both n and h, irrespective of the trained trees/model.", "3.3TRAININGOFTREEFORMERMODELS", "Pre-training It is well-known that learning tree is a challenging problem due to the discrete structure.", "We use recent techniques (Karthikeyan et al., 2021) to help with the optimization, and ensure", "5" ]

{ "arxiv_id": "2208.09015", "language": "en", "page_idx": 4 }

[ [ "0.1136", "0.1332", "0.8594", "0.1332", "0.8594", "0.1509", "0.1136", "0.1509" ], [ "0.2427", "0.1572", "0.7295", "0.1572", "0.7295", "0.1718", "0.2427", "0.1718" ], [ "0.1765", "0.2026", "0.8", "0.202", "0.8", "0.2197", "0.1765", "0.2203" ], [ "0.1912", "0.2229", "0.558", "0.2229", "0.558", "0.2374", "0.1912", "0.2374" ], [ "0.5664", "0.2229", "0.7812", "0.2229", "0.7812", "0.2374", "0.5664", "0.2374" ], [ "0.558", "0.2291", "0.5703", "0.2291", "0.5703", "0.2329", "0.558", "0.2329" ], [ "0.3635", "0.2424", "0.601", "0.2424", "0.601", "0.2568", "0.3635", "0.2568" ], [ "0.01878", "0.254", "0.0645", "0.254", "0.0678", "0.3757", "0.02124", "0.3757" ], [ "0.2632", "0.26", "0.71", "0.26", "0.71", "0.2734", "0.2632", "0.2734" ], [ "0.2361", "0.308", "0.3145", "0.308", "0.3145", "0.3257", "0.2361", "0.3257" ], [ "0.5254", "0.315", "0.584", "0.315", "0.584", "0.327", "0.5254", "0.327" ], [ "0.642", "0.315", "0.6567", "0.315", "0.6567", "0.318", "0.642", "0.318" ], [ "0.697", "0.3164", "0.7466", "0.3164", "0.7466", "0.3276", "0.697", "0.3276" ], [ "0.8325", "0.3113", "0.8735", "0.3113", "0.8735", "0.3213", "0.8325", "0.3213" ], [ "0.5063", "0.3193", "0.5156", "0.3193", "0.5156", "0.3245", "0.5063", "0.3245" ], [ "0.607", "0.3208", "0.6323", "0.3208", "0.6323", "0.3289", "0.607", "0.3289" ], [ "0.8", "0.3176", "0.824", "0.3176", "0.824", "0.3289", "0.8", "0.3289" ], [ "0.8364", "0.3208", "0.864", "0.3208", "0.864", "0.332", "0.8364", "0.332" ], [ "0.5537", "0.3308", "0.6045", "0.327", "0.6055", "0.3372", "0.5557", "0.3416" ], [ "0.774", "0.334", "0.7974", "0.334", "0.7974", "0.3423", "0.774", "0.3423" ], [ "0.834", "0.3308", "0.859", "0.3308", "0.859", "0.3423", "0.834", "0.3423" ], [ "0.0989", "0.3403", "0.4697", "0.3416", "0.4697", "0.356", "0.0989", "0.3547" ], [ "0.5034", "0.3352", "0.5186", "0.3352", "0.5186", "0.3452", "0.5034", "0.3452" ], [ "0.7466", "0.3416", "0.773", "0.3416", "0.773", "0.3499", "0.7466", "0.3499" ], [ "0.839", "0.3416", "0.888", "0.3416", "0.888", "0.3528", "0.839", "0.3528" ], [ "0.5024", "0.3516", "0.523", "0.3516", "0.523", "0.3655", "0.5024", "0.3655" ], [ "0.7183", "0.356", "0.744", "0.356", "0.744", "0.3674", "0.7183", "0.3674" ], [ "0.8364", "0.3528", "0.8667", "0.3528", "0.8667", "0.3643", "0.8364", "0.3643" ], [ "0.835", "0.361", "0.888", "0.3591", "0.8896", "0.3738", "0.8364", "0.3757" ], [ "0.0809", "0.358", "0.4697", "0.358", "0.4697", "0.3718", "0.0809", "0.3718" ], [ "0.02696", "0.363", "0.05554", "0.363", "0.05554", "0.3926", "0.02696", "0.3926" ], [ "0.0801", "0.3726", "0.468", "0.3726", "0.468", "0.3865", "0.0801", "0.3865" ], [ "0.5024", "0.3694", "0.5215", "0.3694", "0.5215", "0.3818", "0.5024", "0.3818" ], [ "0.693", "0.3726", "0.7134", "0.3726", "0.7134", "0.3801", "0.693", "0.3801" ], [ "0.8364", "0.375", "0.875", "0.375", "0.875", "0.385", "0.8364", "0.385" ], [ "0.0809", "0.3882", "0.4697", "0.3882", "0.4697", "0.4028", "0.0809", "0.4028" ], [ "0.8384", "0.3833", "0.8906", "0.3833", "0.8906", "0.3945", "0.8384", "0.3945" ], [ "0.505", "0.3901", "0.5195", "0.3901", "0.5195", "0.3984", "0.505", "0.3984" ], [ "0.5186", "0.3894", "0.5483", "0.3926", "0.5474", "0.3997", "0.517", "0.3965" ], [ "0.0784", "0.4016", "0.468", "0.4028", "0.468", "0.4172", "0.0784", "0.4167" ], [ "0.835", "0.3958", "0.8784", "0.3958", "0.8784", "0.4072", "0.835", "0.4072" ], [ "0.02533", "0.406", "0.05554", "0.406", "0.05554", "0.4507", "0.02533", "0.4507" ], [ "0.5063", "0.4104", "0.5156", "0.4104", "0.5156", "0.4153", "0.5063", "0.4153" ], [ "0.5522", "0.4136", "0.6323", "0.4136", "0.6323", "0.4248", "0.5522", "0.4248" ], [ "0.839", "0.4077", "0.869", "0.4077", "0.869", "0.4167", "0.839", "0.4167" ], [ "0.0801", "0.4185", "0.4697", "0.4185", "0.4697", "0.432", "0.0801", "0.432" ], [ "0.71", "0.414", "0.812", "0.414", "0.812", "0.4248", "0.71", "0.4248" ], [ "0.5024", "0.4268", "0.8906", "0.4268", "0.8906", "0.44", "0.5024", "0.44" ], [ "0.04166", "0.4363", "0.04984", "0.4363", "0.04984", "0.4414", "0.04166", "0.4414" ], [ "0.0784", "0.4343", "0.4673", "0.4343", "0.4673", "0.4475", "0.0784", "0.4475" ], [ "0.5034", "0.4426", "0.864", "0.4426", "0.864", "0.4526", "0.5034", "0.4526" ], [ "0.03186", "0.449", "0.058", "0.449", "0.058", "0.481", "0.03186", "0.481" ], [ "0.0801", "0.449", "0.4697", "0.449", "0.4697", "0.4622", "0.0801", "0.4622" ], [ "0.0801", "0.4634", "0.4673", "0.4634", "0.4673", "0.4766", "0.0801", "0.4766" ], [ "0.5034", "0.461", "0.8906", "0.461", "0.8906", "0.4749", "0.5034", "0.4749" ], [ "0.0801", "0.4778", "0.4673", "0.4778", "0.4673", "0.4912", "0.0801", "0.4912" ], [ "0.501", "0.4766", "0.8906", "0.4778", "0.8906", "0.4912", "0.501", "0.4905" ], [ "0.0801", "0.4937", "0.468", "0.4937", "0.468", "0.507", "0.0801", "0.507" ], [ "0.5024", "0.4924", "0.8906", "0.4924", "0.8906", "0.506", "0.5024", "0.506" ], [ "0.02533", "0.502", "0.05554", "0.502", "0.05554", "0.5547", "0.02533", "0.5547" ], [ "0.0801", "0.509", "0.468", "0.509", "0.468", "0.523", "0.0801", "0.523" ], [ "0.5024", "0.507", "0.892", "0.507", "0.892", "0.521", "0.5024", "0.521" ], [ "0.0801", "0.5234", "0.468", "0.5234", "0.468", "0.537", "0.0801", "0.537" ], [ "0.501", "0.523", "0.892", "0.523", "0.892", "0.536", "0.501", "0.536" ], [ "0.0801", "0.539", "0.4697", "0.539", "0.4697", "0.553", "0.0801", "0.553" ], [ "0.501", "0.536", "0.892", "0.537", "0.892", "0.552", "0.501", "0.5503" ], [ "0.0784", "0.5537", "0.4697", "0.5537", "0.4697", "0.5684", "0.0784", "0.5684" ], [ "0.5024", "0.553", "0.8906", "0.553", "0.8906", "0.5664", "0.5024", "0.5664" ], [ "0.5024", "0.5684", "0.8906", "0.5684", "0.8906", "0.582", "0.5024", "0.582" ], [ "0.02696", "0.556", "0.05392", "0.556", "0.05392", "0.587", "0.02696", "0.587" ], [ "0.0801", "0.5693", "0.468", "0.5693", "0.468", "0.5835", "0.0801", "0.5835" ], [ "0.0784", "0.5835", "0.4673", "0.5835", "0.4673", "0.5977", "0.0784", "0.5977" ], [ "0.5024", "0.5835", "0.892", "0.5835", "0.892", "0.5967", "0.5024", "0.5967" ], [ "0.04004", "0.5884", "0.04822", "0.5884", "0.04822", "0.5913", "0.04004", "0.5913" ], [ "0.02777", "0.5884", "0.05392", "0.5884", "0.05392", "0.6206", "0.02777", "0.6206" ], [ "0.0801", "0.5996", "0.4697", "0.5996", "0.4697", "0.613", "0.0801", "0.613" ], [ "0.501", "0.5977", "0.8906", "0.5977", "0.8906", "0.6113", "0.501", "0.6113" ], [ "0.04166", "0.605", "0.0474", "0.605", "0.0474", "0.6113", "0.04166", "0.6113" ], [ "0.04166", "0.6113", "0.04822", "0.6113", "0.04822", "0.6157", "0.04166", "0.6157" ], [ "0.501", "0.6123", "0.8906", "0.613", "0.8906", "0.6274", "0.501", "0.627" ], [ "0.0809", "0.616", "0.4673", "0.616", "0.4673", "0.627", "0.0809", "0.627" ], [ "0.02777", "0.6143", "0.05392", "0.6143", "0.05392", "0.632", "0.02777", "0.632" ], [ "0.0784", "0.629", "0.4697", "0.63", "0.4697", "0.643", "0.0784", "0.6426" ], [ "0.5024", "0.6274", "0.876", "0.6274", "0.876", "0.6416", "0.5024", "0.6416" ], [ "0.0335", "0.64", "0.05392", "0.64", "0.05392", "0.672", "0.0335", "0.672" ], [ "0.02942", "0.63", "0.05228", "0.63", "0.05228", "0.643", "0.02942", "0.643" ], [ "0.0809", "0.6445", "0.468", "0.6445", "0.468", "0.6577", "0.0809", "0.6577" ], [ "0.523", "0.6445", "0.892", "0.6445", "0.892", "0.659", "0.523", "0.659" ], [ "0.0801", "0.66", "0.4673", "0.66", "0.4673", "0.674", "0.0801", "0.674" ], [ "0.5024", "0.66", "0.8906", "0.66", "0.8906", "0.674", "0.5024", "0.674" ], [ "0.04166", "0.6694", "0.04984", "0.6694", "0.04984", "0.674", "0.04166", "0.674" ], [ "0.04166", "0.674", "0.04984", "0.674", "0.04984", "0.678", "0.04166", "0.678" ], [ "0.0801", "0.675", "0.468", "0.675", "0.468", "0.688", "0.0801", "0.688" ], [ "0.5034", "0.675", "0.892", "0.674", "0.892", "0.688", "0.5034", "0.6895" ], [ "0.0784", "0.688", "0.468", "0.6895", "0.468", "0.704", "0.0784", "0.7026" ], [ "0.03186", "0.6787", "0.05228", "0.678", "0.05554", "0.7056", "0.03513", "0.707" ], [ "0.5024", "0.69", "0.8906", "0.69", "0.8906", "0.704", "0.5024", "0.704" ], [ "0.0809", "0.705", "0.2255", "0.705", "0.2255", "0.7183", "0.0809", "0.7183" ], [ "0.5034", "0.7056", "0.8906", "0.7056", "0.8906", "0.7197", "0.5034", "0.7197" ], [ "0.5034", "0.72", "0.892", "0.72", "0.892", "0.7344", "0.5034", "0.7344" ], [ "0.5024", "0.735", "0.8896", "0.7344", "0.8896", "0.7485", "0.5024", "0.75" ], [ "0.0809", "0.7505", "0.2075", "0.7505", "0.2075", "0.7646", "0.0809", "0.7646" ], [ "0.5034", "0.7524", "0.8906", "0.7524", "0.8906", "0.763", "0.5034", "0.763" ], [ "0.5024", "0.7666", "0.892", "0.7666", "0.892", "0.78", "0.5024", "0.78" ], [ "0.10046", "0.7764", "0.468", "0.7764", "0.468", "0.7905", "0.10046", "0.7905" ], [ "0.5034", "0.7827", "0.8896", "0.7827", "0.8896", "0.7935", "0.5034", "0.7935" ], [ "0.0784", "0.7905", "0.468", "0.7915", "0.468", "0.806", "0.0784", "0.8047" ], [ "0.5024", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5024", "0.81" ], [ "0.0801", "0.8066", "0.468", "0.8066", "0.468", "0.821", "0.0801", "0.821" ], [ "0.505", "0.8125", "0.8896", "0.8125", "0.8896", "0.8228", "0.505", "0.8228" ], [ "0.0801", "0.8228", "0.468", "0.8228", "0.468", "0.836", "0.0801", "0.836" ], [ "0.5024", "0.8257", "0.892", "0.8257", "0.892", "0.8403", "0.5024", "0.8403" ], [ "0.0784", "0.837", "0.468", "0.836", "0.468", "0.851", "0.0784", "0.8516" ], [ "0.5024", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.5024", "0.855" ], [ "0.0972", "0.863", "0.468", "0.863", "0.468", "0.877", "0.0972", "0.877" ], [ "0.5034", "0.858", "0.8896", "0.858", "0.8896", "0.8687", "0.5034", "0.8687" ], [ "0.501", "0.8706", "0.8906", "0.872", "0.8906", "0.886", "0.501", "0.8853" ], [ "0.0825", "0.8774", "0.10455", "0.8774", "0.10455", "0.886", "0.0825", "0.886" ], [ "0.0972", "0.887", "0.2255", "0.887", "0.2255", "0.901", "0.0972", "0.901" ], [ "0.5024", "0.887", "0.8716", "0.887", "0.8716", "0.901", "0.5024", "0.901" ] ]

[ "GPA-3D: Geometry-aware Prototype Alignment for Unsupervised Domain", "Adaptive3DObjectDetectionfromPointClouds", "Ziyu Lil* Jingming Guo2Tongtong Cao2 Liu Bingbing?Wankou Yanglt", "1 School of Automation, Southeast University", "2 Huawei Noah's Ark Lab", "", "(iiziyu, wkyang)@seu.edu.cn", " 3", "(guojingming,caotongtong,liu.bingbing)@huawei.com", "Abstract", "3D AP (%)", "AO", "3D AP (%) ", "Source", "70", "65.64", "22.54", "Only", "59.261.83", "20.9", "SN", "LiDAR-based 3D detection has made great progress in", "60", "20.19", "[CVPR'20]", "50", "18.57", "ST3D", "[CVPR*21]", "recent years. However, the performance of 3D detectors", "6", "is considerably limited when deployed in unseen environ-", "40", "17.24", "ST3D++", "ments, owing to the severe domain gap problem. Existing", "[TPAMI22]", "30", "127.48", "domain adaptive 3D detection methods do not adequately", "GPA-3D", "LAD", "20", "Waymo-to-KITTI", "(Ours)", "consider the problem of the distributional discrepancy in", "Waymo-to-nuScenes", "Figure 1. The performance comparison with previous works [34,", "中", "feature space, thereby hindering generalization of detec-", "37,381.ThedetectionarchitectureisSECOND-loU[35,37]", "SO", "tors across domains.In this work, we propose a novel", "unsupervised domain adaptive 3D detection framework,", "suffer from dramatic performance degradation when train-", "namely Geometry-aware Prototype Alignment (GPA-3D),", "ing data and test data are from different environments, i.e.,", "which explicitly leverages the intrinsic geometric relation-", "domain shift problem [34].Various factors, such as di-", "", "ship from point cloud objects to reduce the feature discrep-", "verse weather conditions, object sizes, laser beams, and", "ancy, thus facilitating cross-domain transferring. Specifi-", "scanning patterns, lead to substantial discrepancies across", "cally, GPA-3D assigns a series of tailored and learnable", "different domains, hindering the transferability of existing", "prototypes to point cloud objects with distinct geometric", "LiDAR-based 3D detectors. Intuitively, fine-tuning the de-", "tectors with adequate data from the target domain could al-", "8", "structures. Each prototype aligns BEV (bird's-eye-view)", "features derived from corresponding point cloud objects", "leviate this issue. However, manually annotating a large", "", "80", "on source and target domains, reducing the distributional", "amount of point cloud scenes is a prohibitively expensive", "", "", "task. Therefore, the research on unsupervised domain adap-", "discrepancy and achieving better adaptation.The evalu-", "3", "ation results obtained on various benchmarks, including", "tation (UDA) for LiDAR-based 3D detection is essential.", ".AI", "2", "Waymo, nuScenes and KITTI, demonstrate the superiority", "Although many works have been proposed to deal with", "of our GPA-3D over the state-of-the-art approaches for dif-", "the UDA for image-based detection [43, 11, 25, 17, 10,", "-", "", "ferent adaptation scenarios. The MindSpore version code", "14, 13, 26, 3], directly applying these methods to 3D point", "will be publicly available at https://github.com/", "e", "cloud detection is insufficient for tackling the domain shifts.", "Liz66666/GPA3D.", "These approaches mainly concentrate on the gaps of light-", "ing and texture variations, which could not be obtained from", "point clouds. While there is only a limited number of liter-", "1. Introduction", "ature[34,37,27.9,40,38,201dealingwiththeUDAon", "LiDAR-based 3D detection. Prior work [34] utilizes the", "As a fundamental research in 3D scene understanding,", "statisticsfromtarget annotation to perform data-level nor-", "3D detection from point clouds has attracted increasing at-", "malization. MLC-Net [20] designs a mean-teacher frame-", "tention due to its essential role in intelligent robotics, aug-", "worktoprovidereliablepseudo-labelstofacilitatetransfer-", "mented reality and autonomous driving [7, 22, 18, 8, 1]. De-", "ring. ST3D [37] and ST3D++ [38] propose a self-training", "spite significant process, state-of-the-art 3D detectors still", "pipeline with a memory bank to collect and refine pseudo-", "Work was jointly done when Ziyu was an intern at Huawei Noah's Ark", "labels.Despitetheirgreat success,thesemethodsdo notad-", "equately consider the problem of distributional discrepancy", "Lab", "*Corresponding Author", "in feature space, hampering the adaptation performance." ]

{ "arxiv_id": "2308.08140", "language": "en", "page_idx": 0 }

[ [ "0.1993", "0.09155", "0.2566", "0.09155", "0.2566", "0.0985", "0.1993", "0.0985" ], [ "0.2957", "0.09155", "0.3635", "0.09155", "0.3635", "0.0985", "0.2957", "0.0985" ], [ "0.3938", "0.0903", "0.4478", "0.0903", "0.4478", "0.101", "0.3938", "0.101" ], [ "0.5024", "0.09216", "0.8906", "0.09216", "0.8906", "0.1061", "0.5024", "0.1061" ], [ "0.5024", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.5024", "0.1218" ], [ "0.125", "0.09344", "0.152", "0.09344", "0.152", "0.09784", "0.125", "0.09784" ], [ "0.501", "0.1225", "0.892", "0.1237", "0.892", "0.137", "0.501", "0.1364" ], [ "0.5024", "0.1383", "0.796", "0.1383", "0.796", "0.1515", "0.5024", "0.1515" ], [ "0.523", "0.1528", "0.819", "0.1528", "0.819", "0.1674", "0.523", "0.1674" ], [ "0.5215", "0.1768", "0.8906", "0.1768", "0.8906", "0.1901", "0.5215", "0.1901" ], [ "0.2214", "0.1963", "0.2507", "0.1995", "0.25", "0.207", "0.2197", "0.2039" ], [ "0.4192", "0.1976", "0.423", "0.1995", "0.4192", "0.2052", "0.4143", "0.2026" ], [ "0.535", "0.1932", "0.8906", "0.1932", "0.8906", "0.207", "0.535", "0.207" ], [ "0.3774", "0.1989", "0.4028", "0.1989", "0.4028", "0.2058", "0.3774", "0.2058" ], [ "0.08826", "0.1989", "0.1259", "0.1989", "0.1259", "0.2058", "0.08826", "0.2058" ], [ "0.3162", "0.2008", "0.339", "0.2008", "0.339", "0.2052", "0.3162", "0.2052" ], [ "0.4207", "0.1989", "0.451", "0.1989", "0.451", "0.2058", "0.4207", "0.2058" ], [ "0.535", "0.209", "0.892", "0.209", "0.892", "0.2229", "0.535", "0.2229" ], [ "0.536", "0.226", "0.8896", "0.226", "0.8896", "0.2361", "0.536", "0.2361" ], [ "0.535", "0.2393", "0.8906", "0.2393", "0.8906", "0.2524", "0.535", "0.2524" ], [ "0.535", "0.2551", "0.8716", "0.2551", "0.8716", "0.2683", "0.535", "0.2683" ], [ "0.10455", "0.2778", "0.1691", "0.2778", "0.1691", "0.2854", "0.10455", "0.2854" ], [ "0.2115", "0.276", "0.2778", "0.276", "0.2778", "0.286", "0.2115", "0.286" ], [ "0.3064", "0.2778", "0.398", "0.2778", "0.398", "0.2854", "0.3064", "0.2854" ], [ "0.523", "0.279", "0.892", "0.279", "0.892", "0.2935", "0.523", "0.2935" ], [ "0.0801", "0.2922", "0.4673", "0.2922", "0.4673", "0.3062", "0.0801", "0.3062" ], [ "0.5337", "0.2942", "0.892", "0.2935", "0.892", "0.308", "0.5337", "0.3093" ], [ "0.0809", "0.3093", "0.468", "0.3093", "0.468", "0.3193", "0.0809", "0.3193" ], [ "0.535", "0.31", "0.8594", "0.31", "0.8594", "0.324", "0.535", "0.324" ], [ "0.0784", "0.3208", "0.468", "0.3208", "0.468", "0.334", "0.0784", "0.334" ], [ "0.0784", "0.3352", "0.468", "0.3352", "0.468", "0.3484", "0.0784", "0.3484" ], [ "0.5215", "0.3333", "0.8906", "0.334", "0.8906", "0.3484", "0.5215", "0.348" ], [ "0.0972", "0.356", "0.468", "0.3567", "0.468", "0.3718", "0.0972", "0.3706" ], [ "0.535", "0.351", "0.892", "0.351", "0.892", "0.3643", "0.535", "0.3643" ], [ "0.5337", "0.3643", "0.8906", "0.3643", "0.8906", "0.3774", "0.5337", "0.3774" ], [ "0.0801", "0.3738", "0.4697", "0.3738", "0.4697", "0.387", "0.0801", "0.387" ], [ "0.535", "0.3806", "0.8906", "0.3806", "0.8906", "0.3945", "0.535", "0.3945" ], [ "0.0784", "0.3882", "0.4722", "0.3882", "0.4722", "0.4016", "0.0784", "0.4016" ], [ "0.536", "0.3977", "0.892", "0.3977", "0.892", "0.4077", "0.536", "0.4077" ], [ "0.0809", "0.4048", "0.468", "0.4048", "0.468", "0.4153", "0.0809", "0.4153" ], [ "0.535", "0.411", "0.8364", "0.411", "0.8364", "0.4248", "0.535", "0.4248" ], [ "0.0801", "0.4185", "0.468", "0.4185", "0.468", "0.432", "0.0801", "0.432" ], [ "0.0784", "0.4343", "0.4697", "0.433", "0.4697", "0.4463", "0.0784", "0.4475" ], [ "0.0801", "0.449", "0.4697", "0.449", "0.4697", "0.4622", "0.0801", "0.4622" ], [ "0.5034", "0.4495", "0.6406", "0.4495", "0.6406", "0.464", "0.5034", "0.464" ], [ "0.0801", "0.464", "0.4714", "0.464", "0.4714", "0.4778", "0.0801", "0.4778" ], [ "0.0801", "0.4792", "0.4697", "0.4792", "0.4697", "0.4924", "0.0801", "0.4924" ], [ "0.5024", "0.4749", "0.892", "0.4749", "0.892", "0.488", "0.5024", "0.488" ], [ "0.0784", "0.4944", "0.4656", "0.4937", "0.4656", "0.507", "0.0784", "0.5083" ], [ "0.5024", "0.4905", "0.892", "0.4905", "0.892", "0.504", "0.5024", "0.504" ], [ "0.0801", "0.509", "0.4697", "0.509", "0.4697", "0.523", "0.0801", "0.523" ], [ "0.5024", "0.506", "0.8936", "0.506", "0.8936", "0.5195", "0.5024", "0.5195" ], [ "0.0784", "0.5234", "0.4697", "0.5234", "0.4697", "0.537", "0.0784", "0.537" ], [ "0.5024", "0.521", "0.892", "0.521", "0.892", "0.534", "0.5024", "0.534" ], [ "0.0801", "0.539", "0.4697", "0.539", "0.4697", "0.553", "0.0801", "0.553" ], [ "0.5024", "0.535", "0.8906", "0.535", "0.8906", "0.549", "0.5024", "0.549" ], [ "0.0801", "0.5537", "0.4697", "0.5537", "0.4697", "0.5674", "0.0801", "0.5674" ], [ "0.5024", "0.5503", "0.8906", "0.5503", "0.8906", "0.5645", "0.5024", "0.5645" ], [ "0.0809", "0.571", "0.468", "0.571", "0.468", "0.5806", "0.0809", "0.5806" ], [ "0.5024", "0.5664", "0.892", "0.5664", "0.892", "0.58", "0.5024", "0.58" ], [ "0.0801", "0.5854", "0.468", "0.5854", "0.468", "0.5986", "0.0801", "0.5986" ], [ "0.5034", "0.582", "0.8906", "0.582", "0.8906", "0.595", "0.5034", "0.595" ], [ "0.09644", "0.5986", "0.4697", "0.5996", "0.4697", "0.6143", "0.09644", "0.613" ], [ "0.501", "0.595", "0.8896", "0.595", "0.8896", "0.6094", "0.501", "0.6094" ], [ "0.0784", "0.613", "0.468", "0.6157", "0.468", "0.63", "0.0784", "0.6274" ], [ "0.5024", "0.6123", "0.892", "0.6123", "0.892", "0.6255", "0.5024", "0.6255" ], [ "0.0801", "0.629", "0.4673", "0.629", "0.4673", "0.643", "0.0801", "0.643" ], [ "0.501", "0.625", "0.892", "0.6255", "0.892", "0.64", "0.501", "0.6396" ], [ "0.501", "0.6396", "0.8906", "0.64", "0.8906", "0.655", "0.501", "0.654" ], [ "0.0801", "0.646", "0.4697", "0.646", "0.4697", "0.659", "0.0801", "0.659" ], [ "0.0801", "0.66", "0.4697", "0.66", "0.4697", "0.674", "0.0801", "0.674" ], [ "0.501", "0.656", "0.8906", "0.657", "0.8906", "0.6704", "0.501", "0.6694" ], [ "0.0768", "0.675", "0.4697", "0.674", "0.4697", "0.688", "0.0768", "0.6895" ], [ "0.5024", "0.672", "0.892", "0.672", "0.892", "0.685", "0.5024", "0.685" ], [ "0.0801", "0.69", "0.468", "0.69", "0.468", "0.704", "0.0801", "0.704" ], [ "0.5024", "0.6875", "0.8906", "0.6875", "0.8906", "0.7007", "0.5024", "0.7007" ], [ "0.0801", "0.7056", "0.4697", "0.7056", "0.4697", "0.7197", "0.0801", "0.7197" ], [ "0.5024", "0.702", "0.892", "0.702", "0.892", "0.7153", "0.5024", "0.7153" ], [ "0.0801", "0.72", "0.4697", "0.72", "0.4697", "0.7344", "0.0801", "0.7344" ], [ "0.5024", "0.7173", "0.8906", "0.7173", "0.8906", "0.731", "0.5024", "0.731" ], [ "0.501", "0.731", "0.892", "0.732", "0.892", "0.747", "0.501", "0.7456" ], [ "0.0809", "0.7373", "0.4673", "0.7373", "0.4673", "0.7476", "0.0809", "0.7476" ], [ "0.5", "0.7456", "0.8936", "0.747", "0.8936", "0.761", "0.5", "0.7603" ], [ "0.0801", "0.752", "0.468", "0.752", "0.468", "0.765", "0.0801", "0.765" ], [ "0.0801", "0.7666", "0.468", "0.7666", "0.468", "0.78", "0.0801", "0.78" ], [ "0.5024", "0.762", "0.892", "0.762", "0.892", "0.776", "0.5024", "0.776" ], [ "0.0768", "0.78", "0.468", "0.781", "0.468", "0.7954", "0.0768", "0.7944" ], [ "0.5034", "0.7764", "0.8896", "0.7764", "0.8896", "0.7905", "0.5034", "0.7905" ], [ "0.0784", "0.797", "0.4714", "0.797", "0.4714", "0.81", "0.0784", "0.81" ], [ "0.5034", "0.7935", "0.8906", "0.7935", "0.8906", "0.8037", "0.5034", "0.8037" ], [ "0.0801", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.0801", "0.8247" ], [ "0.501", "0.8066", "0.8906", "0.806", "0.8906", "0.821", "0.501", "0.8213" ], [ "0.0801", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0801", "0.8403" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.5034", "0.8403", "0.892", "0.8403", "0.892", "0.855", "0.5034", "0.855" ], [ "0.0801", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0801", "0.8706" ], [ "0.5024", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5024", "0.8706" ], [ "0.0801", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0801", "0.8853" ], [ "0.5", "0.8706", "0.892", "0.872", "0.892", "0.886", "0.5", "0.8853" ], [ "0.0768", "0.8853", "0.4697", "0.886", "0.4697", "0.901", "0.0768", "0.8994" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ] ]

[ "visibleparts", "objectcenter★", "direction→", "the background prototype, to produce a mask for decreas-", "ing the impact of noise. IRA displaces pseudo-labels with", "dtruth", "high-quality samples that have similar geometric structures,", "enriching the diversity on the target domain.", "The main contributions of this paper include:", ". We propose a novel UDA framework for LiDAR-", "nrototy", "C", "based 3D detectors,namely Geometry-awareProto-", "ecisio", "sourceO", "rouo", "undary", "type Alignment (GPA-3D). It explicitly integrates geo-", "metric associations intofeaturealignment,effectively", "decreasing the distributional discrepancy and facilitat-", "ing the adaptation of existing point cloud detectors.", "(c)w/oalignme", "(d) vanilla alignr", "(e)w/geometry-aw", ": Noise sample suppression and instance replacement", "Figure 2. (a) Point cloud scene on BEV (bird's-eye-view). (b)", "augmentation are designed to enhance pseudo-labels", "Distinctgeometric structures of point cloud objects.(c-e)Illus-", "in terms of reliability and versatility, respectively.", "tration of the distributional discrepancy. With explicit geometric", "constraints, the features from different domains are better aligned.", ": We conduct comprehensive experiments on Waymo,", "To reduce this discrepancy in 2D UDA task, some ap-", "nuScenes, and KITTI. The encouraging results demon-", "strate the GPA-3D outperforms state-of-the-art meth-", "proaches utilize the class-wise prototypes align features", "ods in various adaptation scenarios. More importantly,", "from different domains [12, 32, 41, 19]. In these works, a", "thankstothe architecture-agnosticdesign,GPA-3Dis", "universalprototypeisemployedtoenforcehighrepresenta-", "flexible to be applied to point cloud detectors.", "tional similarity among features belonging to the same cat-", "egory. However, in the case of 3D scenes, such as vehicles", "on the road, diverse locations and directions can result in", "2. Related Work", "distinct geometric structures, i.e., distributional patterns of", "point clouds, as presented in Fig. 2 (a) and (b). If a uniform", "LiDAR-based 3D Detection.Mainstream point cloud de-", "prototype is applied to objects with completely different ge", "tectors can be broadly divided into two categories: point-", "ometric structures, the efficacy of feature alignment might", "based and grid-based. Point-based methods mainly adopt", "be hindered, as illustrated in Fig. 2 (c-d). We argue that", "the architectures of PointNet [23] and PointNet++ [24] to", "adopting different prototypes to point cloud objects with", "extract features from raw point clouds. PointRCNN [29]", "distinct geometric structures could deal with the problem of", "designs an encoder-decoder backbone to learn the point-", "distributionaldiscrepancybutmoreattentionshouldalsobe", "wise representation. 3DSSD [39] improves the point sam-", "paid to model these geometric structures during adaptation.", "pling operator from the aspect of feature distance. IA-", "Based on the considerations, we propose a novel", "SSD [42] utilizes the instance-aware downsampling to pre-", "UDA framework for LiDAR-based 3D detectors, namely", "serve more foregrounds.On the other hand, grid-based", "Geometry-aware Prototype Alignment (GPA-3D). Con-", "methods first divide point clouds into fixed-size voxels,", "which are then processed via 2D/3D CNN. SECOND [35]", "cretely, we first explore the potential relationships between", "the geometric structures of point cloud objects. During", "adopts sparse 3D convolution for efficient feature learning.", "training, we randomly extract the BEV features of point", "PointPillars [16] proposes a pillar encoding method and", "clouds from both the source and target domains, and sub-", "achieves a good trade-off between speed and performance.", "sequently divide them into distinct groups based on their", "PV-RCNN [28] incorporates the voxel backbone with the", "geometric structures. In this process, BEV features derived", "keypoint branch to learn the representative scene features.", "Other approaches [36, 4, 31] project the point clouds into", "frompointclouds withthe similargeometric structures will", "certain kinds of 2D views, and employ 2D CNN to extract", "be classified into the same group. Each group is then as-", "signed a unique prototype, which enforces high representa-", "the features. In this work, we conduct focused discussions", "tional similarity among the BEV features within that group,", "with SECOND [35] and PointPillars [16] as base detectors.", "as illustrated in Fig. 2 (e). To this end, the soft contrast loss", "Todemonstratethegeneralizationabilityof ourmethod.we", "is devised to pull the intra-group feature-prototype pairs", "also provide the comparisons with PV-RCNN [28] detector.", "closer in the representational space and push the inter-group", "pairs farther away. Additionally, we develop the frame-", "Domain Adaptive Object Detection.A large amount of", "work with two components, namely noise sample suppres-", "literature has been presented in UDA for 2D-image de-", "sion (NSS) and instance replacement augmentation (IRA).", "tection, which can be roughly classified into two groups:", " NSS utilizes the similarities between foreground areas and", "distribution alignment and self-training. Alignment-based" ]

{ "arxiv_id": "2308.08140", "language": "en", "page_idx": 1 }

[ [ "0.0801", "0.09155", "0.4673", "0.0903", "0.4673", "0.1048", "0.0801", "0.1061" ], [ "0.0801", "0.1067", "0.468", "0.1067", "0.468", "0.1206", "0.0801", "0.1206" ], [ "0.0784", "0.1225", "0.4395", "0.1218", "0.4395", "0.1364", "0.0784", "0.137" ], [ "0.511", "0.1206", "0.569", "0.1206", "0.569", "0.1307", "0.511", "0.1307" ], [ "0.5186", "0.1307", "0.55", "0.1307", "0.55", "0.1383", "0.5186", "0.1383" ], [ "0.2051", "0.1427", "0.343", "0.1427", "0.343", "0.1572", "0.2051", "0.1572" ], [ "0.4468", "0.1427", "0.468", "0.1427", "0.468", "0.1559", "0.4468", "0.1559" ], [ "0.5107", "0.1401", "0.5664", "0.1401", "0.5664", "0.1477", "0.5107", "0.1477" ], [ "0.5107", "0.1483", "0.576", "0.1483", "0.576", "0.159", "0.5107", "0.159" ], [ "0.0801", "0.1604", "0.4714", "0.161", "0.4714", "0.1755", "0.0801", "0.1749" ], [ "0.511", "0.1604", "0.585", "0.1604", "0.585", "0.1642", "0.511", "0.1642" ], [ "0.511", "0.1686", "0.5664", "0.1686", "0.5664", "0.1755", "0.511", "0.1755" ], [ "0.0801", "0.178", "0.4673", "0.178", "0.4673", "0.1913", "0.0801", "0.1913" ], [ "0.0801", "0.1925", "0.468", "0.1925", "0.468", "0.2058", "0.0801", "0.2058" ], [ "0.5024", "0.1881", "0.892", "0.1881", "0.892", "0.202", "0.5024", "0.202" ], [ "0.0801", "0.207", "0.468", "0.207", "0.468", "0.2203", "0.0801", "0.2203" ], [ "0.5024", "0.202", "0.8423", "0.202", "0.8423", "0.2153", "0.5024", "0.2153" ], [ "0.0801", "0.2229", "0.468", "0.2229", "0.468", "0.2361", "0.0801", "0.2361" ], [ "0.5034", "0.2267", "0.892", "0.2267", "0.892", "0.2443", "0.5034", "0.2443" ], [ "0.0784", "0.2374", "0.4673", "0.2361", "0.4673", "0.2507", "0.0784", "0.252" ], [ "0.5024", "0.2437", "0.892", "0.2443", "0.892", "0.2588", "0.5024", "0.2583" ], [ "0.0801", "0.2615", "0.407", "0.2615", "0.407", "0.276", "0.0801", "0.276" ], [ "0.5024", "0.2588", "0.8022", "0.26", "0.8022", "0.2747", "0.5024", "0.2734" ], [ "0.0809", "0.286", "0.4673", "0.286", "0.4673", "0.2966", "0.0809", "0.2966" ], [ "0.5034", "0.2842", "0.6787", "0.2842", "0.6787", "0.2974", "0.5034", "0.2974" ], [ "0.0801", "0.2998", "0.468", "0.2998", "0.468", "0.3132", "0.0801", "0.3132" ], [ "0.0784", "0.3113", "0.4714", "0.3113", "0.4714", "0.3289", "0.0784", "0.3289" ], [ "0.5215", "0.307", "0.892", "0.307", "0.892", "0.3208", "0.5215", "0.3208" ], [ "0.5034", "0.3225", "0.8906", "0.3225", "0.8906", "0.3372", "0.5034", "0.3372" ], [ "0.0809", "0.3289", "0.468", "0.3289", "0.468", "0.3435", "0.0809", "0.3435" ], [ "0.501", "0.3384", "0.634", "0.3372", "0.634", "0.3516", "0.501", "0.3528" ], [ "0.0784", "0.3447", "0.468", "0.3435", "0.468", "0.358", "0.0784", "0.3591" ], [ "0.0801", "0.3599", "0.4697", "0.3599", "0.4697", "0.3738", "0.0801", "0.3738" ], [ "0.5034", "0.3643", "0.8823", "0.3643", "0.8823", "0.379", "0.5034", "0.379" ], [ "0.0801", "0.375", "0.468", "0.375", "0.468", "0.3882", "0.0801", "0.3882" ], [ "0.501", "0.379", "0.8896", "0.379", "0.8896", "0.3933", "0.501", "0.3933" ], [ "0.0801", "0.387", "0.4697", "0.387", "0.4697", "0.4048", "0.0801", "0.4048" ], [ "0.5034", "0.3958", "0.6914", "0.3958", "0.6914", "0.4092", "0.5034", "0.4092" ], [ "0.0801", "0.4048", "0.468", "0.4048", "0.468", "0.4192", "0.0801", "0.4192" ], [ "0.574", "0.4136", "0.5884", "0.4136", "0.5884", "0.4216", "0.574", "0.4216" ], [ "0.591", "0.4153", "0.597", "0.4153", "0.597", "0.4192", "0.591", "0.4192" ], [ "0.596", "0.4104", "0.6406", "0.4124", "0.6396", "0.4236", "0.5947", "0.4216" ], [ "0.0801", "0.4204", "0.452", "0.4204", "0.452", "0.435", "0.0801", "0.435" ], [ "0.513", "0.4185", "0.8945", "0.4185", "0.8945", "0.443", "0.513", "0.443" ], [ "0.572", "0.4443", "0.6357", "0.4443", "0.6357", "0.452", "0.572", "0.452" ], [ "0.0809", "0.4539", "0.468", "0.4539", "0.468", "0.4673", "0.0809", "0.4673" ], [ "0.4983", "0.4539", "0.8945", "0.4526", "0.8945", "0.4778", "0.4983", "0.4792" ], [ "0.0801", "0.4685", "0.4673", "0.4685", "0.4673", "0.4817", "0.0801", "0.4817" ], [ "0.5024", "0.4749", "0.892", "0.473", "0.892", "0.4873", "0.5024", "0.4893" ], [ "0.0809", "0.4841", "0.468", "0.4841", "0.468", "0.4976", "0.0809", "0.4976" ], [ "0.5024", "0.4905", "0.8906", "0.488", "0.8906", "0.5024", "0.5024", "0.505" ], [ "0.0801", "0.4988", "0.468", "0.4988", "0.468", "0.512", "0.0801", "0.512" ], [ "0.5024", "0.505", "0.892", "0.505", "0.892", "0.5195", "0.5024", "0.5195" ], [ "0.0801", "0.515", "0.281", "0.515", "0.281", "0.53", "0.0801", "0.53" ], [ "0.501", "0.521", "0.7983", "0.5195", "0.7983", "0.534", "0.501", "0.535" ], [ "0.2133", "0.53", "0.335", "0.532", "0.335", "0.5503", "0.2133", "0.549" ], [ "0.4468", "0.534", "0.4697", "0.534", "0.4697", "0.5474", "0.4468", "0.5474" ], [ "0.625", "0.536", "0.665", "0.536", "0.665", "0.5474", "0.625", "0.5474" ], [ "0.6094", "0.55", "0.6357", "0.55", "0.6357", "0.563", "0.6094", "0.563" ], [ "0.6807", "0.549", "0.8286", "0.549", "0.8286", "0.5664", "0.6807", "0.5664" ], [ "0.869", "0.55", "0.892", "0.55", "0.892", "0.563", "0.869", "0.563" ], [ "0.661", "0.552", "0.6855", "0.552", "0.6855", "0.56", "0.661", "0.56" ], [ "0.0809", "0.5537", "0.468", "0.5537", "0.468", "0.5684", "0.0809", "0.5684" ], [ "0.569", "0.5503", "0.6167", "0.5503", "0.6167", "0.565", "0.569", "0.565" ], [ "0.6465", "0.5537", "0.6587", "0.5537", "0.6587", "0.56", "0.6465", "0.56" ], [ "0.0784", "0.5693", "0.468", "0.571", "0.468", "0.5854", "0.0784", "0.584" ], [ "0.6216", "0.5684", "0.6665", "0.5684", "0.6665", "0.5757", "0.6216", "0.5757" ], [ "0.5034", "0.5776", "0.6406", "0.5776", "0.6406", "0.5923", "0.5034", "0.5923" ], [ "0.6406", "0.577", "0.8906", "0.5776", "0.8906", "0.5923", "0.6406", "0.5913" ], [ "0.0809", "0.5854", "0.4673", "0.5854", "0.4673", "0.5996", "0.0809", "0.5996" ], [ "0.6396", "0.5835", "0.6465", "0.5835", "0.6465", "0.587", "0.6396", "0.587" ], [ "0.501", "0.5923", "0.892", "0.5933", "0.892", "0.608", "0.501", "0.607" ], [ "0.0801", "0.601", "0.468", "0.6016", "0.468", "0.616", "0.0801", "0.6157" ], [ "0.5024", "0.6094", "0.8906", "0.6094", "0.8906", "0.6226", "0.5024", "0.6226" ], [ "0.0801", "0.617", "0.468", "0.617", "0.468", "0.631", "0.0801", "0.631" ], [ "0.501", "0.6226", "0.8906", "0.6235", "0.8906", "0.638", "0.501", "0.637" ], [ "0.0801", "0.6333", "0.4575", "0.6333", "0.4575", "0.6465", "0.0801", "0.6465" ], [ "0.501", "0.6396", "0.6665", "0.637", "0.6665", "0.6514", "0.501", "0.654" ], [ "0.1765", "0.651", "0.3733", "0.6484", "0.3733", "0.6665", "0.1765", "0.6685" ], [ "0.4468", "0.653", "0.468", "0.653", "0.468", "0.6655", "0.4468", "0.6655" ], [ "0.6436", "0.654", "0.6787", "0.654", "0.6787", "0.6655", "0.6436", "0.6655" ], [ "0.5557", "0.6636", "0.5947", "0.6753", "0.5864", "0.6914", "0.5474", "0.6787" ], [ "0.6724", "0.6655", "0.8447", "0.6665", "0.8447", "0.6846", "0.6724", "0.683" ], [ "0.8677", "0.6685", "0.8906", "0.6685", "0.8906", "0.681", "0.8677", "0.681" ], [ "0.0801", "0.686", "0.4673", "0.686", "0.4673", "0.6997", "0.0801", "0.6997" ], [ "0.0801", "0.6704", "0.4697", "0.672", "0.4697", "0.686", "0.0801", "0.685" ], [ "0.6562", "0.6895", "0.6665", "0.6895", "0.6665", "0.6934", "0.6562", "0.6934" ], [ "0.0801", "0.702", "0.468", "0.7007", "0.468", "0.7153", "0.0801", "0.7163" ], [ "0.5024", "0.6978", "0.892", "0.6978", "0.892", "0.711", "0.5024", "0.711" ], [ "0.0801", "0.7173", "0.468", "0.7173", "0.468", "0.731", "0.0801", "0.731" ], [ "0.5024", "0.7134", "0.8906", "0.7134", "0.8906", "0.7266", "0.5024", "0.7266" ], [ "0.0801", "0.733", "0.468", "0.733", "0.468", "0.747", "0.0801", "0.747" ], [ "0.5024", "0.728", "0.892", "0.728", "0.892", "0.741", "0.5024", "0.741" ], [ "0.0801", "0.7476", "0.4673", "0.7476", "0.4673", "0.761", "0.0801", "0.761" ], [ "0.5024", "0.7437", "0.892", "0.7437", "0.892", "0.757", "0.5024", "0.757" ], [ "0.0801", "0.763", "0.468", "0.763", "0.468", "0.778", "0.0801", "0.778" ], [ "0.5024", "0.7583", "0.7915", "0.7583", "0.7915", "0.7715", "0.5024", "0.7715" ], [ "0.584", "0.7764", "0.5923", "0.7764", "0.5923", "0.782", "0.584", "0.782" ], [ "0.597", "0.7764", "0.603", "0.7764", "0.603", "0.7827", "0.597", "0.7827" ], [ "0.603", "0.776", "0.621", "0.776", "0.621", "0.784", "0.603", "0.784" ], [ "0.6465", "0.784", "0.8594", "0.785", "0.8594", "0.8037", "0.6465", "0.803" ], [ "0.0801", "0.778", "0.312", "0.778", "0.312", "0.7925", "0.0801", "0.7925" ], [ "0.513", "0.787", "0.5986", "0.787", "0.5986", "0.802", "0.513", "0.802" ], [ "0.5933", "0.786", "0.6235", "0.7886", "0.621", "0.803", "0.591", "0.8003" ], [ "0.6357", "0.789", "0.653", "0.789", "0.653", "0.7974", "0.6357", "0.7974" ], [ "0.5728", "0.8", "0.6665", "0.803", "0.6665", "0.8174", "0.5728", "0.8145" ], [ "0.0809", "0.81", "0.468", "0.81", "0.468", "0.8247", "0.0809", "0.8247" ], [ "0.5923", "0.821", "0.6045", "0.821", "0.6045", "0.83", "0.5923", "0.83" ], [ "0.6113", "0.8193", "0.6313", "0.8213", "0.63", "0.8306", "0.6094", "0.829" ], [ "0.603", "0.8306", "0.63", "0.8306", "0.63", "0.8496", "0.603", "0.8496" ], [ "0.0801", "0.8247", "0.4673", "0.8257", "0.4673", "0.8403", "0.0801", "0.839" ], [ "0.511", "0.8335", "0.568", "0.83", "0.569", "0.8447", "0.513", "0.8477" ], [ "0.664", "0.832", "0.689", "0.8276", "0.694", "0.8467", "0.669", "0.851" ], [ "0.721", "0.8335", "0.8813", "0.8335", "0.8813", "0.851", "0.721", "0.851" ], [ "0.584", "0.8384", "0.596", "0.834", "0.603", "0.8447", "0.591", "0.8486" ], [ "0.0801", "0.836", "0.4714", "0.836", "0.4714", "0.857", "0.0801", "0.857" ], [ "0.5254", "0.8433", "0.5415", "0.8433", "0.5415", "0.8477", "0.5254", "0.8477" ], [ "0.0784", "0.855", "0.468", "0.856", "0.468", "0.8706", "0.0784", "0.869" ], [ "0.585", "0.8496", "0.722", "0.8516", "0.722", "0.8657", "0.585", "0.863" ], [ "0.0801", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0801", "0.8853" ], [ "0.4983", "0.8687", "0.8945", "0.8613", "0.8945", "0.886", "0.4983", "0.893" ], [ "0.0784", "0.886", "0.4697", "0.886", "0.4697", "0.904", "0.0784", "0.904" ], [ "0.501", "0.886", "0.8906", "0.887", "0.8906", "0.9014", "0.501", "0.901" ] ]

[ "where hi is the backbone with parameters i, and F indi-", "cates the BEV features. After that, a detection head h2 with", "parameters 92 produces the final output, formulated as:", "visiblearea", "directio", "[b, s]i = h2(F:; 02),", "(2)", "★objectcent", "S rotation angle", "where b and s represent the predicted 3D boxes and scores", "Soheaniatinnana", "Soffsetangle", "respectively: A co-training paradigm is applied to progres-", "sively mitigate the domain shift. In each mini-batch, both", "Figure 4. Left: Demonstration of the offset angle. Right: Objects", "the source point clouds Ps and target point clouds Pt are", "with same offset angle share similar geometric structures.", "sent to the detector, and their outputs are supervised by the", "aligned with the foreground prototype gk(ke[1,K]). Sim-", "corresponding ground truth and pseudo-labels, respectively", "ilarly, the background features F, in the last group are", "3.3. Geometry-aware Prototype Alignment", "aligned with the background prototype gk+1.", "Extract..As mentioned in Sec.3.2.for i-th point cloud", "3.4. Soft Contrast Loss", "scenario P, from the source or target domain, LiDAR-based", "detector generates the BEV features F; e RHxWxC, where", "Given a point cloud Pi, our goal is to align its", "fore/background features F+ and F-with the correspond-", "H, W, and C denote the height, width and channel num-", "ing prototypes in g.", "bers of the feature map.We first project the correspond-", "ing ground truth L, or pseudo-labels L, to the BEV feature", "Intra-group Attract. For the foreground features F", "map, and then randomly extract the equal-length sequences", "we pull them closer with the corresponding prototype in g", "F+ RMxC and F- e IRMixC. Here, M, is the length", "which can be formulated as:", "of the feature sequence, F+ and F- represent the fore-", "K", "", "NM", "ground and background features from BEV, respectively.", "Ctt=(1- sim(Ft,gk)1[Qij =k], (5)", "=1=1", "Group.For the extracted foreground features F+, we fur-", "where sim(a,b) = a is the cosine similarity.", "ther divide them into different groups according to their ge-", "I[Qi,j = k] is an indicator function that equals to 1 if", "ometric structures on point clouds. Specifically, for j-th", "Qi,j = k and O otherwise. Similarly, the background fea-", "foreground Ft in the sequence (Gi e [1, Mil), we compute", "tures F- are also required to be pulled to the background", "its offset angle ooff as follows:", "prototype gK+1, which can be calculated as:", "ooff = Oobs - ri.j,", "(3)", "NM", ">", "- sim(F.j,gK+1)).", "(6)", "71", "where ri is the direction, oobs is the observation angle, as", "Catt =", ">", "presented in Fig. 4 (left). Note that the direction ri, is pro-", "=1=1", "Inter-group Repel.", "To enhance the discriminative capac-", "vided from the labels L: and Lt, while the observation an-", "", "ity, we need to push the features away from all prototypes", "gle oobs can be computed according to the central position", "belonging to other groups. For example, the distances be-", "of 3D bounding box. Next, all foreground features are split", "tween background features F and all foreground proto-", "into K groups, and the group index Qi., is formulated as:", "types are minimized via:", "Qi,j = [norm(0o)/8] +1,", "(4)", "VM", "Crep", "Cmar(0, sim(Fr.j, gk)).", "(7)", "the input angles into [0,2元l, and =- 2/K is the inter-", "where norm() is a normalization function that converting", "T", "val of angles between groups. In this way, the foreground", "For foreground features within adjacent groups, their corre-", "features with similar offset angles Ooff are assigned into the", "sponding geometric structures are relatively more similar.", "same group, where their geometric structures are very sim-", "Repelling thesefeatures away is not very necessary,and", "ilar, as demonstrated in Fig. 4 (right). Additionally, the ex-", "might even make the training process unstable. Hence, we", "tracted backgrounds F., are sent into an individual group,", "adopt a more relaxed constraints as follows:", "V", "", "M", " mar(0, sim(F,gk) -m),", "thus totally K + 1 groups are built.", "Ctedi =>", "V", "", "i=ikeAr,", "Prototype Construction. At the beginning of training,", "N", "M", "", "we randomly initialize a series of learnable prototypes g =", "Ctother", "D", "mar(0, sim(F+,gk),", "J", "(gk)K+I e R(K+1)xC, During training, we extract the", "reT", "BEV features F from both source and target domains,", "j=AQi", "and split them into corresponding groups via Eq. 4. In k-", "where m indicates the margin which is set to 0.5 in our R)", "th group, the foreground features Ft, are enforced to be", "periments, Ai,j is the index of the groups adjacent to Qi,j." ]

{ "arxiv_id": "2308.08140", "language": "en", "page_idx": 3 }

[ [ "0.1217", "0.0947", "0.2303", "0.0947", "0.2303", "0.1048", "0.1217", "0.1048" ], [ "0.511", "0.0966", "0.839", "0.0966", "0.839", "0.10986", "0.511", "0.10986" ], [ "0.1438", "0.1111", "0.1691", "0.1111", "0.1691", "0.1193", "0.1438", "0.1193" ], [ "0.1871", "0.1111", "0.2133", "0.1111", "0.2133", "0.1187", "0.1871", "0.1187" ], [ "0.2255", "0.1111", "0.2507", "0.1111", "0.2507", "0.1193", "0.2255", "0.1193" ], [ "0.5215", "0.113", "0.859", "0.113", "0.859", "0.127", "0.5215", "0.127" ], [ "0.4126", "0.1237", "0.4534", "0.1237", "0.4534", "0.1351", "0.4126", "0.1351" ], [ "0.10706", "0.113", "0.125", "0.113", "0.125", "0.11743", "0.10706", "0.11743" ], [ "0.5645", "0.1237", "0.8555", "0.125", "0.8555", "0.1395", "0.5645", "0.1383" ], [ "0.156", "0.1338", "0.1969", "0.1338", "0.1969", "0.1414", "0.156", "0.1414" ], [ "0.5664", "0.137", "0.6895", "0.1383", "0.6895", "0.1497", "0.5664", "0.1483" ], [ "0.08905", "0.1465", "0.2573", "0.1465", "0.2573", "0.1572", "0.08905", "0.1572" ], [ "0.523", "0.1497", "0.6895", "0.1497", "0.6895", "0.1604", "0.523", "0.1604" ], [ "0.509", "0.161", "0.8096", "0.161", "0.8096", "0.1749", "0.509", "0.1749" ], [ "0.4084", "0.1724", "0.4558", "0.1724", "0.4558", "0.1831", "0.4084", "0.1831" ], [ "0.509", "0.1749", "0.621", "0.1749", "0.621", "0.1849", "0.509", "0.1849" ], [ "0.1609", "0.1768", "0.1683", "0.1768", "0.1683", "0.1799", "0.1609", "0.1799" ], [ "0.1013", "0.1768", "0.11926", "0.1768", "0.11926", "0.1812", "0.1013", "0.1812" ], [ "0.2378", "0.178", "0.25", "0.178", "0.25", "0.1812", "0.2378", "0.1812" ], [ "0.4207", "0.1843", "0.4412", "0.1843", "0.4412", "0.1925", "0.4207", "0.1925" ], [ "0.5063", "0.1863", "0.8325", "0.1863", "0.8325", "0.1995", "0.5063", "0.1995" ], [ "0.1609", "0.1976", "0.1805", "0.1976", "0.1805", "0.2052", "0.1609", "0.2052" ], [ "0.509", "0.2008", "0.652", "0.2008", "0.652", "0.2108", "0.509", "0.2108" ], [ "0.546", "0.2122", "0.669", "0.2108", "0.669", "0.2229", "0.546", "0.2235" ], [ "0.1396", "0.2197", "0.1765", "0.2197", "0.1765", "0.2311", "0.1396", "0.2311" ], [ "0.1748", "0.2197", "0.2141", "0.2197", "0.2141", "0.2311", "0.1748", "0.2311" ], [ "0.572", "0.2229", "0.808", "0.2229", "0.808", "0.2361", "0.572", "0.2361" ], [ "0.2263", "0.2235", "0.2443", "0.2235", "0.2443", "0.2279", "0.2263", "0.2279" ], [ "0.10864", "0.2235", "0.11926", "0.2235", "0.11926", "0.2279", "0.10864", "0.2279" ], [ "0.4167", "0.226", "0.4478", "0.226", "0.4478", "0.2374", "0.4167", "0.2374" ], [ "0.11273", "0.2361", "0.2402", "0.2361", "0.2402", "0.2468", "0.11273", "0.2468" ], [ "0.4224", "0.238", "0.4412", "0.238", "0.4412", "0.2424", "0.4224", "0.2424" ], [ "0.572", "0.2374", "0.8096", "0.2361", "0.8096", "0.2507", "0.572", "0.252" ], [ "0.0825", "0.254", "0.4656", "0.254", "0.4656", "0.2644", "0.0825", "0.2644" ], [ "0.572", "0.252", "0.8164", "0.252", "0.8164", "0.2664", "0.572", "0.2664" ], [ "0.0801", "0.2664", "0.468", "0.2664", "0.468", "0.2798", "0.0801", "0.2798" ], [ "0.5474", "0.2664", "0.652", "0.2664", "0.652", "0.279", "0.5474", "0.279" ], [ "0.0801", "0.2798", "0.468", "0.2798", "0.468", "0.2935", "0.0801", "0.2935" ], [ "0.572", "0.2778", "0.773", "0.2778", "0.773", "0.2922", "0.572", "0.2922" ], [ "0.0801", "0.2935", "0.468", "0.2935", "0.468", "0.307", "0.0801", "0.307" ], [ "0.505", "0.2922", "0.572", "0.2922", "0.572", "0.3037", "0.505", "0.3037" ], [ "0.0809", "0.3093", "0.273", "0.3093", "0.273", "0.3193", "0.0809", "0.3193" ], [ "0.0768", "0.3308", "0.4673", "0.332", "0.4673", "0.3499", "0.0768", "0.3484" ], [ "0.5024", "0.3384", "0.8906", "0.339", "0.8906", "0.3542", "0.5024", "0.3528" ], [ "0.0809", "0.351", "0.2239", "0.351", "0.2239", "0.361", "0.0809", "0.361" ], [ "0.5", "0.3542", "0.8506", "0.3542", "0.8506", "0.3687", "0.5", "0.3687" ], [ "0.0801", "0.3643", "0.46", "0.3643", "0.46", "0.3818", "0.0801", "0.3818" ], [ "0.5215", "0.3687", "0.892", "0.3694", "0.892", "0.3838", "0.5215", "0.3833" ], [ "0.3628", "0.3757", "0.3694", "0.3757", "0.3694", "0.379", "0.3628", "0.379" ], [ "0.4468", "0.3818", "0.468", "0.3818", "0.468", "0.3945", "0.4468", "0.3945" ], [ "0.5024", "0.385", "0.8896", "0.385", "0.8896", "0.3984", "0.5024", "0.3984" ], [ "0.0784", "0.3958", "0.398", "0.3945", "0.398", "0.4092", "0.0784", "0.4104" ], [ "0.5034", "0.401", "0.892", "0.401", "0.892", "0.414", "0.5034", "0.414" ], [ "0.0784", "0.4185", "0.3162", "0.4192", "0.3162", "0.4375", "0.0784", "0.4363" ], [ "0.5034", "0.4172", "0.8906", "0.4172", "0.8906", "0.428", "0.5034", "0.428" ], [ "0.5024", "0.4312", "0.8906", "0.4312", "0.8906", "0.4443", "0.5024", "0.4443" ], [ "0.0989", "0.4443", "0.4673", "0.4443", "0.4673", "0.459", "0.0989", "0.459" ], [ "0.501", "0.4443", "0.649", "0.4458", "0.649", "0.4602", "0.501", "0.459" ], [ "0.0801", "0.461", "0.468", "0.461", "0.468", "0.4749", "0.0801", "0.4749" ], [ "0.0784", "0.4749", "0.468", "0.476", "0.468", "0.4905", "0.0784", "0.4893" ], [ "0.5034", "0.4702", "0.7485", "0.4702", "0.7485", "0.4841", "0.5034", "0.4841" ], [ "0.0801", "0.4905", "0.468", "0.4905", "0.468", "0.504", "0.0801", "0.504" ], [ "0.5215", "0.4937", "0.892", "0.4937", "0.892", "0.5083", "0.5215", "0.5083" ], [ "0.0801", "0.506", "0.468", "0.506", "0.468", "0.5195", "0.0801", "0.5195" ], [ "0.5024", "0.509", "0.892", "0.509", "0.892", "0.5234", "0.5024", "0.5234" ], [ "0.0801", "0.521", "0.4697", "0.521", "0.4697", "0.534", "0.0801", "0.534" ], [ "0.5024", "0.5234", "0.892", "0.5234", "0.892", "0.537", "0.5024", "0.537" ], [ "0.076", "0.5327", "0.4722", "0.53", "0.4722", "0.5503", "0.076", "0.5537" ], [ "0.5024", "0.5405", "0.7534", "0.5405", "0.7534", "0.5547", "0.5024", "0.5547" ], [ "0.0801", "0.5503", "0.468", "0.5503", "0.468", "0.5645", "0.0801", "0.5645" ], [ "0.6284", "0.558", "0.767", "0.558", "0.767", "0.5757", "0.6284", "0.5757" ], [ "0.863", "0.56", "0.892", "0.56", "0.892", "0.5728", "0.863", "0.5728" ], [ "0.0801", "0.5664", "0.4697", "0.5664", "0.4697", "0.58", "0.0801", "0.58" ], [ "0.0801", "0.582", "0.468", "0.582", "0.468", "0.595", "0.0801", "0.595" ], [ "0.5024", "0.58", "0.8936", "0.58", "0.8936", "0.5977", "0.5024", "0.5977" ], [ "0.0801", "0.5967", "0.4697", "0.5967", "0.4697", "0.6113", "0.0801", "0.6113" ], [ "0.5024", "0.5967", "0.892", "0.5967", "0.892", "0.61", "0.5024", "0.61" ], [ "0.0801", "0.6123", "0.429", "0.6123", "0.429", "0.6255", "0.0801", "0.6255" ], [ "0.501", "0.6113", "0.8906", "0.61", "0.8906", "0.625", "0.501", "0.6255" ], [ "0.0989", "0.6255", "0.4673", "0.6255", "0.4673", "0.6396", "0.0989", "0.6396" ], [ "0.501", "0.625", "0.892", "0.6255", "0.892", "0.64", "0.501", "0.6396" ], [ "0.0768", "0.64", "0.4697", "0.6396", "0.4697", "0.657", "0.0768", "0.6577" ], [ "0.5024", "0.6416", "0.8677", "0.6416", "0.8677", "0.655", "0.5024", "0.655" ], [ "0.0801", "0.657", "0.468", "0.657", "0.468", "0.672", "0.0801", "0.672" ], [ "0.613", "0.659", "0.7837", "0.661", "0.7837", "0.6787", "0.613", "0.677" ], [ "0.8613", "0.662", "0.892", "0.662", "0.892", "0.675", "0.8613", "0.675" ], [ "0.0801", "0.672", "0.4697", "0.672", "0.4697", "0.686", "0.0801", "0.686" ], [ "0.0784", "0.686", "0.468", "0.6875", "0.468", "0.702", "0.0784", "0.7007" ], [ "0.5024", "0.681", "0.8936", "0.681", "0.8936", "0.6987", "0.5024", "0.6987" ], [ "0.5", "0.696", "0.8936", "0.6963", "0.8936", "0.714", "0.5", "0.7134" ], [ "0.0801", "0.712", "0.3962", "0.712", "0.3962", "0.7266", "0.0801", "0.7266" ], [ "0.572", "0.712", "0.826", "0.7134", "0.826", "0.731", "0.572", "0.73" ], [ "0.8613", "0.7153", "0.8906", "0.7153", "0.8906", "0.7266", "0.8613", "0.7266" ], [ "0.0989", "0.735", "0.468", "0.735", "0.468", "0.7485", "0.0989", "0.7485" ], [ "0.505", "0.735", "0.8896", "0.735", "0.8896", "0.7456", "0.505", "0.7456" ], [ "0.0801", "0.752", "0.468", "0.752", "0.468", "0.765", "0.0801", "0.765" ], [ "0.501", "0.7476", "0.892", "0.7485", "0.892", "0.763", "0.501", "0.762" ], [ "0.0801", "0.7666", "0.4673", "0.7666", "0.4673", "0.78", "0.0801", "0.78" ], [ "0.5024", "0.7646", "0.7837", "0.7646", "0.7837", "0.779", "0.5024", "0.779" ], [ "0.0809", "0.7827", "0.4673", "0.7827", "0.4673", "0.7935", "0.0809", "0.7935" ], [ "0.0801", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0801", "0.81" ], [ "0.5034", "0.7935", "0.6313", "0.7935", "0.6313", "0.808", "0.5034", "0.808" ], [ "0.0801", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0801", "0.8247" ], [ "0.5034", "0.8174", "0.689", "0.8174", "0.689", "0.832", "0.5034", "0.832" ], [ "0.0801", "0.827", "0.4697", "0.827", "0.4697", "0.8403", "0.0801", "0.8403" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.501", "0.8403", "0.892", "0.839", "0.892", "0.854", "0.501", "0.855" ], [ "0.0801", "0.857", "0.4697", "0.857", "0.4697", "0.8706", "0.0801", "0.8706" ], [ "0.501", "0.857", "0.892", "0.857", "0.892", "0.8706", "0.501", "0.8706" ], [ "0.0784", "0.872", "0.4697", "0.8706", "0.4697", "0.8853", "0.0784", "0.886" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0801", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ] ]

[ "Original Pseudo-labels", "Algorithm 1: The learning procedure of GPA-3D", "Ins.2", "Ins.3", "Ins.", "input : labeled source domain Ds, unlabeled target domain Dt,", "Original", "ns", "3D detector e, total epochs T, steps per epoch N, and", "query", "list ofupdate epochs U", "High-quality Pseudo-label Databas", "output: adapted 3D detector §t", "1 Pre-train the network s D according to Eq. (10);", "Random", "2Initialize0←0", "2", "rou", "", "Rep", "3 Generate pseudo-labels and database (Lt,Dt)+ (e,Dt)", "repi", "4 for epoch 1toT do", "forstep1toNdo", "Ins.2", "ins.3", "Sample mini-batches (βs, βt) (Ds, Dt):", "ns", "a.", "Group", "AugmentedPseudo-labels", "Ren", "Instance replacement BtugIRA(βt, Dt);", "Figure5.Illustration of instancereplacement augmentation (IRA)", "update (β°, Bug) according to Eq. (12);", "Left: IRA leverages a group mechanism to displace the original in-", "if epoch e U then", "stances with high-quality candidates. Right: Compared with ran-", "Update pseudo-labels Lt (e, Dt):", "dom replacing, our group mechanism does not interfere the spatial", " :", "context of the point cloud scene.", "i.e., As,j = Qij ± 1. The overall soft contrast loss Ccontra", "database.In this procedure, a parameter pira is adopted to", "can be formulated as:", "regulate the probability of the replacement operation.", "Lcontra=Cat+Catt+BiCtea+β2Ctether+B3Crer", "There are two main merits of IRA. First, the quantity of", "", "(9)", "target data is maintained and the diversity is also enhanced", "where βi, β2 and B3 are the balance coefficients.", "Second, benefiting from the group mechanism, the spatial", "3.5. Noise Sample Suppression", "contexts around thereplaced instances are unchanged and", "no ambiguous or unreasonable case is introduced, as de-", "The pseudo-labels used on the target domain are noisy", "vised in Fig. 5 (right).", "and can lead to the accumulation of errors. To mitigate", "the impact of noise, we propose the noise sample suppres-", "3.7. Overall Training Procedure", "sion (NsS) approach, which generates a noise mask to sup-", "The overall training procedure of GPA-3D is illustrated", "press the magnitude of the gradient descent for the fore-", "in Alg. 1. Following previous works [37, 38], the 3D de-", "ground areas that might be underlying noise.The noise", "tector is first trained on the labeled source domain Ds via", "mask can be represented as S e (a, 1.0)Hxw, where", "minimizing the detection loss Caet as:", "Q (α < 1.0) is the suppression factor to decrease the con-", "Caet = Creg + Cals,", "(10)", "tribution of low-quality samples. In S, the foreground areas", "that have high similarities with the background prototype,", "where the Creg and Cal, indicate the regression and classi-", "i.e., sim(F,gK+1) > 0.3, are assigned to Q, while rest", "fication errors respectively. Next, we use the pre-trained de-", "foreground and background areas are assigned to 1.0.", "tector to generate pseudo-labels L, and the database of IRA", "During training, the noise mask S is multiplied to the", "on the unlabeled target domain Dt. Finally, the co-training", "co-training loss Cco-train, elaborated in Sec. 3.7. With the", "paradigm is employed to further fine-tune the model as:", "progress of training, prototypes will be optimized with bet-", "Leo-train = Caet + Caet.", "(11)", "ter representative capability, which enables NsS to suppress", "the noise more reliably and facilitate the training procedure.", "where Cdet is the detection loss on target data, same as in", "Eq. 10. The overall adaptation loss Cadapt is calculated via:", "3.6. Instance Replacement Augmentation", "Cadapt = β. Lcontra + S. Ceo-train,", "(12)", "Those uncertain pseudo-labels (with scores of 0.2 ~ 0.5)", "where B is thetotal weight of the soft contrastloss.andS", "are usually ignored in training. Despite inaccurate, they", "is the noise mask of NSs. For more details of the training", "might provide partial localization information. To this end", "procedure, please refer to the supplements.", "we devise the instance replacement augmentation (IRA)", "module. As shown in Fig. 5 (left), we first pick the pseudo-", "4.Experiments", "labels with scores over 0.5 to construct a high-quality", "4.1. Experimental Setup", "database, which utilizes the group mechanism as Eq. 4 to", "divide the picked instances into groups belonging to differ-", "Datasets.We evaluate the GPA-3D on widely used au-", "ent geometric structures. During training, we calculate the", "tonomous driving benchmarks including Waymo [30],", "group indexes for the uncertain pseudo-labels, and replace", "nuScenes [2], and KITTI [6]. These datasets exhibit signif-", "them with instances having same group indexes from the", "icant diversities in foreground patterns and LiDAR beams," ]

{ "arxiv_id": "2308.08140", "language": "en", "page_idx": 4 }

[ [ "0.0801", "0.09155", "0.468", "0.09155", "0.468", "0.1061", "0.0801", "0.1061" ], [ "0.517", "0.09216", "0.892", "0.09216", "0.892", "0.1061", "0.517", "0.1061" ], [ "0.0809", "0.108", "0.468", "0.108", "0.468", "0.1218", "0.0809", "0.1218" ], [ "0.5156", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.5156", "0.1218" ], [ "0.0809", "0.1237", "0.468", "0.1237", "0.468", "0.137", "0.0809", "0.137" ], [ "0.517", "0.125", "0.8784", "0.125", "0.8784", "0.1351", "0.517", "0.1351" ], [ "0.0809", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0809", "0.1515" ], [ "0.5024", "0.1364", "0.8906", "0.137", "0.8906", "0.1515", "0.5024", "0.1509" ], [ "0.0809", "0.1528", "0.468", "0.1528", "0.468", "0.1666", "0.0809", "0.1666" ], [ "0.5156", "0.1528", "0.8906", "0.1528", "0.8906", "0.1666", "0.5156", "0.1666" ], [ "0.0809", "0.1692", "0.468", "0.1692", "0.468", "0.1799", "0.0809", "0.1799" ], [ "0.517", "0.1686", "0.892", "0.1686", "0.892", "0.1818", "0.517", "0.1818" ], [ "0.0809", "0.1818", "0.4673", "0.1818", "0.4673", "0.1957", "0.0809", "0.1957" ], [ "0.5156", "0.1831", "0.6685", "0.1831", "0.6685", "0.1963", "0.5156", "0.1963" ], [ "0.0801", "0.1976", "0.4673", "0.1976", "0.4673", "0.2108", "0.0801", "0.2108" ], [ "0.5024", "0.1976", "0.892", "0.1976", "0.892", "0.2108", "0.5024", "0.2108" ], [ "0.0801", "0.2134", "0.4697", "0.2134", "0.4697", "0.2267", "0.0801", "0.2267" ], [ "0.5156", "0.2134", "0.892", "0.2134", "0.892", "0.2267", "0.5156", "0.2267" ], [ "0.0801", "0.2291", "0.3096", "0.2291", "0.3096", "0.2424", "0.0801", "0.2424" ], [ "0.5156", "0.2279", "0.892", "0.2279", "0.892", "0.2412", "0.5156", "0.2412" ], [ "0.0989", "0.2456", "0.468", "0.2456", "0.468", "0.26", "0.0989", "0.26" ], [ "0.5156", "0.2437", "0.8906", "0.2424", "0.8906", "0.2568", "0.5156", "0.2583" ], [ "0.0784", "0.2615", "0.4673", "0.26", "0.4673", "0.2747", "0.0784", "0.276" ], [ "0.5156", "0.2588", "0.841", "0.2588", "0.841", "0.2727", "0.5156", "0.2727" ], [ "0.0801", "0.2766", "0.468", "0.2766", "0.468", "0.2903", "0.0801", "0.2903" ], [ "0.0801", "0.2917", "0.468", "0.2917", "0.468", "0.3062", "0.0801", "0.3062" ], [ "0.5034", "0.2874", "0.6406", "0.2874", "0.6406", "0.3018", "0.5034", "0.3018" ], [ "0.0809", "0.307", "0.468", "0.307", "0.468", "0.3208", "0.0809", "0.3208" ], [ "0.5024", "0.31", "0.892", "0.3125", "0.892", "0.327", "0.5024", "0.3245" ], [ "0.0801", "0.3213", "0.468", "0.3213", "0.468", "0.3352", "0.0801", "0.3352" ], [ "0.5024", "0.3276", "0.8906", "0.3276", "0.8906", "0.3416", "0.5024", "0.3416" ], [ "0.0809", "0.3384", "0.4673", "0.3384", "0.4673", "0.3484", "0.0809", "0.3484" ], [ "0.5034", "0.3435", "0.892", "0.3435", "0.892", "0.3567", "0.5034", "0.3567" ], [ "0.0801", "0.3516", "0.4673", "0.3516", "0.4673", "0.3655", "0.0801", "0.3655" ], [ "0.5024", "0.358", "0.892", "0.358", "0.892", "0.3718", "0.5024", "0.3718" ], [ "0.0801", "0.3662", "0.4697", "0.3662", "0.4697", "0.3801", "0.0801", "0.3801" ], [ "0.505", "0.375", "0.8906", "0.375", "0.8906", "0.385", "0.505", "0.385" ], [ "0.0801", "0.3833", "0.468", "0.3833", "0.468", "0.3965", "0.0801", "0.3965" ], [ "0.501", "0.387", "0.892", "0.3882", "0.892", "0.4028", "0.501", "0.4016" ], [ "0.0784", "0.3977", "0.4697", "0.3977", "0.4697", "0.411", "0.0784", "0.411" ], [ "0.5024", "0.4028", "0.892", "0.4016", "0.892", "0.4167", "0.5024", "0.4172" ], [ "0.0784", "0.411", "0.4697", "0.4124", "0.4697", "0.4268", "0.0784", "0.4255" ], [ "0.5024", "0.4185", "0.8906", "0.4185", "0.8906", "0.432", "0.5024", "0.432" ], [ "0.0809", "0.4287", "0.4673", "0.4287", "0.4673", "0.4395", "0.0809", "0.4395" ], [ "0.5024", "0.4343", "0.892", "0.4343", "0.892", "0.4475", "0.5024", "0.4475" ], [ "0.0784", "0.4426", "0.4697", "0.4414", "0.4697", "0.4558", "0.0784", "0.457" ], [ "0.5024", "0.449", "0.892", "0.449", "0.892", "0.4622", "0.5024", "0.4622" ], [ "0.0801", "0.4583", "0.468", "0.4583", "0.468", "0.4717", "0.0801", "0.4717" ], [ "0.5024", "0.4634", "0.8906", "0.4634", "0.8906", "0.4766", "0.5024", "0.4766" ], [ "0.0809", "0.4749", "0.4673", "0.4749", "0.4673", "0.4849", "0.0809", "0.4849" ], [ "0.5024", "0.4792", "0.892", "0.4792", "0.892", "0.4924", "0.5024", "0.4924" ], [ "0.0801", "0.488", "0.468", "0.488", "0.468", "0.502", "0.0801", "0.502" ], [ "0.5034", "0.4937", "0.8906", "0.4937", "0.8906", "0.507", "0.5034", "0.507" ], [ "0.0801", "0.5024", "0.468", "0.5024", "0.468", "0.5166", "0.0801", "0.5166" ], [ "0.5024", "0.509", "0.8906", "0.5083", "0.8906", "0.523", "0.5024", "0.5234" ], [ "0.0801", "0.518", "0.3635", "0.518", "0.3635", "0.532", "0.0801", "0.532" ], [ "0.5034", "0.5244", "0.892", "0.5244", "0.892", "0.5386", "0.5034", "0.5386" ], [ "0.0989", "0.535", "0.4673", "0.535", "0.4673", "0.549", "0.0989", "0.549" ], [ "0.5024", "0.5386", "0.892", "0.539", "0.892", "0.5537", "0.5024", "0.553" ], [ "0.0784", "0.55", "0.468", "0.5503", "0.468", "0.565", "0.0784", "0.5645" ], [ "0.505", "0.556", "0.8906", "0.556", "0.8906", "0.5664", "0.505", "0.5664" ], [ "0.0809", "0.5684", "0.4673", "0.5684", "0.4673", "0.579", "0.0809", "0.579" ], [ "0.5034", "0.5693", "0.892", "0.5693", "0.892", "0.5835", "0.5034", "0.5835" ], [ "0.0801", "0.582", "0.4673", "0.582", "0.4673", "0.595", "0.0801", "0.595" ], [ "0.5024", "0.5854", "0.8906", "0.5854", "0.8906", "0.5986", "0.5024", "0.5986" ], [ "0.0801", "0.5967", "0.4697", "0.5967", "0.4697", "0.61", "0.0801", "0.61" ], [ "0.5034", "0.6016", "0.8906", "0.6016", "0.8906", "0.6123", "0.5034", "0.6123" ], [ "0.0801", "0.6113", "0.4673", "0.61", "0.4673", "0.625", "0.0801", "0.6255" ], [ "0.5034", "0.616", "0.8896", "0.616", "0.8896", "0.627", "0.5034", "0.627" ], [ "0.0809", "0.627", "0.4673", "0.627", "0.4673", "0.64", "0.0809", "0.64" ], [ "0.5034", "0.631", "0.8906", "0.631", "0.8906", "0.6416", "0.5034", "0.6416" ], [ "0.0801", "0.6426", "0.468", "0.6426", "0.468", "0.656", "0.0801", "0.656" ], [ "0.5024", "0.6445", "0.892", "0.6445", "0.892", "0.6577", "0.5024", "0.6577" ], [ "0.0801", "0.657", "0.468", "0.657", "0.468", "0.6704", "0.0801", "0.6704" ], [ "0.5024", "0.66", "0.892", "0.66", "0.892", "0.674", "0.5024", "0.674" ], [ "0.0784", "0.672", "0.468", "0.6704", "0.468", "0.685", "0.0784", "0.686" ], [ "0.501", "0.674", "0.892", "0.675", "0.892", "0.6895", "0.501", "0.688" ], [ "0.0801", "0.686", "0.4673", "0.686", "0.4673", "0.6997", "0.0801", "0.6997" ], [ "0.5024", "0.69", "0.892", "0.69", "0.892", "0.705", "0.5024", "0.705" ], [ "0.0809", "0.702", "0.4673", "0.702", "0.4673", "0.7153", "0.0809", "0.7153" ], [ "0.5034", "0.7056", "0.892", "0.7056", "0.892", "0.7197", "0.5034", "0.7197" ], [ "0.0801", "0.7173", "0.468", "0.7173", "0.468", "0.731", "0.0801", "0.731" ], [ "0.5024", "0.72", "0.892", "0.72", "0.892", "0.7344", "0.5024", "0.7344" ], [ "0.0809", "0.732", "0.4673", "0.732", "0.4673", "0.7456", "0.0809", "0.7456" ], [ "0.5034", "0.7373", "0.8896", "0.7373", "0.8896", "0.7476", "0.5034", "0.7476" ], [ "0.0801", "0.7476", "0.468", "0.7476", "0.468", "0.761", "0.0801", "0.761" ], [ "0.5024", "0.752", "0.892", "0.752", "0.892", "0.765", "0.5024", "0.765" ], [ "0.0801", "0.762", "0.4673", "0.762", "0.4673", "0.776", "0.0801", "0.776" ], [ "0.5034", "0.7666", "0.892", "0.7666", "0.892", "0.78", "0.5034", "0.78" ], [ "0.0801", "0.778", "0.468", "0.778", "0.468", "0.7915", "0.0801", "0.7915" ], [ "0.5024", "0.781", "0.857", "0.781", "0.857", "0.7944", "0.5024", "0.7944" ], [ "0.0801", "0.7935", "0.4697", "0.7935", "0.4697", "0.8066", "0.0801", "0.8066" ], [ "0.5195", "0.7954", "0.892", "0.7944", "0.892", "0.8096", "0.5195", "0.81" ], [ "0.0801", "0.808", "0.468", "0.808", "0.468", "0.8213", "0.0801", "0.8213" ], [ "0.5024", "0.811", "0.892", "0.81", "0.892", "0.8237", "0.5024", "0.8247" ], [ "0.0801", "0.8237", "0.468", "0.8237", "0.468", "0.837", "0.0801", "0.837" ], [ "0.5024", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.5024", "0.8403" ], [ "0.0784", "0.836", "0.4697", "0.8384", "0.4697", "0.853", "0.0784", "0.851" ], [ "0.5024", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.5024", "0.855" ], [ "0.0809", "0.853", "0.468", "0.853", "0.468", "0.866", "0.0809", "0.866" ], [ "0.5034", "0.858", "0.8906", "0.858", "0.8906", "0.8687", "0.5034", "0.8687" ], [ "0.0801", "0.8687", "0.384", "0.8687", "0.384", "0.882", "0.0801", "0.882" ], [ "0.501", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.501", "0.8853" ], [ "0.0801", "0.886", "0.468", "0.886", "0.468", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5024", "0.901" ] ]

[ "every pair of AUs in the spatial domain(e.g., via transformer", "tively/specifically considers facial dynamic correspond-", "encoder with the self-attention operation [13] and graph", "ing to each AU at each spatial scale, as each AUs' activa-", "edges learned by the same cross-attention operation [29],", "tionexhibitheterogeneityinbothrangeandmagnitude.", "without giving explicit consideration to the natural hierar-", ":The proposed HSR is the first module that hierarchically", "chical relationship among AUs (Problem 1). More specif-", "learns local and cross-regional spatio-temporal relation-", "ically,AUscorrespondingtothesame/closefacialregions", "ship, while previous approaches fail to consider such hi-", "frequently show stronger associations than AUs located in", "erarchical relationship.", "different facial regions(illustrated in Fig. 1 (a)), as AUs lo-", ": Experimental results show that our MDHR is the new", "calized to the same/close facial regions may be influenced", "state-of-the-art on the widely-used AU recognition bench-", "by some shared facial muscles [1].", "mark datasets: BP4D [59] and DISFA [31], where the", "Besides the spatial cues, some studies additionally model", "proposed MFD and HSR modules positively and comple-", "temporal dynamic between facial frames to enhance AU", "mentarily contributed to this decent performance.", "recognition performances [23, 24, 47]. A typical solution", "is applying common temporal models (e.g., Long-Short-", "2.Related Work", "Term-Memory (LSTM) [3, 18], Spatio-Temporal Graph", "Static face image-based methods: Existing approaches", "[37, 50] and 3D Convolution Neural Networks (CNNs) [5])", "frequently predict AUs' status based on static facial dis-", "to process the extractedframe-level static facial/AU fea-", "plays. Given the anatomical definition of AUs, many of", "tures. However, these temporal modeling strategies are in-", "them [6, 13, 14, 18, 19, 42] attempted to recognize each", "sensitive to subtle facial muscle movements. While other", "AUbased onafacepatchdefinedby automaticallyde-", "approaches [12, 43, 56] (e.g., optical flow and dynamic", "tected facial landmarks or other prior settings.For ex-", "image) can explicitly capture facial motions, they still fail", "ample, Zhao et al. [61] proposed a patch-based DRML", "to consider that facial muscle movements corresponding", "that learns AU representations robust to variations inher-", "to different AUs'activation could exhibit heterogeneity in", "ent within local facial regions. EAC-Net [19] proposed a", "both range and magnitude (Problem 2), e.g., AU25 involves", "cropping layer to learn individual AU's representation from", "large-scale deformations of the mouth region, while AU2", "small AU-specific areas. Furthermore, JAA-Net [38] jointly", "arerepresented by subtlemusclemovements surrounding", "conducted AU recognition and face alignment, where the", "eyebrows (illustrated in Fig. 1 (b)). In other words, fa-", "predicted facial landmarks are used to localize each AU re-", "cial dynamic of a certain spatial scale could contribute un-", "gion. To take global facial contextual cues into considera-", "equally to the recognition of different AUs.", "tion, alternative approaches [23, 29, 36, 39, 40] learn each", "In this paper, we propose a novel Multi-scale Dynamic", "AU's representation holistically from the full face image,", "and Hierarchical Relationship (MDHR)modeling approach", "where spatial attention mechanismshave been widely ex-", "forAUrecognition,which:(i)hierarchicallymodels spatio-", "plored. Shao et al. [36] employed adaptive channel-wise", "temporal relationship among AUs; and (i) adaptively con-", "and spatial attention strategy to enforce the model focus-", "siders facial dynamic at various spatial scales for each AU's", "ing on AU-relatedlocalfeaturesfromthe globalface.:Li", "recognition.Our MDHR consists of two key modules.", "etal.[20l proposed a self-diversified multi-channel atten-", "The Multi-scale Facial Dynamic Modelling (MFD) mod-", "tion to seek a more robust attention between the global fa-", "ule that adaptively emphasizes AU-related facial dynamic", "cial representation and each target AU. As AUs are mu-", "at multiple spatial scales (i.e., computing differences be-", "tually related [60], recent approaches [2, 13, 32, 57] also", "tween neighboring frames' features maps output from dif-", "specifically modelled the underlying relationship among", "ferent backbone layers), ensuring both obvious and sub-", "them. For exampple, LP-Net [32] applied LSTMs to capture", "tle AU-related facial dynamic can be captured in an eff-", "AU relationship. Jacob et al.[13] proposed a transformer-", "cient manner (addressing Problem 2). Then, a Hierarchi-", "style AU correlation network. In addition, graph-based", "cal Spatio-temporal AU Relationship Modelling (HSR)", "strategieshave been frequently investigated to model AU", "module is introduced to hierarchically model relationship", "relationship[17, 26, 29, 46], where graph nodes have been", "among spatio-temporal AU features in a two-stage man-", "frequently used to represent target AUs while edges explic-", "ner, where the first stage individually models relationship", "itly define the relationship between every pair of AUs.", "among AUs within the same/close facial region at both", "Spatial-temporal methods: Since facial dynamic also", "feature extraction and AU prediction levels, and the sec-", "provide crucial cues forAU recognition[4],LSTM hasbeen", "ond stage explicitly learns the relationship between pairs", "frequently employed by early studies [4, 15] to model tem-", "of AUs located in different facial regions via graph edges", "poral dynamic between static facial features extracted from", "(addressing Problem 1). The main contributions and nov-", "adjacent frames.Tofurther explore spatio-temporal rela-", "elties of this paper are summarised as follows:", "tionship among AUs, a Spatio-temporal Graph Neural Net-", "· The proposed MFD is the first module that adap-", "work (GNN) [37] and a Heterogeneous Spatio-temporal Re-" ]

{ "arxiv_id": "2404.06443", "language": "en", "page_idx": 1 }

[ [ "0.0809", "0.09155", "0.468", "0.09155", "0.468", "0.1061", "0.0809", "0.1061" ], [ "0.5024", "0.089", "0.892", "0.089", "0.892", "0.1067", "0.5024", "0.1067" ], [ "0.0801", "0.108", "0.4697", "0.108", "0.4697", "0.1218", "0.0801", "0.1218" ], [ "0.5034", "0.1092", "0.6284", "0.1092", "0.6284", "0.1193", "0.5034", "0.1193" ], [ "0.0809", "0.125", "0.468", "0.125", "0.468", "0.1351", "0.0809", "0.1351" ], [ "0.0784", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0784", "0.1515" ], [ "0.5034", "0.1364", "0.8823", "0.1364", "0.8823", "0.1497", "0.5034", "0.1497" ], [ "0.0809", "0.1547", "0.468", "0.1547", "0.468", "0.1654", "0.0809", "0.1654" ], [ "0.5024", "0.154", "0.8716", "0.154", "0.8716", "0.1674", "0.5024", "0.1674" ], [ "0.0801", "0.1674", "0.468", "0.1674", "0.468", "0.1818", "0.0801", "0.1818" ], [ "0.5034", "0.1692", "0.8037", "0.1692", "0.8037", "0.1843", "0.5034", "0.1843" ], [ "0.0801", "0.1831", "0.468", "0.1831", "0.468", "0.1963", "0.0801", "0.1963" ], [ "0.513", "0.1875", "0.652", "0.1875", "0.652", "0.2008", "0.513", "0.2008" ], [ "0.0801", "0.1976", "0.4697", "0.1976", "0.4697", "0.2108", "0.0801", "0.2108" ], [ "0.513", "0.2058", "0.528", "0.2058", "0.528", "0.214", "0.513", "0.214" ], [ "0.5474", "0.2026", "0.8906", "0.2026", "0.8906", "0.2166", "0.5474", "0.2166" ], [ "0.0801", "0.214", "0.4697", "0.214", "0.4697", "0.2247", "0.0801", "0.2247" ], [ "0.5483", "0.2153", "0.8228", "0.2153", "0.8228", "0.2329", "0.5483", "0.2329" ], [ "0.814", "0.2184", "0.888", "0.2153", "0.888", "0.2267", "0.814", "0.2299" ], [ "0.0801", "0.2279", "0.468", "0.2279", "0.468", "0.2412", "0.0801", "0.2412" ], [ "0.5146", "0.238", "0.523", "0.238", "0.523", "0.2437", "0.5146", "0.2437" ], [ "0.5474", "0.2329", "0.8716", "0.2329", "0.8716", "0.2468", "0.5474", "0.2468" ], [ "0.0801", "0.2437", "0.4673", "0.2437", "0.4673", "0.2568", "0.0801", "0.2568" ], [ "0.546", "0.2468", "0.694", "0.2474", "0.694", "0.262", "0.546", "0.2615" ], [ "0.0809", "0.26", "0.4673", "0.26", "0.4673", "0.2708", "0.0809", "0.2708" ], [ "0.544", "0.262", "0.6235", "0.2644", "0.6235", "0.279", "0.544", "0.2766" ], [ "0.6626", "0.2644", "0.7466", "0.2644", "0.7466", "0.276", "0.6626", "0.276" ], [ "0.7817", "0.2632", "0.892", "0.2632", "0.892", "0.2778", "0.7817", "0.2778" ], [ "0.0801", "0.2734", "0.4697", "0.2734", "0.4697", "0.2874", "0.0801", "0.2874" ], [ "0.517", "0.2651", "0.524", "0.2708", "0.517", "0.276", "0.5107", "0.2708" ], [ "0.546", "0.276", "0.863", "0.276", "0.863", "0.2935", "0.546", "0.2935" ], [ "0.0801", "0.2886", "0.468", "0.2886", "0.468", "0.3018", "0.0801", "0.3018" ], [ "0.511", "0.2942", "0.5825", "0.2942", "0.5825", "0.3062", "0.511", "0.3062" ], [ "0.0801", "0.3037", "0.468", "0.3037", "0.468", "0.3176", "0.0801", "0.3176" ], [ "0.509", "0.308", "0.8364", "0.3062", "0.8364", "0.3208", "0.509", "0.3225" ], [ "0.834", "0.31", "0.8667", "0.31", "0.8667", "0.3176", "0.834", "0.3176" ], [ "0.0809", "0.3208", "0.4673", "0.3208", "0.4673", "0.3308", "0.0809", "0.3308" ], [ "0.5312", "0.3213", "0.751", "0.3213", "0.751", "0.336", "0.5312", "0.336" ], [ "0.0809", "0.3352", "0.4656", "0.3352", "0.4656", "0.3452", "0.0809", "0.3452" ], [ "0.0809", "0.351", "0.4673", "0.351", "0.4673", "0.361", "0.0809", "0.361" ], [ "0.5024", "0.358", "0.819", "0.3599", "0.819", "0.375", "0.5024", "0.3726" ], [ "0.0801", "0.3643", "0.4697", "0.3643", "0.4697", "0.3774", "0.0801", "0.3774" ], [ "0.0801", "0.379", "0.468", "0.379", "0.468", "0.3926", "0.0801", "0.3926" ], [ "0.5024", "0.3833", "0.8906", "0.3833", "0.8906", "0.3965", "0.5024", "0.3965" ], [ "0.0825", "0.3958", "0.4673", "0.3958", "0.4673", "0.406", "0.0825", "0.406" ], [ "0.5034", "0.3984", "0.892", "0.3984", "0.892", "0.4124", "0.5034", "0.4124" ], [ "0.0801", "0.4104", "0.3513", "0.4104", "0.3513", "0.4236", "0.0801", "0.4236" ], [ "0.501", "0.414", "0.8906", "0.414", "0.8906", "0.428", "0.501", "0.428" ], [ "0.5024", "0.428", "0.8906", "0.4287", "0.8906", "0.443", "0.5024", "0.4426" ], [ "0.0784", "0.4343", "0.21", "0.4375", "0.21", "0.455", "0.0784", "0.452" ], [ "0.5034", "0.4443", "0.8906", "0.4443", "0.8906", "0.4583", "0.5034", "0.4583" ], [ "0.0809", "0.4634", "0.4436", "0.4634", "0.4436", "0.4778", "0.0809", "0.4778" ], [ "0.501", "0.4583", "0.8906", "0.459", "0.8906", "0.4734", "0.501", "0.473" ], [ "0.0768", "0.4734", "0.4697", "0.473", "0.4697", "0.4944", "0.0768", "0.4956" ], [ "0.5034", "0.476", "0.8906", "0.476", "0.8906", "0.486", "0.5034", "0.486" ], [ "0.0801", "0.4937", "0.4673", "0.4937", "0.4673", "0.5083", "0.0801", "0.5083" ], [ "0.5024", "0.4893", "0.8936", "0.4893", "0.8936", "0.5024", "0.5024", "0.5024" ], [ "0.0801", "0.509", "0.468", "0.509", "0.468", "0.523", "0.0801", "0.523" ], [ "0.5024", "0.504", "0.892", "0.504", "0.892", "0.5176", "0.5024", "0.5176" ], [ "0.0801", "0.5234", "0.468", "0.5234", "0.468", "0.537", "0.0801", "0.537" ], [ "0.5034", "0.521", "0.8906", "0.521", "0.8906", "0.531", "0.5034", "0.531" ], [ "0.0801", "0.539", "0.468", "0.539", "0.468", "0.553", "0.0801", "0.553" ], [ "0.5024", "0.534", "0.892", "0.534", "0.892", "0.5474", "0.5024", "0.5474" ], [ "0.0801", "0.553", "0.4673", "0.553", "0.4673", "0.5674", "0.0801", "0.5674" ], [ "0.5024", "0.55", "0.892", "0.55", "0.892", "0.563", "0.5024", "0.563" ], [ "0.0801", "0.5693", "0.4697", "0.5693", "0.4697", "0.584", "0.0801", "0.584" ], [ "0.5024", "0.565", "0.653", "0.5645", "0.653", "0.5776", "0.5024", "0.579" ], [ "0.0801", "0.5806", "0.468", "0.5806", "0.468", "0.5986", "0.0801", "0.5986" ], [ "0.5195", "0.579", "0.892", "0.5776", "0.892", "0.5923", "0.5195", "0.5933" ], [ "0.0801", "0.5996", "0.4673", "0.5996", "0.4673", "0.613", "0.0801", "0.613" ], [ "0.5024", "0.5947", "0.892", "0.5947", "0.892", "0.6094", "0.5024", "0.6094" ], [ "0.0784", "0.6157", "0.4673", "0.6157", "0.4673", "0.629", "0.0784", "0.629" ], [ "0.5034", "0.61", "0.8906", "0.61", "0.8906", "0.6235", "0.5034", "0.6235" ], [ "0.505", "0.627", "0.5728", "0.627", "0.5728", "0.637", "0.505", "0.637" ], [ "0.5884", "0.625", "0.892", "0.625", "0.892", "0.6396", "0.5884", "0.6396" ], [ "0.5024", "0.64", "0.892", "0.64", "0.892", "0.654", "0.5024", "0.654" ], [ "0.0801", "0.63", "0.468", "0.63", "0.468", "0.643", "0.0801", "0.643" ], [ "0.0768", "0.643", "0.4697", "0.6416", "0.4697", "0.659", "0.0768", "0.661" ], [ "0.501", "0.6514", "0.8936", "0.6514", "0.8936", "0.6694", "0.501", "0.6694" ], [ "0.0784", "0.659", "0.468", "0.6577", "0.468", "0.6724", "0.0784", "0.674" ], [ "0.0801", "0.675", "0.4697", "0.675", "0.4697", "0.688", "0.0801", "0.688" ], [ "0.501", "0.6694", "0.892", "0.6685", "0.892", "0.683", "0.501", "0.6846" ], [ "0.0801", "0.69", "0.4697", "0.69", "0.4697", "0.704", "0.0801", "0.704" ], [ "0.5024", "0.685", "0.8906", "0.685", "0.8906", "0.6987", "0.5024", "0.6987" ], [ "0.0801", "0.705", "0.468", "0.705", "0.468", "0.7183", "0.0801", "0.7183" ], [ "0.5024", "0.7007", "0.8896", "0.7007", "0.8896", "0.714", "0.5024", "0.714" ], [ "0.0801", "0.72", "0.4697", "0.72", "0.4697", "0.7344", "0.0801", "0.7344" ], [ "0.5034", "0.7153", "0.892", "0.7153", "0.892", "0.729", "0.5034", "0.729" ], [ "0.0801", "0.735", "0.4697", "0.735", "0.4697", "0.7485", "0.0801", "0.7485" ], [ "0.501", "0.731", "0.892", "0.731", "0.892", "0.744", "0.501", "0.744" ], [ "0.0801", "0.75", "0.468", "0.7505", "0.468", "0.765", "0.0801", "0.7646" ], [ "0.5024", "0.7456", "0.892", "0.7456", "0.892", "0.7603", "0.5024", "0.7603" ], [ "0.0809", "0.7666", "0.4673", "0.7666", "0.4673", "0.78", "0.0809", "0.78" ], [ "0.5024", "0.761", "0.818", "0.761", "0.818", "0.7744", "0.5024", "0.7744" ], [ "0.0801", "0.781", "0.468", "0.781", "0.468", "0.7944", "0.0801", "0.7944" ], [ "0.642", "0.7827", "0.751", "0.781", "0.751", "0.8003", "0.642", "0.803" ], [ "0.871", "0.784", "0.892", "0.787", "0.8896", "0.802", "0.8677", "0.799" ], [ "0.0801", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0801", "0.81" ], [ "0.0809", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.0809", "0.8247" ], [ "0.501", "0.806", "0.8936", "0.806", "0.8936", "0.827", "0.501", "0.827" ], [ "0.0809", "0.8276", "0.4673", "0.8276", "0.4673", "0.8384", "0.0809", "0.8384" ], [ "0.5", "0.8228", "0.892", "0.8237", "0.892", "0.8413", "0.5", "0.8403" ], [ "0.0801", "0.839", "0.468", "0.8403", "0.468", "0.855", "0.0801", "0.854" ], [ "0.5024", "0.839", "0.892", "0.8403", "0.892", "0.855", "0.5024", "0.854" ], [ "0.0801", "0.856", "0.468", "0.856", "0.468", "0.869", "0.0801", "0.869" ], [ "0.5", "0.855", "0.7363", "0.8516", "0.7363", "0.869", "0.5", "0.8726" ], [ "0.728", "0.856", "0.8896", "0.856", "0.8896", "0.869", "0.728", "0.869" ], [ "0.0784", "0.869", "0.468", "0.872", "0.468", "0.886", "0.0784", "0.884" ], [ "0.501", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.501", "0.8853" ], [ "0.0784", "0.886", "0.4673", "0.8853", "0.4673", "0.8994", "0.0784", "0.901" ], [ "0.5034", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5034", "0.901" ] ]

[ "lation Learning Network (HSTR-Net) [47] have been pro-", "put T frames as P1, P2,..., pT (Pt = (pt,... , pN], line", "posed, both of which first construct a set of spatial graphs", "6 in Algorithm 1).", "tomodel static AU relationshipattheframe-level,and then", "individually model each AU's temporal dynamic by consid-", "Algorithm 1 Pipeline of the proposed approach (MDHR)", "eringits corresponding spatial graphnodefeatures across", "Input : T consecutive facial frames S = ffl, ... fT", "all frames. In addition, Li et al. [21] applied a transformer", "Output: N AU's predictions of each frame ft", "to learn both spatial AU dependencies and temporal inter-", "1: for t = 1 to T do", "frame contexts by representing the inter-AU and inter-frame", "2:", "Generating multi-scale static global representations", "correlations within amulti-head attention matrix.Besides", "Xi, X2,... XL - Backbone(ft-k,..., f", "ft... ft+k", "such standard temporal model-based feature-level dynamic", "3", "Generating global spatio-temporal features Gt", "modelling,other solutions[12,22,43,53,54] also have", "MFD(X1, X2 ...XL)", "beeninvestigated.For example,two auxiliary AU related", "Generating", "hierarchical", " spatio-temporal", "tasks (e.g., ROI inpainting and optical flow estimation) are", "X", "relationship-aware AU features Vt ,- HSR(Gt)", "jointly conducted in [52] to enhance the regional features", "5: end for", "and encode the facial dynamic into the global facial rep-", "6: Predicting N AUs of all frames Pl, P2....", "PT", "resentation.respectively.Morerecently,Yang etal.[56]", "SC(TCN(VI.... , Vt,..., VT))", "introducedatemporaldifferencenetwork(TDN)thatex-", "tractfacial dynamic at a specific spatial scale.Despite the", "3.1. Multi-scale facial dynamic modelling", "progress made by approaches discussed above, to the best", "of our knowledge, none of them has specifically modelled", "Inspired by the fact that facial muscle movements are con-", "AU-related multi-scalefacial dynamic and thehierarchical", "tinuous and smooth while each AU exhibit heterogeneity in", "spatio-temporal relationship among AUs.", "their range of motions and magnitudes [1], we propose a", "novel MFD module to model the preceding and proceed-", "3. Methodology", "ing temporal evolution of the target face at multiple spa-", "Overview: Given T consecutive facial frames S", "tial scales. It includes a multi-scale Temporal Differecing", "(f\", f2,,fT) e IRTxCxHxW,our approach jointly", "blockthatfirstcomputesdifferencesbetweenglobalfacial", "predicts multiple (N) AUs occurrence at the th facial", "features extracted from every pair of neighboring frames at", "frame ft (t = 1, 2, .- , T) by taking not only the ft but", "multiple spatial scales. The obtained multi-scale facial dy-", "also its adjacent frames into consideration. As illustrated", "namic features are then masked by a set of weighting ma-", "in Fig. 2 and Algorithm 1, our MDHR starts with utiliz-", "trices learned by our adaptive weighting block, aiming to", "ing a backbone (e.g., CNN or Transformer) to jointly ex-", "emphasize the informative cues for target AUs at multiple", "tract static facial features from ft and its adjacent frames", "spatio-temporal scales.", "At - {ft-, ..., ft-1, ft+1,..\", ft+k]. For each frame,", "Multi-scale Temporal Differecing block: This block", "multi-scale static facial features are produced by L-1 back-", "is made up of multiple Temporal Differecing (TD) layers", "bone hidden layers and the output layer (the Lth layer).", "followed by convolution layers, which takes feature maps", "X=ft", ".,rt,. ,rt+hl = 1, 2, ..,L] produced", "by multiple (L -- 1) hidden layers and the output layer of the", "Thus, L static facial feature sets corresponding to 2k + 1", "frames Xi = {rt-h,..., at,... ,at+kj = i, 2,.., L]", "backbone as the input, where ct e RCi x Hi xWi denotes the", "are generated (line 2 in Algorithm 1). Then, these multi-", "scale features are fed to the Multi-scale Facial Dynamic", "feature map corresponding to the th facial frame generated", "Modelling (MFD) module, targeting at not only explic-", "from the lth backbone hidden layer (i.e., Ci, Hi, and Wi rep-", "itly capturing facial dynamic at multiple spatial scales, but", "resent the channel, height and width of the rt, respectively)", "also adaptively combining these multi-scale facial dynamic", "Here, the lth TD layer conducts point-to-point subtraction", "features with the static feature ct. (line 3 in Algorithm", "on feature maps produced by the lth hidden layer between", "1). Based on the spatio-temporal full face representation", "neighboring frames, aiming to capture facial dynamic at a", "Gt learned by MFD, a Hierarchical Spatio-temporal AU", "certain spatial scale. This can be formulated as:", "Relationship Modelling (HSR) module further adaptively", "di = rt - at-1", "(1)", "models the hierarchical spatio-temporal relationship among", "AUs in a two-stage manner, where the spatial distribution of", "Thus, a dynamic feature map d, RCxHixWi representing", "thetarget AUs on the human face is considered, resulting in", "the facial dynamic between ft and ft-1 at the ln spatial", "N individual AU representations Vt ={ot,..., ot} (line", "scale are produced from the l TD layer. As a result, L sets", "4 in Algorithm 1). Finally, a Temporal Convolution Net-", "of dynamic features D, = {dt-k+i", "-I,. ,dt,+. ,dt+kll -", "works (TCN) [27] with similarity calculating (SC) strategy", "1, 2, ., L} are obtained to represent facial dynamic at L", "[29] are employed to predict N AUs' occurrences of the in-", "different scales. After that, we introduce L step convolution" ]

{ "arxiv_id": "2404.06443", "language": "en", "page_idx": 2 }

[ [ "0.0801", "0.0903", "0.468", "0.09155", "0.468", "0.1061", "0.0801", "0.1048" ], [ "0.501", "0.09216", "0.8", "0.09216", "0.8", "0.1061", "0.501", "0.1061" ], [ "0.1209", "0.1092", "0.1978", "0.1092", "0.1978", "0.1206", "0.1209", "0.1206" ], [ "0.597", "0.113", "0.796", "0.113", "0.796", "0.1307", "0.597", "0.1307" ], [ "0.869", "0.11554", "0.892", "0.11554", "0.892", "0.1288", "0.869", "0.1288" ], [ "0.0809", "0.132", "0.468", "0.132", "0.468", "0.1459", "0.0809", "0.1459" ], [ "0.505", "0.137", "0.892", "0.137", "0.892", "0.1509", "0.505", "0.1509" ], [ "0.0801", "0.1465", "0.468", "0.1465", "0.468", "0.1604", "0.0801", "0.1604" ], [ "0.5034", "0.1528", "0.8906", "0.1528", "0.8906", "0.1666", "0.5034", "0.1666" ], [ "0.0801", "0.1622", "0.4697", "0.161", "0.4697", "0.1755", "0.0801", "0.1768" ], [ "0.5034", "0.1686", "0.892", "0.1686", "0.892", "0.1818", "0.5034", "0.1818" ], [ "0.0801", "0.178", "0.468", "0.178", "0.468", "0.1913", "0.0801", "0.1913" ], [ "0.5024", "0.1831", "0.892", "0.1831", "0.892", "0.1963", "0.5024", "0.1963" ], [ "0.0801", "0.1925", "0.4697", "0.1925", "0.4697", "0.2058", "0.0801", "0.2058" ], [ "0.5024", "0.1989", "0.859", "0.1989", "0.859", "0.2122", "0.5024", "0.2122" ], [ "0.0801", "0.207", "0.4697", "0.207", "0.4697", "0.2203", "0.0801", "0.2203" ], [ "0.523", "0.2134", "0.892", "0.2134", "0.892", "0.2267", "0.523", "0.2267" ], [ "0.0801", "0.2229", "0.468", "0.2229", "0.468", "0.2361", "0.0801", "0.2361" ], [ "0.5024", "0.2291", "0.892", "0.2291", "0.892", "0.2424", "0.5024", "0.2424" ], [ "0.0801", "0.2374", "0.468", "0.2374", "0.468", "0.2507", "0.0801", "0.2507" ], [ "0.5034", "0.2437", "0.8906", "0.2437", "0.8906", "0.2568", "0.5034", "0.2568" ], [ "0.0809", "0.2524", "0.468", "0.2524", "0.468", "0.2664", "0.0809", "0.2664" ], [ "0.5024", "0.2588", "0.8906", "0.2588", "0.8906", "0.2727", "0.5024", "0.2727" ], [ "0.0801", "0.2683", "0.3137", "0.2683", "0.3137", "0.2822", "0.0801", "0.2822" ], [ "0.5024", "0.2734", "0.8906", "0.2734", "0.8906", "0.2874", "0.5024", "0.2874" ], [ "0.0989", "0.2827", "0.468", "0.2842", "0.468", "0.2986", "0.0989", "0.2974" ], [ "0.5034", "0.289", "0.8906", "0.289", "0.8906", "0.303", "0.5034", "0.303" ], [ "0.0809", "0.2998", "0.4697", "0.2998", "0.4697", "0.3132", "0.0809", "0.3132" ], [ "0.5024", "0.303", "0.8906", "0.3037", "0.8906", "0.318", "0.5024", "0.3176" ], [ "0.0801", "0.315", "0.468", "0.315", "0.468", "0.3289", "0.0801", "0.3289" ], [ "0.5034", "0.318", "0.892", "0.318", "0.892", "0.332", "0.5034", "0.332" ], [ "0.0784", "0.3289", "0.468", "0.33", "0.468", "0.3447", "0.0784", "0.3435" ], [ "0.5034", "0.334", "0.888", "0.334", "0.888", "0.348", "0.5034", "0.348" ], [ "0.0801", "0.3452", "0.468", "0.3452", "0.468", "0.3591", "0.0801", "0.3591" ], [ "0.501", "0.3499", "0.8906", "0.3484", "0.8906", "0.3623", "0.501", "0.363" ], [ "0.0784", "0.361", "0.4697", "0.3599", "0.4697", "0.3738", "0.0784", "0.375" ], [ "0.5024", "0.3643", "0.8906", "0.3643", "0.8906", "0.3774", "0.5024", "0.3774" ], [ "0.0801", "0.3757", "0.4714", "0.3757", "0.4714", "0.3894", "0.0801", "0.3894" ], [ "0.5024", "0.3801", "0.892", "0.3801", "0.892", "0.3933", "0.5024", "0.3933" ], [ "0.0801", "0.3894", "0.468", "0.3894", "0.468", "0.404", "0.0801", "0.404" ], [ "0.5034", "0.3945", "0.892", "0.3945", "0.892", "0.4077", "0.5034", "0.4077" ], [ "0.0801", "0.406", "0.468", "0.406", "0.468", "0.4192", "0.0801", "0.4192" ], [ "0.5024", "0.4104", "0.888", "0.4104", "0.888", "0.4236", "0.5024", "0.4236" ], [ "0.0784", "0.4192", "0.468", "0.4204", "0.468", "0.435", "0.0784", "0.4343" ], [ "0.5034", "0.4248", "0.8906", "0.4248", "0.8906", "0.4382", "0.5034", "0.4382" ], [ "0.0809", "0.4363", "0.468", "0.4363", "0.468", "0.4495", "0.0809", "0.4495" ], [ "0.501", "0.4382", "0.8813", "0.4395", "0.8813", "0.4539", "0.501", "0.4526" ], [ "0.0801", "0.452", "0.4673", "0.4495", "0.4673", "0.4634", "0.0801", "0.4653" ], [ "0.5557", "0.459", "0.835", "0.4583", "0.835", "0.4797", "0.5557", "0.481" ], [ "0.0825", "0.4666", "0.4673", "0.4666", "0.4673", "0.4797", "0.0825", "0.4797" ], [ "0.0825", "0.481", "0.4697", "0.481", "0.4697", "0.4944", "0.0825", "0.4944" ], [ "0.0801", "0.4968", "0.2197", "0.4968", "0.2197", "0.5103", "0.0801", "0.5103" ], [ "0.564", "0.5005", "0.5703", "0.5005", "0.5703", "0.5083", "0.564", "0.5083" ], [ "0.6445", "0.4956", "0.7363", "0.4944", "0.7363", "0.513", "0.6445", "0.5146" ], [ "0.58", "0.504", "0.603", "0.504", "0.603", "0.507", "0.58", "0.507" ], [ "0.8677", "0.507", "0.8945", "0.507", "0.8945", "0.5234", "0.8677", "0.5234" ], [ "0.626", "0.5166", "0.665", "0.5166", "0.665", "0.528", "0.626", "0.528" ], [ "0.721", "0.532", "0.8013", "0.532", "0.8013", "0.5503", "0.721", "0.5503" ], [ "0.4143", "0.547", "0.4558", "0.547", "0.4558", "0.561", "0.4143", "0.561" ], [ "0.558", "0.5425", "0.6543", "0.5425", "0.6543", "0.5566", "0.558", "0.5566" ], [ "0.1896", "0.5474", "0.2632", "0.5474", "0.2632", "0.5596", "0.1896", "0.5596" ], [ "0.09314", "0.5474", "0.1969", "0.5474", "0.1969", "0.5625", "0.09314", "0.5625" ], [ "0.273", "0.5474", "0.3162", "0.5474", "0.3162", "0.5596", "0.273", "0.5596" ], [ "0.3245", "0.553", "0.3472", "0.553", "0.3472", "0.5566", "0.3245", "0.5566" ], [ "0.357", "0.5503", "0.392", "0.5503", "0.392", "0.558", "0.357", "0.558" ], [ "0.6562", "0.553", "0.6724", "0.553", "0.6724", "0.556", "0.6562", "0.556" ], [ "0.699", "0.553", "0.826", "0.5503", "0.826", "0.5684", "0.699", "0.571" ], [ "0.4468", "0.5674", "0.4673", "0.5674", "0.4673", "0.58", "0.4468", "0.58" ], [ "0.0801", "0.58", "0.468", "0.5806", "0.468", "0.595", "0.0801", "0.5947" ], [ "0.5024", "0.577", "0.892", "0.577", "0.892", "0.5947", "0.5024", "0.5947" ], [ "0.0801", "0.595", "0.468", "0.595", "0.468", "0.6094", "0.0801", "0.6094" ], [ "0.5024", "0.5933", "0.888", "0.5923", "0.888", "0.607", "0.5024", "0.608" ], [ "0.0801", "0.6113", "0.4697", "0.6113", "0.4697", "0.625", "0.0801", "0.625" ], [ "0.5024", "0.61", "0.8906", "0.61", "0.8906", "0.6235", "0.5024", "0.6235" ], [ "0.0784", "0.6255", "0.468", "0.625", "0.468", "0.6396", "0.0784", "0.64" ], [ "0.5034", "0.627", "0.8896", "0.627", "0.8896", "0.637", "0.5034", "0.637" ], [ "0.0801", "0.6416", "0.468", "0.6416", "0.468", "0.655", "0.0801", "0.655" ], [ "0.5024", "0.6396", "0.8906", "0.6396", "0.8906", "0.653", "0.5024", "0.653" ], [ "0.0801", "0.657", "0.468", "0.657", "0.468", "0.6704", "0.0801", "0.6704" ], [ "0.5024", "0.655", "0.8906", "0.655", "0.8906", "0.6694", "0.5024", "0.6694" ], [ "0.0784", "0.672", "0.4697", "0.6704", "0.4697", "0.685", "0.0784", "0.686" ], [ "0.5024", "0.6694", "0.892", "0.6694", "0.892", "0.683", "0.5024", "0.683" ], [ "0.0768", "0.683", "0.4697", "0.686", "0.4697", "0.704", "0.0768", "0.7007" ], [ "0.501", "0.6816", "0.8936", "0.683", "0.8936", "0.7007", "0.501", "0.6997" ], [ "0.0801", "0.7026", "0.4673", "0.7026", "0.4673", "0.7163", "0.0801", "0.7163" ], [ "0.5034", "0.7007", "0.8423", "0.7007", "0.8423", "0.714", "0.5034", "0.714" ], [ "0.0809", "0.7173", "0.4697", "0.7173", "0.4697", "0.731", "0.0809", "0.731" ], [ "0.501", "0.7217", "0.6396", "0.7227", "0.6396", "0.7373", "0.501", "0.7363" ], [ "0.0801", "0.733", "0.468", "0.733", "0.468", "0.747", "0.0801", "0.747" ], [ "0.0809", "0.75", "0.4429", "0.75", "0.4429", "0.7603", "0.0809", "0.7603" ], [ "0.5034", "0.7456", "0.8906", "0.7456", "0.8906", "0.7603", "0.5034", "0.7603" ], [ "0.0989", "0.7646", "0.468", "0.7646", "0.468", "0.778", "0.0989", "0.778" ], [ "0.5024", "0.761", "0.892", "0.761", "0.892", "0.7744", "0.5024", "0.7744" ], [ "0.0801", "0.778", "0.4697", "0.779", "0.4697", "0.7935", "0.0801", "0.7925" ], [ "0.5024", "0.7764", "0.8906", "0.7764", "0.8906", "0.7905", "0.5024", "0.7905" ], [ "0.0801", "0.7944", "0.4697", "0.7944", "0.4697", "0.808", "0.0801", "0.808" ], [ "0.5024", "0.7915", "0.892", "0.7915", "0.892", "0.8047", "0.5024", "0.8047" ], [ "0.0801", "0.81", "0.4697", "0.81", "0.4697", "0.8237", "0.0801", "0.8237" ], [ "0.5034", "0.806", "0.8906", "0.806", "0.8906", "0.8193", "0.5034", "0.8193" ], [ "0.0801", "0.8257", "0.4697", "0.8257", "0.4697", "0.8403", "0.0801", "0.8403" ], [ "0.5024", "0.8213", "0.892", "0.8213", "0.892", "0.8354", "0.5024", "0.8354" ], [ "0.0801", "0.8413", "0.4697", "0.8413", "0.4697", "0.855", "0.0801", "0.855" ], [ "0.5024", "0.836", "0.824", "0.836", "0.824", "0.851", "0.5024", "0.851" ], [ "0.0801", "0.857", "0.4697", "0.857", "0.4697", "0.8706", "0.0801", "0.8706" ], [ "0.5034", "0.86", "0.568", "0.86", "0.568", "0.875", "0.5034", "0.875" ], [ "0.6045", "0.86", "0.8896", "0.86", "0.8896", "0.875", "0.6045", "0.875" ], [ "0.0784", "0.8706", "0.4714", "0.8706", "0.4714", "0.888", "0.0784", "0.888" ], [ "0.574", "0.867", "0.581", "0.867", "0.581", "0.8735", "0.574", "0.8735" ], [ "0.569", "0.8774", "0.619", "0.8774", "0.619", "0.888", "0.569", "0.888" ], [ "0.0801", "0.886", "0.4673", "0.886", "0.4673", "0.901", "0.0801", "0.901" ], [ "0.863", "0.887", "0.892", "0.887", "0.892", "0.8994", "0.863", "0.8994" ] ]

[ "3.2. Hierarchical spatio-temporal AU relationship", "ematically, this process can be formulated as:", "modelling", "Ptub = o(FCsub(GAP(Gsub))", "(8)", "Our HSR module hierarchically models the spatio-temporal", "where denotes the Softmax function and FCsub denotes a", "relationship among AUs by specifically considering their", "fully connected layer. As a result, training this branch en-", "spatial distribution on the face, as association among AUs", "forces the network encoding underlying local AU relation-", "in the same/close facial region could be stronger than AUs", "ship to each sliced regional feature, allowing AFE to extract", "located in different facial regions [1]. It consists of two", "enhanced AU-relevant features from regional features.", "stages: the local AU relationship modelling stage first", "Cross-regional AU relationship modeling: Besides", "models the relationship among AUs located in the same fa-", "spatially adjacent AUs, each AU's activation may also as-", "cial region, and then the cross-regional AU relationship", "sociate with AUs located in other facial regions [17]. Con-", "modeling stage adaptively explore the relationship between", "sequently, this stage aims to enhance the recognition per-", "AU pairs of different facial regions.", "formance by additionally capturing such cross-regional AU", "Local AU Relationship Modelling: This stage specif-", "spatio-temporal dependencies within the given face image.", "ically models relationship among AUs located in the same", "It treats each local relationship-aware spatio-temporal AU", "facial regions at both their features extraction level and pre-", "feature ut extracted in the previous stage as a node, and", "diction level. It builds on the assumption that constraining", "adaptively connects it with all activated AU nodes belong", "each AU feature's extraction to its spatially correlated facial", "ing to other facial regions (i.e.,AU activation status are de-", "regions could partially avoid the negative impacts/noises", "cided by AU predictions of the first stage). This edge con-", "caused by irrelevant facial regions [19]. Particularly, it first", "nection definition is inspired by the finding that activated", "divides the spatio-temporal facial feature Gt e Rexhxw ex-", "AUs usually have more influences on other AUs [29]. As a", "tracted by MFD module into three subsets corresponding to", "result, the relationship of each cross regional AU pair is ex", "three slightly overlapped facial regions: (1) the upper re-", "plicitly represented through a graph edge, and further mod-", "gion encompassing eyebrows and eyes; (2) the middle re-", "elled via a Graph Attention Network (GAT) [49] layer as:", "gion containing the nose and cheeks; and (3)thelowerre-", "ern,m = LeakyReLU (rT [Wut I Wum])", "gion covering the mouth and chin (Illustrated in Fig. 1).", "This is achieved by directly slicing the feature Gt along the", "height dimension as:", "", "Can.mWut", "", "(9)", "mENt", "exp (en,m)", "th:hl", "Subject to: α", "=Gt[0:", "Gtp,Gtmid,Gtow", "hl,Gt", "", "-hl.C", ".", "Egen, exp (et.g)", "(7)", "where the height h of the Gt is 7 in our implementation,", "where en,m is a graph edge defines the impacts of the muh", "thus we empirically choose this best partition setting. Af-", "AU node to the nth AU node in the th frame: W E Rbxb de-", "ter that, N AU-specific Feature Extractors (AFE) (each is", "notes a shared linear transformation applied to every node", "made up of a convolution layer with kernel size of 1 × 1", "feature: is the concatenation operation: @ is an activa-", "and a Global Average Pooling (GAP) layer) are employed,", "tion function and r E R2b denotes the weight of an atten-", "where the nth extractor learns a local relationship-aware fea-", "tion operation.Nt is the set of the neighbours of the cur-", "ture ut e IR1xb (b denotes the dimension of an AU vector)", "rent node. Subsequently, N local and global hierarchical", "from its corresponding sliced regional feature (Gtp, Gmid", "relationship-aware AU features Vt = [ot, , - , o] are", "or Gtow), representing the nth AU's status at the tth frame.", "generated to describe N target AUs in the th frame.", "Consequently, each AU feature is extracted in the context of", "3.3. Loss function", "its spatially adjacent AUs (i.e., modelling AU relationship", "ofthesamefacialregionatthefeatureextractionlevel)", "As AU recognition constitutes a multi-label binary classifi-", "In addition, the spatio-temporal relationship among AUs", "cation task, with most AUs inactivated the across majority", "of the same facial region are also modelled at their predic-", "of frames (please refer to Supplementary Material for de-", "tion level, where an auxiliary branch (Aux) is added at the", "tails), an asymmetric loss function [29] is employed. Given", "training phase. It is trained to predict an AU occurrence", "the input consecutive T facial frames with N target AUs.", "combination Ysub = (ysub.1,..,ytub.2Nsub} (i.e., Yub is a", "the loss function CAu for supervising all AUs' recognition", "one-hot vector and Nsub is the number of the target AUs", "(i.e., output by the TCN/SC layers) is defined as:", "in the corresponding sub-region) from each sliced regional", "CAU = -", "wn [yt log(pt) + pt(1 - yt) log(1 -pt)", "feature Gsub E {Gup, Gmid, Gtow], which jointly describes", "", "n=l t=l", "all AUs' occurrence status within each facial region. Math-", "(10)" ]

{ "arxiv_id": "2404.06443", "language": "en", "page_idx": 4 }

[ [ "0.1626", "0.1332", "0.806", "0.1332", "0.806", "0.1509", "0.1626", "0.1509" ], [ "0.2418", "0.1812", "0.738", "0.1799", "0.738", "0.1976", "0.2418", "0.1989" ], [ "0.08826", "0.1963", "0.886", "0.1989", "0.886", "0.2166", "0.08826", "0.214" ], [ "0.2769", "0.2153", "0.6978", "0.2166", "0.6978", "0.2343", "0.2769", "0.2329" ], [ "0.1111", "0.238", "0.8594", "0.238", "0.8594", "0.252", "0.1111", "0.252" ], [ "0.0237", "0.2588", "0.05963", "0.2588", "0.05963", "0.3176", "0.0237", "0.3176" ], [ "0.2385", "0.2874", "0.3137", "0.2874", "0.3137", "0.3018", "0.2385", "0.3018" ], [ "0.5786", "0.3005", "0.7295", "0.2917", "0.7363", "0.3499", "0.585", "0.3591" ], [ "0.03186", "0.31", "0.0621", "0.3132", "0.05228", "0.3687", "0.02287", "0.3655" ], [ "0.03757", "0.3113", "0.0433", "0.3113", "0.0433", "0.3164", "0.03757", "0.3164" ], [ "0.0989", "0.3193", "0.468", "0.3193", "0.468", "0.3333", "0.0989", "0.3333" ], [ "0.7183", "0.3225", "0.7295", "0.3225", "0.7295", "0.332", "0.7183", "0.332" ], [ "0.0809", "0.3352", "0.4673", "0.3352", "0.4673", "0.3484", "0.0809", "0.3484" ], [ "0.0801", "0.351", "0.4673", "0.351", "0.4673", "0.3655", "0.0801", "0.3655" ], [ "0.79", "0.351", "0.822", "0.351", "0.822", "0.358", "0.79", "0.358" ], [ "0.6724", "0.361", "0.6953", "0.361", "0.6953", "0.3655", "0.6724", "0.3655" ], [ "0.847", "0.361", "0.8784", "0.361", "0.8784", "0.3655", "0.847", "0.3655" ], [ "0.02777", "0.377", "0.05554", "0.377", "0.05554", "0.3958", "0.02777", "0.3958" ], [ "0.0784", "0.3655", "0.4697", "0.3643", "0.4697", "0.379", "0.0784", "0.3801" ], [ "0.0801", "0.3806", "0.468", "0.3806", "0.468", "0.3945", "0.0801", "0.3945" ], [ "0.5034", "0.3833", "0.8906", "0.3833", "0.8906", "0.3933", "0.5034", "0.3933" ], [ "0.0801", "0.3958", "0.4714", "0.3945", "0.4714", "0.4092", "0.0801", "0.4104" ], [ "0.5034", "0.3977", "0.8896", "0.3977", "0.8896", "0.4077", "0.5034", "0.4077" ], [ "0.02696", "0.4072", "0.05554", "0.4072", "0.05554", "0.4343", "0.02696", "0.4343" ], [ "0.0809", "0.4124", "0.468", "0.4124", "0.468", "0.4224", "0.0809", "0.4224" ], [ "0.5024", "0.4104", "0.892", "0.4104", "0.892", "0.4236", "0.5024", "0.4236" ], [ "0.0784", "0.4255", "0.468", "0.4255", "0.468", "0.4395", "0.0784", "0.4395" ], [ "0.5024", "0.4236", "0.8906", "0.4248", "0.8906", "0.4363", "0.5024", "0.435" ], [ "0.0809", "0.4426", "0.468", "0.4426", "0.468", "0.4526", "0.0809", "0.4526" ], [ "0.5034", "0.4382", "0.892", "0.4382", "0.892", "0.452", "0.5034", "0.452" ], [ "0.0335", "0.4495", "0.05554", "0.4495", "0.05554", "0.4797", "0.0335", "0.4797" ], [ "0.5024", "0.452", "0.892", "0.452", "0.892", "0.4653", "0.5024", "0.4653" ], [ "0.0809", "0.4583", "0.468", "0.4583", "0.468", "0.4685", "0.0809", "0.4685" ], [ "0.501", "0.464", "0.892", "0.4653", "0.892", "0.4792", "0.501", "0.4778" ], [ "0.0801", "0.4717", "0.468", "0.4717", "0.468", "0.4849", "0.0801", "0.4849" ], [ "0.0801", "0.486", "0.468", "0.486", "0.468", "0.5", "0.0801", "0.5" ], [ "0.5034", "0.481", "0.8896", "0.481", "0.8896", "0.4912", "0.5034", "0.4912" ], [ "0.02777", "0.4956", "0.05392", "0.4956", "0.05392", "0.5146", "0.02777", "0.5146" ], [ "0.501", "0.4937", "0.761", "0.4937", "0.761", "0.507", "0.501", "0.507" ], [ "0.0801", "0.502", "0.4697", "0.502", "0.4697", "0.515", "0.0801", "0.515" ], [ "0.03513", "0.509", "0.05148", "0.509", "0.05148", "0.5244", "0.03513", "0.5244" ], [ "0.02777", "0.5195", "0.05554", "0.5195", "0.05554", "0.537", "0.02777", "0.537" ], [ "0.0801", "0.5176", "0.4673", "0.5176", "0.4673", "0.531", "0.0801", "0.531" ], [ "0.5024", "0.523", "0.892", "0.523", "0.892", "0.536", "0.5024", "0.536" ], [ "0.02942", "0.535", "0.05228", "0.535", "0.05228", "0.5474", "0.02942", "0.5474" ], [ "0.0801", "0.531", "0.468", "0.531", "0.468", "0.5444", "0.0801", "0.5444" ], [ "0.501", "0.536", "0.892", "0.537", "0.892", "0.552", "0.501", "0.5503" ], [ "0.02533", "0.5435", "0.05637", "0.5435", "0.05637", "0.58", "0.02533", "0.58" ], [ "0.0801", "0.547", "0.468", "0.547", "0.468", "0.561", "0.0801", "0.561" ], [ "0.505", "0.5537", "0.8896", "0.5537", "0.8896", "0.5645", "0.505", "0.5645" ], [ "0.0801", "0.563", "0.4673", "0.563", "0.4673", "0.5737", "0.0801", "0.5737" ], [ "0.5024", "0.5684", "0.8667", "0.5684", "0.8667", "0.582", "0.5024", "0.582" ], [ "0.0801", "0.577", "0.468", "0.577", "0.468", "0.5903", "0.0801", "0.5903" ], [ "0.5215", "0.582", "0.8906", "0.5835", "0.8906", "0.5977", "0.5215", "0.5967" ], [ "0.02942", "0.5923", "0.05228", "0.5923", "0.05228", "0.606", "0.02942", "0.606" ], [ "0.0801", "0.5923", "0.468", "0.5923", "0.468", "0.606", "0.0801", "0.606" ], [ "0.02942", "0.601", "0.05392", "0.601", "0.05392", "0.6235", "0.02942", "0.6235" ], [ "0.5034", "0.5986", "0.892", "0.5986", "0.892", "0.6123", "0.5034", "0.6123" ], [ "0.0801", "0.607", "0.468", "0.607", "0.468", "0.6206", "0.0801", "0.6206" ], [ "0.03757", "0.619", "0.04413", "0.625", "0.03757", "0.631", "0.02942", "0.625" ], [ "0.5024", "0.6143", "0.8906", "0.6143", "0.8906", "0.6274", "0.5024", "0.6274" ], [ "0.02287", "0.6226", "0.06046", "0.6226", "0.06046", "0.712", "0.02287", "0.712" ], [ "0.0801", "0.6226", "0.4697", "0.6226", "0.4697", "0.636", "0.0801", "0.636" ], [ "0.5024", "0.629", "0.8906", "0.629", "0.8906", "0.6426", "0.5024", "0.6426" ], [ "0.0784", "0.637", "0.4673", "0.636", "0.4673", "0.651", "0.0784", "0.6514" ], [ "0.5034", "0.6445", "0.8906", "0.6445", "0.8906", "0.6577", "0.5034", "0.6577" ], [ "0.0825", "0.654", "0.4673", "0.654", "0.4673", "0.664", "0.0825", "0.664" ], [ "0.501", "0.659", "0.8896", "0.659", "0.8896", "0.6724", "0.501", "0.6724" ], [ "0.0801", "0.6685", "0.423", "0.6685", "0.423", "0.6816", "0.0801", "0.6816" ], [ "0.5034", "0.675", "0.8906", "0.675", "0.8906", "0.688", "0.5034", "0.688" ], [ "0.5024", "0.6895", "0.8906", "0.6895", "0.8906", "0.7026", "0.5024", "0.7026" ], [ "0.0825", "0.705", "0.2075", "0.705", "0.2075", "0.7183", "0.0825", "0.7183" ], [ "0.5034", "0.705", "0.892", "0.705", "0.892", "0.7183", "0.5034", "0.7183" ], [ "0.5034", "0.7197", "0.892", "0.7197", "0.892", "0.733", "0.5034", "0.733" ], [ "0.0989", "0.729", "0.468", "0.729", "0.468", "0.7437", "0.0989", "0.7437" ], [ "0.5034", "0.735", "0.892", "0.735", "0.892", "0.7485", "0.5034", "0.7485" ], [ "0.0809", "0.744", "0.468", "0.744", "0.468", "0.7583", "0.0809", "0.7583" ], [ "0.5024", "0.75", "0.8906", "0.75", "0.8906", "0.763", "0.5024", "0.763" ], [ "0.0801", "0.7603", "0.468", "0.7603", "0.468", "0.7734", "0.0801", "0.7734" ], [ "0.5034", "0.7666", "0.8906", "0.7666", "0.8906", "0.7764", "0.5034", "0.7764" ], [ "0.0784", "0.7734", "0.4697", "0.7744", "0.4697", "0.789", "0.0784", "0.7886" ], [ "0.5024", "0.78", "0.892", "0.78", "0.892", "0.7944", "0.5024", "0.7944" ], [ "0.0801", "0.7905", "0.468", "0.7905", "0.468", "0.8037", "0.0801", "0.8037" ], [ "0.5034", "0.7954", "0.8906", "0.7954", "0.8906", "0.8096", "0.5034", "0.8096" ], [ "0.0801", "0.806", "0.468", "0.806", "0.468", "0.8193", "0.0801", "0.8193" ], [ "0.5034", "0.811", "0.734", "0.811", "0.734", "0.8247", "0.5034", "0.8247" ], [ "0.0801", "0.821", "0.468", "0.821", "0.468", "0.834", "0.0801", "0.834" ], [ "0.523", "0.827", "0.892", "0.827", "0.892", "0.8403", "0.523", "0.8403" ], [ "0.0801", "0.8354", "0.468", "0.8354", "0.468", "0.8486", "0.0801", "0.8486" ], [ "0.5024", "0.8413", "0.8906", "0.8413", "0.8906", "0.855", "0.5024", "0.855" ], [ "0.0809", "0.8516", "0.4656", "0.8516", "0.4656", "0.8623", "0.0809", "0.8623" ], [ "0.5034", "0.857", "0.892", "0.857", "0.892", "0.8706", "0.5034", "0.8706" ], [ "0.0972", "0.8755", "0.2239", "0.8755", "0.2239", "0.8896", "0.0972", "0.8896" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0989", "0.887", "0.4412", "0.887", "0.4412", "0.9014", "0.0989", "0.9014" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ], [ "0.4836", "0.928", "0.4902", "0.928", "0.4902", "0.933", "0.4836", "0.933" ] ]

[ "LiBRe: A Practical Bayesian Approach to Adversarial Detection", "Zhijie Deng', Xiao Yang', Shizhen Xu?, Hang Sul*, Jun Zhul*", "1 Dept. of Comp. Sci. and Tech., BNRist Center, Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab", "1 Tsinghua University, Beijing, 100084, China2 RealAI", "[dzj17,yangxiao19)@mails.tsinghua.edu.cn,shizhen.xu@realai.ai,(suhangss,dcszj)@tsinghua.edu.cn", "12", "Abstract", "", "e", "", "Despite theirappealingflexibility, deep neural networks", "", "(DNNs) are vulnerable against adversarial examples. Vari-", "ous adversarial defense strategies have been proposed to re-", "orediction", "tiea", "tlreipr", "3", "solve this problem, but they typically demonstrate restricted", "practicability owing to unsurmountable compromise on uni-", "Figurel:Given apre-trainedDNN,LiBRe convertsits lastfew", "versality, effectiveness, or efficiencyIn this work, we pro-", "layers(excludingthetask-dependent outputhead)tobeBayesian", "G", "poseamorepracticalapproach,LightweightBavesianRe-", "and reuses the pre-trained parameters.Then, LiBRe launches", "finement (LiBRe), in the spirit of leveraging Bayesian neu-", "several-round adversarial detection-oriented fine-tuning to render", "ral networks (BNNs)for adversarial detection.Empow-", "the posterior effective for prediction and meanwhile appropriate", "'SO", "for adversarial detection. In the inference phase, LiBRe estimates", "eredbythetaskandattackagnosticmodelingunderBayes", "the predictive uncertainty and task-dependent predictions of the", "principle, LiBRe can endow a variety of pre-trained task-", "dependent DNNs with the ability of defending heteroge-", "input concurrently,where theformeris used for adversarial detec-", "2", "tion and determines the fidelity of the latter.", "neous adversarial attacks at a low cost. We develop and", "V", "5", "integrate advanced learning techniques to make LiBRe ap-", "versarial examples have been witnessed in a wide spectrum", "3", "propriate for adversarial detection. Concretely, we build", "of practical systems [52, 12], raising an urgent requirement", "8", "the few-layer deep ensemble variational and adopt the pre-", "foradvanced techniques to achieverobust and reliabledeci-", "training & fine-tuning workflowtoboosttheeffectiveness", "sion making, especially in safety-critical scenarios [13].", "and efficiency of LiBRe.We further provide a novel in-", "Though increasing methods have been developed to", "3", "sight to realise adversarial detection-oriented uncertainty", "0", "tackle adversarial examples [42, 69, 25, 18, 68], they are", "correction without inefficiently crafting adversarial exam-", "", "not problemless. On on hand, as one of the most popu-", "AX", "ples during training. Extensive empirical studies covering", "lar adversarial defenses, adversarial training [42, 69] intro-", "a wide range of scenarios verify the practicability of LiBRe.", "duces adversarial examples into training to explicitly tailor", "Wealso conductthoroughablationstudiesto evidence the", "the decision boundaries, which, yet, causes added training", "superiority of our modeling and learning strategies.", "overheads and typically leads to degraded predictive per-", "formance on clean examples. On the other hand, adversar-", "1.Introduction", "ial detection methods bypass the drawbacks of modifying", "the original DNNs by deploying a workflow to detect the", "The blooming development of deep neural networks", "adversarial examples ahead of decision making, by virtue", "(DNNs) has brought great success in extensive industrial", "of auxiliary classifiers [44, 18, 68, 5] or designed statis-", "applications, such as image classification [23], face recog-", "tics [14,40l.Yet.theyareusually developedfor specific", "nition [9] and object detection [50]. However, despite", "tasks (e.g., image classification [68, 32, 18]) or for specific", "their promising expressiveness, DNNs are highly vulnera-", "adversarial attacks [39], lacking the flexibility to effectively", "ble to adversarial examples [57, 19], which are generated by", "generalize to other tasks or attacks.", "adding human-imperceptible perturbations upon clean ex-", "By regarding the adversarial example as a special case", "amples to deliberately cause misclassification, partly due to", "of out-of-distribution (OOD) data, Bayesian neural net-", "their non-linearand black-box nature.Thethreatsfrom ad-", "works (BNNs) have shown promise in adversarial detec-", "*Corresponding author", "tion [14, 38, 54]. In theory, the predictive uncertainty ac-", "'Code at https://github.com/thudzj/scalableBDL.", "quired under Bayes principle suffices for detecting hetero-", "1" ]

{ "arxiv_id": "2103.14835", "language": "en", "page_idx": 0 }

[ [ "0.0809", "0.09216", "0.4697", "0.09216", "0.4697", "0.1061", "0.0809", "0.1061" ], [ "0.5024", "0.09155", "0.892", "0.0903", "0.892", "0.1048", "0.5024", "0.1061" ], [ "0.0784", "0.108", "0.4697", "0.1067", "0.4697", "0.1206", "0.0784", "0.1218" ], [ "0.5034", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.5034", "0.1218" ], [ "0.0825", "0.125", "0.468", "0.125", "0.468", "0.1351", "0.0825", "0.1351" ], [ "0.5034", "0.1225", "0.8906", "0.1225", "0.8906", "0.137", "0.5034", "0.137" ], [ "0.0801", "0.137", "0.468", "0.137", "0.468", "0.1509", "0.0801", "0.1509" ], [ "0.501", "0.1383", "0.892", "0.1364", "0.892", "0.1509", "0.501", "0.1528" ], [ "0.0809", "0.1528", "0.468", "0.1528", "0.468", "0.1666", "0.0809", "0.1666" ], [ "0.5034", "0.1547", "0.8906", "0.1547", "0.8906", "0.1654", "0.5034", "0.1654" ], [ "0.0801", "0.1674", "0.4697", "0.1674", "0.4697", "0.1812", "0.0801", "0.1812" ], [ "0.501", "0.1674", "0.8896", "0.1674", "0.8896", "0.1812", "0.501", "0.1812" ], [ "0.0801", "0.1831", "0.4697", "0.1831", "0.4697", "0.1963", "0.0801", "0.1963" ], [ "0.5034", "0.1831", "0.8906", "0.1831", "0.8906", "0.1963", "0.5034", "0.1963" ], [ "0.0784", "0.1963", "0.4617", "0.1976", "0.4617", "0.2122", "0.0784", "0.2108" ], [ "0.5024", "0.1963", "0.8906", "0.1976", "0.8906", "0.2122", "0.5024", "0.2108" ], [ "0.0989", "0.2166", "0.4673", "0.2166", "0.4673", "0.2299", "0.0989", "0.2299" ], [ "0.501", "0.2108", "0.8613", "0.2134", "0.8613", "0.2279", "0.501", "0.226" ], [ "0.0801", "0.2311", "0.4697", "0.2311", "0.4697", "0.2443", "0.0801", "0.2443" ], [ "0.5024", "0.2393", "0.6406", "0.2393", "0.6406", "0.254", "0.5024", "0.254" ], [ "0.0801", "0.2468", "0.468", "0.2468", "0.468", "0.2615", "0.0801", "0.2615" ], [ "0.0801", "0.262", "0.468", "0.262", "0.468", "0.276", "0.0801", "0.276" ], [ "0.523", "0.2615", "0.8906", "0.2615", "0.8906", "0.276", "0.523", "0.276" ], [ "0.0801", "0.2766", "0.468", "0.2766", "0.468", "0.2903", "0.0801", "0.2903" ], [ "0.5024", "0.2766", "0.8906", "0.2766", "0.8906", "0.2903", "0.5024", "0.2903" ], [ "0.0784", "0.2917", "0.4697", "0.2903", "0.4697", "0.305", "0.0784", "0.3062" ], [ "0.5034", "0.2935", "0.8896", "0.2935", "0.8896", "0.3037", "0.5034", "0.3037" ], [ "0.0801", "0.307", "0.468", "0.307", "0.468", "0.3208", "0.0801", "0.3208" ], [ "0.5034", "0.307", "0.892", "0.307", "0.892", "0.3208", "0.5034", "0.3208" ], [ "0.0809", "0.3213", "0.4697", "0.3213", "0.4697", "0.336", "0.0809", "0.336" ], [ "0.501", "0.3213", "0.8906", "0.3213", "0.8906", "0.3352", "0.501", "0.3352" ], [ "0.0825", "0.3384", "0.468", "0.3384", "0.468", "0.3484", "0.0825", "0.3484" ], [ "0.5024", "0.3372", "0.892", "0.3372", "0.892", "0.351", "0.5024", "0.351" ], [ "0.0809", "0.3516", "0.4673", "0.3516", "0.4673", "0.3655", "0.0809", "0.3655" ], [ "0.5024", "0.3528", "0.8906", "0.3528", "0.8906", "0.3662", "0.5024", "0.3662" ], [ "0.0809", "0.3674", "0.4697", "0.3674", "0.4697", "0.3806", "0.0809", "0.3806" ], [ "0.5034", "0.3687", "0.892", "0.3687", "0.892", "0.3818", "0.5034", "0.3818" ], [ "0.0825", "0.3838", "0.4656", "0.3838", "0.4656", "0.3945", "0.0825", "0.3945" ], [ "0.505", "0.3838", "0.886", "0.3838", "0.886", "0.3945", "0.505", "0.3945" ], [ "0.0809", "0.3977", "0.4673", "0.3977", "0.4673", "0.411", "0.0809", "0.411" ], [ "0.501", "0.3977", "0.8896", "0.3977", "0.8896", "0.411", "0.501", "0.411" ], [ "0.0801", "0.4124", "0.4697", "0.4124", "0.4697", "0.4255", "0.0801", "0.4255" ], [ "0.5034", "0.4124", "0.8906", "0.4124", "0.8906", "0.4255", "0.5034", "0.4255" ], [ "0.0801", "0.428", "0.4697", "0.428", "0.4697", "0.4414", "0.0801", "0.4414" ], [ "0.501", "0.428", "0.892", "0.4268", "0.892", "0.44", "0.501", "0.4414" ], [ "0.0801", "0.443", "0.468", "0.443", "0.468", "0.457", "0.0801", "0.457" ], [ "0.501", "0.4414", "0.892", "0.4426", "0.892", "0.457", "0.501", "0.4558" ], [ "0.0801", "0.4583", "0.4673", "0.4583", "0.4673", "0.4717", "0.0801", "0.4717" ], [ "0.5034", "0.4583", "0.892", "0.4583", "0.892", "0.4717", "0.5034", "0.4717" ], [ "0.0809", "0.4749", "0.468", "0.4749", "0.468", "0.4849", "0.0809", "0.4849" ], [ "0.505", "0.4749", "0.892", "0.4749", "0.892", "0.4849", "0.505", "0.4849" ], [ "0.0801", "0.488", "0.4697", "0.488", "0.4697", "0.502", "0.0801", "0.502" ], [ "0.5024", "0.488", "0.892", "0.488", "0.892", "0.502", "0.5024", "0.502" ], [ "0.0801", "0.5024", "0.4673", "0.5024", "0.4673", "0.5166", "0.0801", "0.5166" ], [ "0.5024", "0.504", "0.8906", "0.504", "0.8906", "0.5176", "0.5024", "0.5176" ], [ "0.0801", "0.518", "0.4673", "0.518", "0.4673", "0.532", "0.0801", "0.532" ], [ "0.5024", "0.518", "0.888", "0.518", "0.888", "0.532", "0.5024", "0.532" ], [ "0.0809", "0.5327", "0.468", "0.5327", "0.468", "0.547", "0.0809", "0.547" ], [ "0.523", "0.534", "0.8906", "0.534", "0.8906", "0.5474", "0.523", "0.5474" ], [ "0.0809", "0.549", "0.468", "0.549", "0.468", "0.5625", "0.0809", "0.5625" ], [ "0.5034", "0.55", "0.892", "0.55", "0.892", "0.563", "0.5034", "0.563" ], [ "0.0801", "0.563", "0.4697", "0.563", "0.4697", "0.577", "0.0801", "0.577" ], [ "0.5024", "0.563", "0.892", "0.563", "0.892", "0.577", "0.5024", "0.577" ], [ "0.0809", "0.579", "0.4673", "0.579", "0.4673", "0.5923", "0.0809", "0.5923" ], [ "0.5034", "0.579", "0.8906", "0.579", "0.8906", "0.5923", "0.5034", "0.5923" ], [ "0.0801", "0.5933", "0.273", "0.5933", "0.273", "0.607", "0.0801", "0.607" ], [ "0.5034", "0.5933", "0.8896", "0.5933", "0.8896", "0.607", "0.5034", "0.607" ], [ "0.0989", "0.61", "0.4673", "0.6113", "0.4673", "0.6255", "0.0989", "0.625" ], [ "0.5024", "0.6094", "0.892", "0.6094", "0.892", "0.6226", "0.5024", "0.6226" ], [ "0.0801", "0.627", "0.4673", "0.627", "0.4673", "0.6416", "0.0801", "0.6416" ], [ "0.5034", "0.6255", "0.8896", "0.6255", "0.8896", "0.636", "0.5034", "0.636" ], [ "0.0801", "0.6426", "0.468", "0.6426", "0.468", "0.656", "0.0801", "0.656" ], [ "0.5024", "0.6396", "0.892", "0.6396", "0.892", "0.653", "0.5024", "0.653" ], [ "0.0801", "0.656", "0.4673", "0.656", "0.4673", "0.6694", "0.0801", "0.6694" ], [ "0.5034", "0.656", "0.8906", "0.656", "0.8906", "0.6665", "0.5034", "0.6665" ], [ "0.0784", "0.6704", "0.468", "0.672", "0.468", "0.686", "0.0784", "0.685" ], [ "0.5024", "0.6694", "0.8906", "0.6694", "0.8906", "0.683", "0.5024", "0.683" ], [ "0.0801", "0.6875", "0.4697", "0.6875", "0.4697", "0.7007", "0.0801", "0.7007" ], [ "0.5024", "0.685", "0.8906", "0.685", "0.8906", "0.6987", "0.5024", "0.6987" ], [ "0.0784", "0.7026", "0.4673", "0.702", "0.4673", "0.7153", "0.0784", "0.7163" ], [ "0.501", "0.6987", "0.74", "0.6997", "0.74", "0.714", "0.501", "0.7134" ], [ "0.0784", "0.7163", "0.468", "0.7173", "0.468", "0.732", "0.0784", "0.731" ], [ "0.501", "0.7227", "0.6284", "0.7246", "0.6284", "0.742", "0.501", "0.7407" ], [ "0.0809", "0.733", "0.468", "0.733", "0.468", "0.747", "0.0809", "0.747" ], [ "0.0809", "0.75", "0.4673", "0.75", "0.4673", "0.7603", "0.0809", "0.7603" ], [ "0.5215", "0.7476", "0.8906", "0.7476", "0.8906", "0.761", "0.5215", "0.761" ], [ "0.0801", "0.763", "0.4697", "0.763", "0.4697", "0.7764", "0.0801", "0.7764" ], [ "0.5024", "0.763", "0.892", "0.763", "0.892", "0.7764", "0.5024", "0.7764" ], [ "0.0809", "0.78", "0.468", "0.78", "0.468", "0.7905", "0.0809", "0.7905" ], [ "0.5024", "0.779", "0.892", "0.779", "0.892", "0.7925", "0.5024", "0.7925" ], [ "0.0801", "0.7935", "0.4697", "0.7935", "0.4697", "0.8066", "0.0801", "0.8066" ], [ "0.5024", "0.7935", "0.751", "0.7935", "0.751", "0.8066", "0.5024", "0.8066" ], [ "0.0801", "0.8096", "0.429", "0.8096", "0.429", "0.8228", "0.0801", "0.8228" ], [ "0.5034", "0.8174", "0.6914", "0.8174", "0.6914", "0.832", "0.5034", "0.832" ], [ "0.0989", "0.8257", "0.4697", "0.8257", "0.4697", "0.839", "0.0989", "0.839" ], [ "0.0809", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0809", "0.855" ], [ "0.523", "0.8403", "0.8906", "0.8403", "0.8906", "0.855", "0.523", "0.855" ], [ "0.0784", "0.855", "0.4697", "0.856", "0.4697", "0.8706", "0.0784", "0.869" ], [ "0.5024", "0.857", "0.892", "0.857", "0.892", "0.8706", "0.5024", "0.8706" ], [ "0.0801", "0.872", "0.4714", "0.872", "0.4714", "0.8853", "0.0801", "0.8853" ], [ "0.5024", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5024", "0.8853" ], [ "0.0768", "0.8853", "0.4697", "0.886", "0.4697", "0.901", "0.0768", "0.8994" ], [ "0.5024", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5024", "0.901" ], [ "0.4805", "0.925", "0.4927", "0.925", "0.4927", "0.936", "0.4805", "0.936" ] ]

[ "geneous adversarial examples in various tasks. However, in", "ference (i.e., testing) speed than typical BNNs thanks to the", "practice, BNNs without a sharpened posterior often present", "adoption of lightweight variational. We can achieve further", "systematicallyworseperformance thantheirdeterministic", "speedup by exploring the potential of parallel computing,", "counterparts [61l; also relatively low-cost Bayesian infer-", "giving rise to inference speed close to the DNN in the same", "ence methods frequently suffer from mode collapse and", "setting.Extensive experiments in scenariosranging from", "hence unreliable uncertainty [15]. BNNs requirement of", "image classification, face recognition, to object detection", "more expertise for implementation and more efforts for", "confirm these claims and testify the superiority of LiBRe.", "training than DNNs further undermine their practicability.", "We further perform thorough ablation studies to deeply un-", "In the work, we aim to develop a more practical adversar-", "derstand the adopted modeling and learning strategies.", "ial detection approach by overcoming the aforementioned", "2. Related Work", "issues of BNNs. We propose Lightweight Bayesian Refine-", "ment (LiBRe), depicted in Fig. 1, to reach a good balance", "Detecting adversarial examples to bypass their safety", "among predictive performance, quality of uncertainty esti-", "threats has attracted increasing attention recently.Many", "mates and learning effciency. Concretely, LiBRe follows", "worksaim atdistinguishing adversarial examples from be-", "the stochastic variational inference pipeline [2], but is em-", "nign ones via an auxiliary classifier applied on statistical", "powered by two non-trivial designs: (i) To achieve efficient", "features [18, 68, 5, 7, 65]. [21] introduces an extra class in", "learning withhigh-guality outcomes, we devise the Few-", "the classifier for adversarial examples. Some recent works", "IAyer Deep Ensemble (FADE) variational, which is remi-", "exploit neighboring statistics to construct more powerful", "niscent of Deep Ensemble [31], one of the most effective", "detection algorithms: [32] fits a Gaussian mixture model of", "Bayesianmarginalizationmethods[62],andmeanwhilein-", "thenetworkresponses,andresortstotheMahalanbobis dis", "spired by the scalable last-layer Bayesian inference [29].", "tance for adversarial detection in the inference phase; [40]", "Namely, FADE only performs deep ensemble in the last few", "introduces the more advanced local intrinsic dimensional-", "layers of a model due to their crucial role for determining", "ity to describe the distance distribution and observes better", "model behaviour, while keeps the other layers determinis-", "results. RCE [47] is developed with the promise of leading", "tic. To encourage various ensemble candidates to capture", "to an enhanced distance between adversarial and normal im-", "diversefunctionmodes,wedevelopa stochasticity-injected", "agesforkerneldensity[14baseddetection.However,most", "learning principle for FADE, which also benefits to reduce", "of the aforementioned methods are restricted in the classi-", "the gradient variance of the parameters. (ii) To further ease", "fication scope, and the detectors trained against certain at-", "and accelerate the learning, we propose a Bayesian refine-", "tacks may not effectively generalize to unseen attacks [39]", "ment paradigm, where we initialize the parameters of FADE", "Bayesian deep learning [20, 60, 2, 1, 36, 26] provides", "with the parameters of its pre-trained deterministic coun-", "us with a more theoretically appealing way to adversarial", "terpart, thanks to the high alignment between FADE and", "detection. However, though the existing BNNs manage to", "point estimate. We then perform fine-tuning to constantly", "perceive adversarial examples [14, 49,54, 38, 48, 33], they", "improve the FADE posterior.", "are typically limited in terms of training efficiency, predic-", "As revealed by [22], the uncertainty quantification purely", "tive performance, etc., and thus cannot effectively scale up", "acquired from Bayes principle may be unreliable for per-", "to real-world settings. More severely,the uncertainty esti-", "ceiving adversarial examples, thus it is indispensable to pur-", "mates given by the BNNs for adversarial examples are not", "sue an adversarial detection-oriented uncertainty correction.", "alwaysreliable[22l,owingtothelackofparticulardesigns", "For universality, we place no assumption on the adversarial", "for adversarial detection. In this work, we address these is-", "examples to detect, so we cannot take the common strat-", "sues with elaborated techniques and establish a more prac-", "egy of integrating the adversarial examples crafted by spe-", "tical adversarial detection approach.", "cific attacks into detector training [40]. Alternatively, we", "3.Background", "cheaply create uniformly perturbed examples and demand", "highpredictiveuncertaintyonthemduringBayesianrefine-", "In this section, we motivate Lightweight Bayesian Re-", "ment to make the model be sensitive to data with any style of", "finement (LiBRe) by briefly reviewing the background of", "perturbation.Though suchacorrectionrendersthelearned", "adversarial defense, and then describe the general workflow", "posterior slightly deviated from the true Bayesian one, it can", "of Bayesian neural networks (BNNs).", "significantly boost adversarial detection performance", "3.1.Adversarial Defense", "The task and attack agnostic designs enable LiBRe to", "quickly and cheaply endow a pre-trained task-dependent", "Typically, let D = {(ci, yi))n-, denote a collection of n", "DNN with the ability to detect various adversarial examples", "training samples with &; e Rd and yi E as the input data", "when facing new tasks, as testified by our empirical studies", "and label, respectively. A deep neural network (DNN) pa-", "in Sec 5. Furthermore, LiBRe has significantly higher in-", "rameterized by w E RP is frequently trained via maximum", "2" ]

{ "arxiv_id": "2103.14835", "language": "en", "page_idx": 1 }

[ [ "0.0801", "0.09216", "0.2852", "0.09216", "0.2852", "0.1061", "0.0801", "0.1061" ], [ "0.5034", "0.09155", "0.892", "0.09155", "0.892", "0.1048", "0.5034", "0.1048" ], [ "0.5024", "0.108", "0.892", "0.108", "0.892", "0.1218", "0.5024", "0.1218" ], [ "0.1772", "0.1187", "0.2115", "0.11615", "0.2133", "0.1288", "0.179", "0.1307" ], [ "0.2018", "0.127", "0.3179", "0.1256", "0.3179", "0.1401", "0.2018", "0.1414" ], [ "0.3423", "0.1276", "0.4019", "0.125", "0.4028", "0.1395", "0.343", "0.1427" ], [ "0.4478", "0.1276", "0.4714", "0.1276", "0.4714", "0.1395", "0.4478", "0.1395" ], [ "0.501", "0.1225", "0.8906", "0.1225", "0.8906", "0.137", "0.501", "0.137" ], [ "0.1478", "0.1288", "0.1814", "0.1288", "0.1814", "0.1383", "0.1478", "0.1383" ], [ "0.165", "0.137", "0.1969", "0.137", "0.1969", "0.1459", "0.165", "0.1459" ], [ "0.3308", "0.1401", "0.339", "0.1401", "0.339", "0.1433", "0.3308", "0.1433" ], [ "0.5063", "0.1547", "0.584", "0.159", "0.5825", "0.1736", "0.505", "0.1692" ], [ "0.761", "0.1559", "0.8906", "0.1559", "0.8906", "0.1704", "0.761", "0.1704" ], [ "0.6313", "0.1578", "0.722", "0.1559", "0.722", "0.1704", "0.6313", "0.1724" ], [ "0.0784", "0.1572", "0.468", "0.1559", "0.468", "0.1704", "0.0784", "0.1718" ], [ "0.0784", "0.1718", "0.4697", "0.1724", "0.4697", "0.1875", "0.0784", "0.1863" ], [ "0.871", "0.1812", "0.892", "0.1812", "0.892", "0.1932", "0.871", "0.1932" ], [ "0.0809", "0.1895", "0.468", "0.1895", "0.468", "0.1995", "0.0809", "0.1995" ], [ "0.523", "0.1945", "0.892", "0.1945", "0.892", "0.2084", "0.523", "0.2084" ], [ "0.0809", "0.2039", "0.4656", "0.2039", "0.4656", "0.214", "0.0809", "0.214" ], [ "0.501", "0.2102", "0.8906", "0.2102", "0.8906", "0.2235", "0.501", "0.2235" ], [ "0.0784", "0.2166", "0.4697", "0.2172", "0.4697", "0.2317", "0.0784", "0.2311" ], [ "0.5024", "0.2247", "0.892", "0.2247", "0.892", "0.238", "0.5024", "0.238" ], [ "0.0784", "0.2329", "0.3545", "0.2311", "0.3545", "0.2456", "0.0784", "0.2474" ], [ "0.7417", "0.254", "0.7573", "0.254", "0.7573", "0.2615", "0.7417", "0.2615" ], [ "0.759", "0.2524", "0.7695", "0.2524", "0.7695", "0.2568", "0.759", "0.2568" ], [ "0.5156", "0.2588", "0.7363", "0.2588", "0.7363", "0.2734", "0.5156", "0.2734" ], [ "0.7646", "0.2588", "0.892", "0.2556", "0.892", "0.2708", "0.7646", "0.2734" ], [ "0.4478", "0.26", "0.4714", "0.26", "0.4714", "0.2734", "0.4478", "0.2734" ], [ "0.1396", "0.2556", "0.406", "0.2583", "0.406", "0.276", "0.1396", "0.2734" ], [ "0.2361", "0.2734", "0.2646", "0.2734", "0.2646", "0.281", "0.2361", "0.281" ], [ "0.7344", "0.2676", "0.7656", "0.2651", "0.767", "0.2734", "0.7363", "0.276" ], [ "0.755", "0.276", "0.773", "0.276", "0.773", "0.2827", "0.755", "0.2827" ], [ "0.0801", "0.2935", "0.4697", "0.2935", "0.4697", "0.307", "0.0801", "0.307" ], [ "0.5034", "0.2935", "0.892", "0.2935", "0.892", "0.308", "0.5034", "0.308" ], [ "0.0801", "0.3093", "0.4697", "0.3093", "0.4697", "0.3225", "0.0801", "0.3225" ], [ "0.5024", "0.3093", "0.8906", "0.31", "0.8906", "0.324", "0.5024", "0.3225" ], [ "0.0784", "0.324", "0.4697", "0.3225", "0.4697", "0.3372", "0.0784", "0.3384" ], [ "0.5024", "0.3245", "0.892", "0.3245", "0.892", "0.3384", "0.5024", "0.3384" ], [ "0.0801", "0.339", "0.4697", "0.339", "0.4697", "0.3528", "0.0801", "0.3528" ], [ "0.5024", "0.3403", "0.892", "0.3403", "0.892", "0.3542", "0.5024", "0.3542" ], [ "0.0825", "0.356", "0.3228", "0.356", "0.3228", "0.3662", "0.0825", "0.3662" ], [ "0.5", "0.3542", "0.841", "0.3547", "0.841", "0.3694", "0.5", "0.3687" ], [ "0.10046", "0.3694", "0.468", "0.3694", "0.468", "0.3833", "0.10046", "0.3833" ], [ "0.523", "0.3706", "0.892", "0.3706", "0.892", "0.3838", "0.523", "0.3838" ], [ "0.0809", "0.3865", "0.4673", "0.3865", "0.4673", "0.3965", "0.0809", "0.3965" ], [ "0.5024", "0.385", "0.8906", "0.385", "0.8906", "0.3984", "0.5024", "0.3984" ], [ "0.0784", "0.3984", "0.4697", "0.3997", "0.4697", "0.414", "0.0784", "0.4136" ], [ "0.5034", "0.4016", "0.8896", "0.4016", "0.8896", "0.4124", "0.5034", "0.4124" ], [ "0.0801", "0.4153", "0.4697", "0.4153", "0.4697", "0.4287", "0.0801", "0.4287" ], [ "0.5024", "0.4153", "0.8906", "0.4153", "0.8906", "0.4287", "0.5024", "0.4287" ], [ "0.0784", "0.43", "0.4697", "0.4287", "0.4697", "0.443", "0.0784", "0.4443" ], [ "0.5024", "0.4312", "0.892", "0.4312", "0.892", "0.4443", "0.5024", "0.4443" ], [ "0.0801", "0.4458", "0.4697", "0.4458", "0.4697", "0.459", "0.0801", "0.459" ], [ "0.5024", "0.4458", "0.892", "0.4458", "0.892", "0.459", "0.5024", "0.459" ], [ "0.0809", "0.4622", "0.468", "0.4622", "0.468", "0.473", "0.0809", "0.473" ], [ "0.501", "0.4602", "0.892", "0.461", "0.892", "0.4749", "0.501", "0.4734" ], [ "0.0784", "0.476", "0.4697", "0.476", "0.4697", "0.4893", "0.0784", "0.4893" ], [ "0.5024", "0.476", "0.892", "0.476", "0.892", "0.4893", "0.5024", "0.4893" ], [ "0.0809", "0.4924", "0.4673", "0.4924", "0.4673", "0.5024", "0.0809", "0.5024" ], [ "0.5", "0.4912", "0.8906", "0.4905", "0.8906", "0.504", "0.5", "0.505" ], [ "0.0801", "0.506", "0.4697", "0.506", "0.4697", "0.5195", "0.0801", "0.5195" ], [ "0.5034", "0.5083", "0.8906", "0.5083", "0.8906", "0.518", "0.5034", "0.518" ], [ "0.0784", "0.521", "0.468", "0.521", "0.468", "0.535", "0.0784", "0.535" ], [ "0.5024", "0.5215", "0.892", "0.5215", "0.892", "0.535", "0.5024", "0.535" ], [ "0.0809", "0.537", "0.4673", "0.537", "0.4673", "0.5474", "0.0809", "0.5474" ], [ "0.5024", "0.536", "0.8906", "0.536", "0.8906", "0.55", "0.5024", "0.55" ], [ "0.0801", "0.5503", "0.468", "0.5503", "0.468", "0.5645", "0.0801", "0.5645" ], [ "0.5024", "0.5503", "0.8906", "0.5503", "0.8906", "0.5645", "0.5024", "0.5645" ], [ "0.0784", "0.565", "0.3857", "0.565", "0.3857", "0.579", "0.0784", "0.579" ], [ "0.5024", "0.5664", "0.892", "0.5664", "0.892", "0.58", "0.5024", "0.58" ], [ "0.505", "0.5835", "0.888", "0.5835", "0.888", "0.5933", "0.505", "0.5933" ], [ "0.0801", "0.5903", "0.3179", "0.5903", "0.3179", "0.605", "0.0801", "0.605" ], [ "0.5024", "0.5967", "0.694", "0.5967", "0.694", "0.61", "0.5024", "0.61" ], [ "0.0972", "0.6143", "0.4697", "0.6143", "0.4697", "0.6274", "0.0972", "0.6274" ], [ "0.6562", "0.6235", "0.773", "0.6255", "0.773", "0.6445", "0.6562", "0.6426" ], [ "0.794", "0.6235", "0.835", "0.6235", "0.835", "0.6396", "0.794", "0.6396" ], [ "0.0801", "0.63", "0.4697", "0.63", "0.4697", "0.643", "0.0801", "0.643" ], [ "0.536", "0.627", "0.5947", "0.629", "0.5947", "0.6426", "0.536", "0.64" ], [ "0.5923", "0.6333", "0.6313", "0.638", "0.6284", "0.6484", "0.589", "0.643" ], [ "0.871", "0.63", "0.888", "0.63", "0.888", "0.6396", "0.871", "0.6396" ], [ "0.0784", "0.643", "0.468", "0.6445", "0.468", "0.659", "0.0784", "0.6577" ], [ "0.6445", "0.6465", "0.661", "0.6465", "0.661", "0.651", "0.6445", "0.651" ], [ "0.774", "0.6445", "0.7837", "0.643", "0.785", "0.6484", "0.7754", "0.6494" ], [ "0.0801", "0.66", "0.468", "0.66", "0.468", "0.674", "0.0801", "0.674" ], [ "0.501", "0.659", "0.892", "0.661", "0.892", "0.6787", "0.501", "0.677" ], [ "0.0801", "0.675", "0.468", "0.675", "0.468", "0.688", "0.0801", "0.688" ], [ "0.5", "0.6753", "0.8936", "0.677", "0.8936", "0.6943", "0.5", "0.6934" ], [ "0.0801", "0.69", "0.4697", "0.69", "0.4697", "0.704", "0.0801", "0.704" ], [ "0.501", "0.6924", "0.8906", "0.6934", "0.8906", "0.7085", "0.501", "0.707" ], [ "0.0801", "0.7056", "0.4697", "0.7056", "0.4697", "0.7197", "0.0801", "0.7197" ], [ "0.5024", "0.709", "0.819", "0.709", "0.819", "0.7227", "0.5024", "0.7227" ], [ "0.0801", "0.72", "0.468", "0.72", "0.468", "0.7344", "0.0801", "0.7344" ], [ "0.0784", "0.7344", "0.468", "0.735", "0.468", "0.75", "0.0784", "0.7485" ], [ "0.5034", "0.7373", "0.8057", "0.7373", "0.8057", "0.752", "0.5034", "0.752" ], [ "0.0801", "0.7505", "0.4355", "0.7505", "0.4355", "0.7646", "0.0801", "0.7646" ], [ "0.0972", "0.765", "0.468", "0.7646", "0.468", "0.779", "0.0972", "0.78" ], [ "0.523", "0.761", "0.8906", "0.761", "0.8906", "0.7744", "0.523", "0.7744" ], [ "0.0801", "0.781", "0.468", "0.781", "0.468", "0.7944", "0.0801", "0.7944" ], [ "0.501", "0.7764", "0.892", "0.778", "0.892", "0.7915", "0.501", "0.7905" ], [ "0.0801", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0801", "0.81" ], [ "0.5024", "0.7935", "0.8906", "0.7935", "0.8906", "0.8066", "0.5024", "0.8066" ], [ "0.0801", "0.811", "0.4697", "0.811", "0.4697", "0.8247", "0.0801", "0.8247" ], [ "0.5024", "0.808", "0.892", "0.808", "0.892", "0.8213", "0.5024", "0.8213" ], [ "0.0784", "0.827", "0.4697", "0.827", "0.4697", "0.8403", "0.0784", "0.8403" ], [ "0.5024", "0.8237", "0.892", "0.8237", "0.892", "0.837", "0.5024", "0.837" ], [ "0.0801", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0801", "0.855" ], [ "0.5024", "0.8384", "0.892", "0.8384", "0.892", "0.8516", "0.5024", "0.8516" ], [ "0.0801", "0.857", "0.4697", "0.857", "0.4697", "0.8706", "0.0801", "0.8706" ], [ "0.5024", "0.853", "0.8896", "0.853", "0.8896", "0.866", "0.5024", "0.866" ], [ "0.0784", "0.872", "0.4697", "0.872", "0.4697", "0.8853", "0.0784", "0.8853" ], [ "0.517", "0.8735", "0.892", "0.875", "0.892", "0.8896", "0.517", "0.888" ], [ "0.0801", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0801", "0.901" ], [ "0.5024", "0.886", "0.5825", "0.888", "0.5825", "0.9014", "0.5024", "0.8994" ], [ "0.4795", "0.924", "0.492", "0.924", "0.492", "0.936", "0.4795", "0.936" ] ]

[ "a posteriori estimation (MAP):", "Generally,in variational BNNs, we introduce a variational", "distribution q(w0) with parameters and maximize the", "1n", "logp(yilai;w)", "log p(w),", "(1)", "evidence lower bound (ELBO) for learning (scaled by 1/n):", "max", "n", "", "max Eg(w0)", "DkL(q(w|0)/p(w))", "log p(yiai; w", "where p(ya; w)refers to the predictive distribution of the", "DNNmodel.By setting the prior p(w)as an isotropic", "(3)", "Gaussian, the second term amounts to the L2 (weight de-", "Inference. The obtained posterior q(w)? offers us", "cay)regularizerwithatunablecoefficientinoptimization.", "the opportunities to predict robustly.For computational", "Generally speaking, the adversarial example corresponding", "tractability, we usually estimate the posterior predictive via:", "to (aci, yi) against the model is defined as", "1", "1", "p(y|a, D) = Eg(wjo) [p(yla; w)) ~", "p(yla; w(t)), (4)", "(2)", "aadv = ri + arg minlogp(yilai+ &; w),", "8.ES", "T2", "t=", "where S = fs :ll < } is the valid perturbation set", "where w(t) ~ q(wo),t = 1, .., T denote the Monte Carlo", "with e > O as the perturbation budget and l-l as some norm", "(MC)samples.Inotherwords,theBNNassembles thepre-", "(e.g., l.).Extensive attack methods have been developed", "dictions yielded by all likely models to make more reliable", "with promise to solve the above minimization problem [19,", "and calibrated decisions, in stark contrast to the DNN which", "41,4,58l.basedongradientsornot.", "only cares about the most possible parameter point.", "The central goal of adversarial defense is to protect the", "Measure of uncertainty. For adversarial detection, we", "modelfrommakingundesirabledecisionsfortheadversar-", "are interested in the epistemic uncertainty which is indica-", "ial examples adv.A representative line of work approaches", "tive of covariate shift.A superior choice of uncertainty met-", "this objective by augmenting the training data with on-the-", "ric is the softmax variance given its previous success for ad-", "fly generated adversarial examples and forcing the model to", "versarial detection in image classification [14] and insight-", "yield correct predictions on them [42, 69]. But their limited", "ful theoretical support [54]. However, the softmax output", "trainingefficiencyandcompromisingperformanceonclean", "of themodel maybe less attractiveduring inference (e.g.,", "data pose a major obstacle for real-world adoption. As an", "in open-set face recognition), letting alone that not all the", "alternative.adversarial detection methods focus on distin-", "computer vision tasks can be formulated as pure classifi-", "guishing the adversarial examples from the normal ones so", "cationproblems (e.g.,object detection).Tomakethemet-", "as to bypass the potentially harmful outcomes of making", "ric faithful and readily applicable to diverse scenarios, we", "decisions for adversarial examples [44.5,40l.However.", "concern the predictive variance of the hidden feature z cor-", "satisfactory transferability to unseen attacks and tasks be-", "responding to c, by mildly assuming the information flow", "yond image classification remains elusive [39]", "inside the model as -→ z → y. We utilize an unbiased", "varianceestimatorand summarizethe varianceof all coor-", "3.2. Bayesian Neural Networks", "dinates of z into a scalar via:", "In essence, the problem of distinguishing adversarial", "() - T(I", "2(013", "examples from benign ones can be viewed as a special-", "U(c)=", "T_1", "(5)", "ized out-of-distribution (OOD) detection problem of partic-", "", "", "ular concern in safety-sensitive scenarios -- with the model", "where z(t) denotes the features of a under parameter sam-", "trained on the clean data, we expect to identify the adver-", "ple w(t) ~ q(w0),t = 1,., T, with IIll2 as l2 norm. It", "sarial examples from a shifted data manifold, though the", "is natural to simultaneously make prediction and quantify", "shift magnitude may be subtle and human-imperceptible. In", "uncertainty via Eq. (4) and Eq. (5) when testing.", "this sense, we naturally introduce BNNs into the picture at-", "tributed to their principled OOD detection capacity along", "4.Lightweight Bayesian Refinement", "with the equivalent flexibility for data fitting as DNNs.", "Modeling and training. Typically, a BNN is specified", "Despite their theoretical appealingness, BNNs are sel-", "by a parameter prior p(w) and an NN-instantiated data like-", "dom adopted for real-world adversarial detection, owing to", "lihood p(Dw). We are interested in the parameter posterior", "a wide range of concerns on their training efficiency, pre-", "p(w D) instead of a point estimate as in DNN. It is known", "dictive performance, quality of uncertainty estimates, and", "that precisely deriving the posterior is intractable owing to", "inference speed. In this section, we provide detailed and", "the high non-linearity of neural networks. Among the wide", "novel strategies to relieve these concerns and build the prac-", "spectrum of approximate Bayesian inference methods, vari-", "tical Lightweight Bayesian Refinement (LiBRe) framework.", "ational BNNs are particularly attractive due to their close re-", "2We use q(w|) equivalently with p(w|D) in the following if there is", "semblance to standard backprop [20, 2, 37, 55, 56, 53, 46].", "no misleading.", "3" ]

{ "arxiv_id": "2103.14835", "language": "en", "page_idx": 2 }

[ [ "0.0801", "0.09216", "0.468", "0.09216", "0.468", "0.1061", "0.0801", "0.1061" ], [ "0.511", "0.09784", "0.8096", "0.09784", "0.8096", "0.108", "0.511", "0.108" ], [ "0.0801", "0.108", "0.4697", "0.108", "0.4697", "0.1218", "0.0801", "0.1218" ], [ "0.5195", "0.11426", "0.846", "0.113", "0.846", "0.1276", "0.5195", "0.1288" ], [ "0.0801", "0.1237", "0.4697", "0.1237", "0.4697", "0.137", "0.0801", "0.137" ], [ "0.568", "0.13", "0.853", "0.13", "0.853", "0.1401", "0.568", "0.1401" ], [ "0.0784", "0.1364", "0.4697", "0.1383", "0.4697", "0.1528", "0.0784", "0.1509" ], [ "0.5073", "0.1427", "0.7812", "0.1427", "0.7812", "0.1572", "0.5073", "0.1572" ], [ "0.0809", "0.1547", "0.4673", "0.1547", "0.4673", "0.1654", "0.0809", "0.1654" ], [ "0.5073", "0.1578", "0.859", "0.1578", "0.859", "0.1718", "0.5073", "0.1718" ], [ "0.0784", "0.1666", "0.4697", "0.1674", "0.4697", "0.1818", "0.0784", "0.1812" ], [ "0.5254", "0.1704", "0.789", "0.1718", "0.789", "0.1901", "0.5254", "0.1895" ], [ "0.0801", "0.1831", "0.468", "0.1831", "0.468", "0.1976", "0.0801", "0.1976" ], [ "0.5073", "0.1901", "0.6772", "0.1901", "0.6772", "0.2039", "0.5073", "0.2039" ], [ "0.0801", "0.1989", "0.4478", "0.1989", "0.4478", "0.2122", "0.0801", "0.2122" ], [ "0.544", "0.2039", "0.8022", "0.2008", "0.8022", "0.2184", "0.544", "0.2216" ], [ "0.0907", "0.2197", "0.4697", "0.2216", "0.4697", "0.2424", "0.0907", "0.2405" ], [ "0.5728", "0.2203", "0.8203", "0.2203", "0.8203", "0.2311", "0.5728", "0.2311" ], [ "0.5728", "0.2329", "0.831", "0.2329", "0.831", "0.2468", "0.5728", "0.2468" ], [ "0.5107", "0.2393", "0.5146", "0.2374", "0.5186", "0.2424", "0.513", "0.2443" ], [ "0.0809", "0.2488", "0.468", "0.2488", "0.468", "0.262", "0.0809", "0.262" ], [ "0.5107", "0.25", "0.5186", "0.2507", "0.517", "0.2568", "0.509", "0.2556" ], [ "0.5728", "0.2468", "0.8545", "0.2468", "0.8545", "0.2615", "0.5728", "0.2615" ], [ "0.0784", "0.2632", "0.4697", "0.2644", "0.4697", "0.279", "0.0784", "0.2778" ], [ "0.572", "0.26", "0.863", "0.26", "0.863", "0.2734", "0.572", "0.2734" ], [ "0.572", "0.2734", "0.865", "0.2747", "0.865", "0.289", "0.572", "0.2886" ], [ "0.0801", "0.2798", "0.255", "0.2798", "0.255", "0.2935", "0.0801", "0.2935" ], [ "0.5107", "0.281", "0.517", "0.281", "0.517", "0.2842", "0.5107", "0.2842" ], [ "0.1748", "0.3037", "0.1896", "0.3037", "0.1896", "0.3113", "0.1748", "0.3113" ], [ "0.1953", "0.3005", "0.4714", "0.303", "0.4714", "0.324", "0.1953", "0.3213" ], [ "0.1013", "0.307", "0.1642", "0.305", "0.1642", "0.3193", "0.1013", "0.3213" ], [ "0.5024", "0.31", "0.8906", "0.3093", "0.8906", "0.324", "0.5024", "0.3245" ], [ "0.1642", "0.3164", "0.1936", "0.3164", "0.1936", "0.3276", "0.1642", "0.3276" ], [ "0.1912", "0.324", "0.2507", "0.324", "0.2507", "0.3352", "0.1912", "0.3352" ], [ "0.5024", "0.327", "0.8906", "0.327", "0.8906", "0.3403", "0.5024", "0.3403" ], [ "0.076", "0.3372", "0.4714", "0.3403", "0.4714", "0.361", "0.076", "0.358" ], [ "0.5024", "0.3416", "0.892", "0.3416", "0.892", "0.3547", "0.5024", "0.3547" ], [ "0.076", "0.3547", "0.4714", "0.3547", "0.4714", "0.379", "0.076", "0.379" ], [ "0.5024", "0.3567", "0.892", "0.3567", "0.892", "0.3706", "0.5024", "0.3706" ], [ "0.5", "0.3694", "0.892", "0.3694", "0.892", "0.387", "0.5", "0.387" ], [ "0.0784", "0.377", "0.4673", "0.3757", "0.4673", "0.3901", "0.0784", "0.3914" ], [ "0.0801", "0.3926", "0.468", "0.3926", "0.468", "0.406", "0.0801", "0.406" ], [ "0.5024", "0.3894", "0.892", "0.3894", "0.892", "0.4028", "0.5024", "0.4028" ], [ "0.0809", "0.4092", "0.4631", "0.4092", "0.4631", "0.4192", "0.0809", "0.4192" ], [ "0.5024", "0.404", "0.892", "0.404", "0.892", "0.4172", "0.5024", "0.4172" ], [ "0.0989", "0.4224", "0.468", "0.4224", "0.468", "0.4363", "0.0989", "0.4363" ], [ "0.505", "0.4204", "0.8906", "0.4204", "0.8906", "0.4312", "0.505", "0.4312" ], [ "0.4983", "0.4287", "0.8945", "0.428", "0.8945", "0.449", "0.4983", "0.4495" ], [ "0.0801", "0.4382", "0.4697", "0.4382", "0.4697", "0.452", "0.0801", "0.452" ], [ "0.501", "0.4443", "0.8936", "0.4463", "0.8936", "0.464", "0.501", "0.4622" ], [ "0.0801", "0.4526", "0.4697", "0.4526", "0.4697", "0.4666", "0.0801", "0.4666" ], [ "0.501", "0.4622", "0.79", "0.4634", "0.79", "0.4778", "0.501", "0.4766" ], [ "0.7812", "0.4622", "0.884", "0.459", "0.884", "0.473", "0.7812", "0.476" ], [ "0.0801", "0.4685", "0.468", "0.4685", "0.468", "0.4817", "0.0801", "0.4817" ], [ "0.5", "0.4766", "0.8936", "0.4749", "0.8936", "0.4924", "0.5", "0.4944" ], [ "0.0784", "0.483", "0.4697", "0.4817", "0.4697", "0.4968", "0.0784", "0.4976" ], [ "0.0784", "0.4968", "0.4697", "0.4988", "0.4697", "0.5166", "0.0784", "0.5146" ], [ "0.5034", "0.4944", "0.8906", "0.4944", "0.8906", "0.5083", "0.5034", "0.5083" ], [ "0.501", "0.5103", "0.8906", "0.5083", "0.8906", "0.523", "0.501", "0.5244" ], [ "0.0801", "0.5166", "0.468", "0.5166", "0.468", "0.53", "0.0801", "0.53" ], [ "0.5", "0.5234", "0.8936", "0.5244", "0.8936", "0.5425", "0.5", "0.5415" ], [ "0.0784", "0.531", "0.468", "0.53", "0.468", "0.5435", "0.0784", "0.5444" ], [ "0.0801", "0.547", "0.3865", "0.547", "0.3865", "0.56", "0.0801", "0.56" ], [ "0.501", "0.5425", "0.8906", "0.5405", "0.8906", "0.5537", "0.501", "0.556" ], [ "0.0989", "0.56", "0.4673", "0.5596", "0.4673", "0.5737", "0.0989", "0.5747" ], [ "0.501", "0.5566", "0.892", "0.556", "0.892", "0.571", "0.501", "0.5713" ], [ "0.0801", "0.5747", "0.4697", "0.5747", "0.4697", "0.5923", "0.0801", "0.5923" ], [ "0.5024", "0.5728", "0.892", "0.5728", "0.892", "0.586", "0.5024", "0.586" ], [ "0.0801", "0.5933", "0.468", "0.5933", "0.468", "0.607", "0.0801", "0.607" ], [ "0.5024", "0.5884", "0.892", "0.5884", "0.892", "0.6016", "0.5024", "0.6016" ], [ "0.0784", "0.607", "0.468", "0.607", "0.468", "0.6206", "0.0784", "0.6206" ], [ "0.5024", "0.6016", "0.689", "0.603", "0.689", "0.616", "0.5024", "0.6157" ], [ "0.0801", "0.6226", "0.468", "0.6226", "0.468", "0.636", "0.0801", "0.636" ], [ "0.5024", "0.629", "0.6313", "0.631", "0.6313", "0.646", "0.5024", "0.643" ], [ "0.0801", "0.638", "0.4697", "0.638", "0.4697", "0.6514", "0.0801", "0.6514" ], [ "0.0801", "0.653", "0.4697", "0.653", "0.4697", "0.6665", "0.0801", "0.6665" ], [ "0.5195", "0.6514", "0.892", "0.653", "0.892", "0.6675", "0.5195", "0.6665" ], [ "0.0784", "0.6685", "0.468", "0.6675", "0.468", "0.681", "0.0784", "0.6816" ], [ "0.501", "0.6675", "0.8906", "0.6675", "0.8906", "0.681", "0.501", "0.681" ], [ "0.0809", "0.685", "0.468", "0.685", "0.468", "0.696", "0.0809", "0.696" ], [ "0.5024", "0.6846", "0.892", "0.6846", "0.892", "0.6978", "0.5024", "0.6978" ], [ "0.0801", "0.6987", "0.4714", "0.6987", "0.4714", "0.712", "0.0801", "0.712" ], [ "0.5024", "0.6987", "0.892", "0.6987", "0.892", "0.712", "0.5024", "0.712" ], [ "0.0801", "0.714", "0.4656", "0.714", "0.4656", "0.728", "0.0801", "0.728" ], [ "0.501", "0.714", "0.88", "0.714", "0.88", "0.728", "0.501", "0.728" ], [ "0.0972", "0.728", "0.4697", "0.73", "0.4697", "0.7437", "0.0972", "0.741" ], [ "0.5215", "0.732", "0.8906", "0.732", "0.8906", "0.7456", "0.5215", "0.7456" ], [ "0.0801", "0.7437", "0.4697", "0.7437", "0.4697", "0.757", "0.0801", "0.757" ], [ "0.5024", "0.7476", "0.892", "0.7476", "0.892", "0.761", "0.5024", "0.761" ], [ "0.0801", "0.759", "0.4697", "0.759", "0.4697", "0.773", "0.0801", "0.773" ], [ "0.5024", "0.762", "0.892", "0.762", "0.892", "0.776", "0.5024", "0.776" ], [ "0.0784", "0.7734", "0.4697", "0.7734", "0.4697", "0.787", "0.0784", "0.787" ], [ "0.5034", "0.779", "0.8906", "0.779", "0.8906", "0.789", "0.5034", "0.789" ], [ "0.0801", "0.789", "0.468", "0.789", "0.468", "0.803", "0.0801", "0.803" ], [ "0.5034", "0.7935", "0.892", "0.7935", "0.892", "0.8066", "0.5034", "0.8066" ], [ "0.0801", "0.8037", "0.468", "0.8037", "0.468", "0.8174", "0.0801", "0.8174" ], [ "0.5024", "0.808", "0.892", "0.808", "0.892", "0.8213", "0.5024", "0.8213" ], [ "0.0784", "0.8193", "0.4697", "0.8184", "0.4697", "0.832", "0.0784", "0.8335" ], [ "0.5024", "0.8237", "0.8906", "0.8237", "0.8906", "0.837", "0.5024", "0.837" ], [ "0.0809", "0.836", "0.468", "0.836", "0.468", "0.8467", "0.0809", "0.8467" ], [ "0.501", "0.8384", "0.8906", "0.8384", "0.8906", "0.8516", "0.501", "0.8516" ], [ "0.0801", "0.8496", "0.468", "0.8496", "0.468", "0.863", "0.0801", "0.863" ], [ "0.5024", "0.854", "0.822", "0.854", "0.822", "0.867", "0.5024", "0.867" ], [ "0.0809", "0.866", "0.468", "0.866", "0.468", "0.877", "0.0809", "0.877" ], [ "0.5215", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5215", "0.8853" ], [ "0.0972", "0.887", "0.281", "0.887", "0.281", "0.901", "0.0972", "0.901" ], [ "0.5024", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5024", "0.901" ], [ "0.4795", "0.925", "0.4927", "0.925", "0.4927", "0.936", "0.4795", "0.936" ] ]

[ "limit the learned models from generalizing to unseen at-", "Algorithml:LightweightBayesianRefinement", "tacks. Instead, we propose an adversarial example free un-", "Input: pre-trained DNN parameters wt, weight decay", "certainty correction strategy by considering a superset of the", "coefficient入,threshold,trade-offcoefficienta", "adversarial examples.We feed uniformly perturbed training", "1 Initialize [w,ni and w based on wt", "instances (whichencompass allkinds ofadversarial exam-", "2Build optimizers opt, and opt_6 with weight decay /C", "ples) into the BNN and demand relatively high predictive", "and A for (w(y-1 and w) respectively", "uncertainty on them. Formally, with etrain as the training", "3 for epoch = I, 2, ., E do", "perturbation budget, we perturb a mini-batch of data via", "for mini-batch B = ((ri, y.)l= in D do", "ai= ai + i,, id u(-train,Etrain),i=l,. B]. (9)", "Estimate the log-likelihood C*viaEq.(8)", "Uniformly perturb the clean data via Eq. (9)", "", "Then we calculate the uncertainty measure U cheaply with", "7", "Estimate the uncertainty penalty R via Eq. (10)", "T = 2 MC samples, and regularize the outcome via solving", "Backward the gradients of C* + αR via autodiff", "Perform 1-step gradient ascent with opt, & opt_b", "the following margin loss:", "", "1", "min(/≥(,1) = ≥(e2)|/2,),(10)", "maxR =", "we expect to take all the C parameter candidates into ac-", "IBI", "(ai-yi)EB", "count for prediction to thoroughly exploit their heteroge-", "where z(c.) refers to the features of a; given parameter", "neous predictive behaviour, i.e., T = C. Naively sequen-", "sample w() = (we), w, with,dunif[1,)", "tially calculating the outcomes under each parameter can-", "didate w(e) is viable, but we can achieve further speedup", "and ci1 + ci2,i = 1,..,JB,j = 1,2. is a tun-", "able threshold. Surprisingly, this regularization remarkably", "by unleashing the potential of parallel computing.Take the", "booststheadversarialdetectionperformance(seeSec5.4)", "convolution layer in the front of the Bayesian sub-module as", "Efficient learning by refining pre-trained DNNs.", "anexample(weabuse somenotationshere):Givenabatch", "of features ain E Rbxixhxw and C convolution kernels", "Though from-scratch BNN training is feasible, a recent", "w(c) e Roxixkxk,c = 1, ., C, we first repeat acin at the", "work demonstrate that it probably incurs worse predic-", "channel dimension for C times, getting cn", "=RbxCixhxu", "tive performance than a fairly trained DNN [61]. There-", "and concatenate [w(c)j-1 as w' e IRCoxixkxk. Then,", "fore, given the alignment between the posterior parameters", " = {w(), w(l),., w(C)] and their DNN counterparts,", "we estimate the outcomes in parallel via group convolution:", "aout = conv(αin, w,groups = C), and the outcome cor-", "we suggest to perform cost-effective Bayesian refinement", "responding to w(e) is a() = aou[;, co - o : co, .]. The", "upon a pre-trained DNN model, which renders our work-", "flow more appropriate for large-scale learning.", "cooperation between FADEvariational and the above strat-", "With the pre-training DNN parameters denoted as wt ", "egy makes our inference time close to that of the DNNs in", "[wt, wthl, we initialize wo) as wl, and w(@) as w, for", "the same setting (see Sec 5.4), while only our approach en-", "c = 1, .., C. Continuing from this, we fine-tune the vari-", "joys the benefits from Bayes principle and is able to achieve", "ational parameters to maximize C* + R4 under weight", "robust adversarial detection.", "decay regularizers with suitable coefficients to realise ad-", "5.Experiments", "versarial detection-oriented posterior inference. The whole", "algorithmic procedure is presented in Algorithm 1. Such", "To verify if LiBRE couldquicklyandeconomically equip", "a practical and economical refinement significantlybenefits", "the pre-trained DNNs with principled adversarial detection", "fromtheprevalenceof open-sourceDNNmodelzoo,andis", "ability in various scenarios, we perform extensive empirical", "promised to maintain non-degraded predictive performance", "studies covering ImageNet classification [8], open-set face", "by the well-evaluated pre-training & fine-tuning workflow.", "recognition [66], and object detection [35] in this section.", "Inference speedup.After learning,a wide criticism on", "General setup. We fetch the pre-trained DNNs avail-", "BNNs is their requirement for longer inference time than", "able online, and inherit all their settings for the Bayesian", "DNNs. This is because BNNs leverage a collection of MC", "refinement unless otherwise stated. We use C - 20 can-", "samples to marginalize the posterior for prediction and un-", "didatesforFADEacrossscenarios.TheFADEposterioris", "certainty quantification, as shown in Eq. (4) and Eq. (5).", "generally employed for the parameters of the last convolu-", "However, such a problem is desirably alleviated in our ap-", "tion block (e.g., the last residual block for ImageNet and", "proach thanks to the adoption of the FADE variational. The", "face tasks or the feature output heads for object detection).", "mainpart ofthemodel remains deterministic,allowing us", " We take the immediate output of the Bayesian sub-module", "to perform only once forward propagation to reach the entry", "as z for estimating feature variance uncertainty.", "of the Bayesian sub-module.In the Bayesian sub-module,", "Attacks. We adopt some popular attacks to craft ad-", "4α refers to a trade-off coefficient.", "versarial examples under l2 and lo threat models, includ-", "5" ]

{ "arxiv_id": "2103.14835", "language": "en", "page_idx": 4 }

[ [ "0.1724", "0.03284", "0.4795", "0.0322", "0.4795", "0.04672", "0.1724", "0.04736" ], [ "0.1748", "0.09973", "0.8228", "0.101", "0.8228", "0.1193", "0.1748", "0.1187" ], [ "0.1748", "0.127", "0.763", "0.1276", "0.763", "0.1459", "0.1748", "0.1445" ], [ "0.1847", "0.1704", "0.8037", "0.1704", "0.8037", "0.1849", "0.1847", "0.1849" ], [ "0.183", "0.1863", "0.5605", "0.1863", "0.5605", "0.1995", "0.183", "0.1995" ], [ "0.1847", "0.2026", "0.6953", "0.2026", "0.6953", "0.2172", "0.1847", "0.2172" ], [ "0.183", "0.2197", "0.3652", "0.2203", "0.3652", "0.2349", "0.183", "0.2343" ], [ "0.452", "0.2524", "0.546", "0.2524", "0.546", "0.2676", "0.452", "0.2676" ], [ "0.02287", "0.2664", "0.06046", "0.2664", "0.06046", "0.4028", "0.02287", "0.4028" ], [ "0.2336", "0.281", "0.767", "0.281", "0.767", "0.2942", "0.2336", "0.2942" ], [ "0.232", "0.2942", "0.7656", "0.2942", "0.7656", "0.308", "0.232", "0.308" ], [ "0.2336", "0.3093", "0.7656", "0.3093", "0.7656", "0.3225", "0.2336", "0.3225" ], [ "0.2336", "0.3225", "0.7656", "0.3225", "0.7656", "0.336", "0.2336", "0.336" ], [ "0.2336", "0.336", "0.7656", "0.336", "0.7656", "0.3499", "0.2336", "0.3499" ], [ "0.2345", "0.3516", "0.7646", "0.3516", "0.7646", "0.3623", "0.2345", "0.3623" ], [ "0.2336", "0.3643", "0.7656", "0.3643", "0.7656", "0.3774", "0.2336", "0.3774" ], [ "0.2345", "0.3774", "0.767", "0.3774", "0.767", "0.3914", "0.2345", "0.3914" ], [ "0.0196", "0.3958", "0.05963", "0.3945", "0.0621", "0.4849", "0.02287", "0.486" ], [ "0.232", "0.3901", "0.767", "0.3914", "0.767", "0.406", "0.232", "0.4048" ], [ "0.232", "0.406", "0.7656", "0.406", "0.7656", "0.4192", "0.232", "0.4192" ], [ "0.2336", "0.4192", "0.7656", "0.4192", "0.7656", "0.433", "0.2336", "0.433" ], [ "0.232", "0.4343", "0.7656", "0.4343", "0.7656", "0.4475", "0.232", "0.4475" ], [ "0.2336", "0.4475", "0.7656", "0.4475", "0.7656", "0.461", "0.2336", "0.461" ], [ "0.2345", "0.461", "0.7656", "0.461", "0.7656", "0.4749", "0.2345", "0.4749" ], [ "0.2345", "0.4766", "0.763", "0.4766", "0.763", "0.4873", "0.2345", "0.4873" ], [ "0.02287", "0.4873", "0.05963", "0.4873", "0.05963", "0.7056", "0.02287", "0.7056" ], [ "0.1814", "0.5146", "0.1887", "0.5195", "0.183", "0.5244", "0.1748", "0.518" ], [ "0.2009", "0.513", "0.3367", "0.513", "0.3367", "0.528", "0.2009", "0.528" ], [ "0.1748", "0.5405", "0.824", "0.5405", "0.824", "0.5537", "0.1748", "0.5537" ], [ "0.1748", "0.5537", "0.824", "0.5537", "0.824", "0.5684", "0.1748", "0.5684" ], [ "0.1748", "0.5684", "0.824", "0.5684", "0.824", "0.582", "0.1748", "0.582" ], [ "0.1732", "0.5835", "0.824", "0.5835", "0.824", "0.5967", "0.1732", "0.5967" ], [ "0.1748", "0.5967", "0.824", "0.5967", "0.824", "0.61", "0.1748", "0.61" ], [ "0.1748", "0.61", "0.824", "0.61", "0.824", "0.6235", "0.1748", "0.6235" ], [ "0.1732", "0.6235", "0.824", "0.6226", "0.824", "0.637", "0.1732", "0.638" ], [ "0.1732", "0.636", "0.7563", "0.637", "0.7563", "0.6514", "0.1732", "0.651" ], [ "0.1732", "0.657", "0.824", "0.6577", "0.824", "0.6724", "0.1732", "0.672" ], [ "0.1732", "0.6724", "0.826", "0.672", "0.826", "0.686", "0.1732", "0.6875" ], [ "0.1748", "0.686", "0.826", "0.686", "0.826", "0.6997", "0.1748", "0.6997" ], [ "0.1748", "0.7007", "0.824", "0.7007", "0.824", "0.714", "0.1748", "0.714" ], [ "0.1748", "0.714", "0.824", "0.714", "0.824", "0.728", "0.1748", "0.728" ], [ "0.1748", "0.728", "0.824", "0.728", "0.824", "0.741", "0.1748", "0.741" ], [ "0.1748", "0.742", "0.826", "0.742", "0.826", "0.756", "0.1748", "0.756" ], [ "0.1732", "0.754", "0.826", "0.756", "0.826", "0.771", "0.1732", "0.768" ], [ "0.1732", "0.768", "0.8267", "0.7695", "0.8267", "0.7827", "0.1732", "0.782" ], [ "0.1748", "0.784", "0.826", "0.784", "0.826", "0.7974", "0.1748", "0.7974" ], [ "0.1748", "0.7974", "0.824", "0.7974", "0.824", "0.811", "0.1748", "0.811" ], [ "0.1748", "0.811", "0.824", "0.811", "0.824", "0.8247", "0.1748", "0.8247" ], [ "0.1732", "0.8257", "0.785", "0.8257", "0.785", "0.839", "0.1732", "0.839" ], [ "0.1748", "0.845", "0.8228", "0.845", "0.8228", "0.86", "0.1748", "0.86" ], [ "0.1765", "0.8623", "0.8228", "0.8623", "0.8228", "0.8726", "0.1765", "0.8726" ], [ "0.1748", "0.875", "0.8228", "0.875", "0.8228", "0.888", "0.1748", "0.888" ], [ "0.1732", "0.888", "0.824", "0.888", "0.824", "0.903", "0.1732", "0.903" ], [ "0.1969", "0.911", "0.653", "0.911", "0.653", "0.9253", "0.1969", "0.9253" ], [ "0.4927", "0.9497", "0.505", "0.9497", "0.505", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2023", "MAESTRO: OPEN-ENDED ENVIRONMENT DESIGN", "FOR MULTI-AGENT REINFORCEMENT LEARNING", "Mikayel Samvelyan12 Akbir Khan? Michael Dennis3 Minqi Jiang12 Jack Parker-Holder4", "Jakob Foerster Roberta Raileanu Tim Rocktaschel?", "1Meta AI 2University College London 3UC Berkeley 4University of Oxford", "samvelyan@meta.com", "ABSTRACT", "", "Open-ended learning methods that automatically generate a curriculum of increas-", "ingly challenging tasks serve as a promising avenue toward generally capable", "reinforcement learning agents.Existing methods adapt curricula independently", "over either environment parameters (in single-agent settings) or co-player policies", "(in multi-agent settings). However, the strengths and weaknesses of co-players can", "manifestthemselves differentlydepending on environmental features.It isthus", "crucial to consider the dependency between the environment and co-player when", "shaping a curriculum in multi-agent domains. In this work, we use this insight", "[OT's]", "and extend Unsupervised Environment Design (UED) to multi-agent environments.", "We then introduce Multi-Agent Environment Design Strategist for Open-Ended", "Learning (MAESTRO), the first multi-agent UED approach for two-player zero-sum", "settings. MAESTRO efficiently produces adversarial, joint curricula over both envi-", "ronments and co-players and attains minimax-regret guarantees at Nash equilibrium", "Our experiments show that MAESTRO outperforms a number of strong baselines on", "competitivetwo-playergames,spanningdiscreteandcontinuous controlsettings.", "1AL00507AE", "", "INTRODUCTION", "The past few years have seen a series of remarkable achievements in producing deep reinforcement", "learning (RL) agents with expert (Vinyals et al., 2019; Berner et al., 2019; Wurman et al., 2022) and", "superhuman (Silver et al., 2016; Schrittwieser et al., 2020) performance in challenging competitive", "games. Central to these successes are adversarial training processes that result in curricula creating", "new challenges at the frontier of an agent's capabilities (Leibo et al., 2019; Yang et al., 2021). Such", "automatic curricula, or autocurricula, can improve the sample efficiency and generality of trained", "policies (Open Ended Learning Team et al., 2021), as well as induce an open-ended learning process", "(Balduzzi et al., 2019; Stanley et al., 2017) that continues to endlessly robustify an agent.", "Autocurricula have been effective in multi-agent RL for adapting to different co-players in competitive", "games (Leibo et al., 2019; Garnelo et al., 2021; Baker et al., 2019; Bansal et al., 2018; Feng et al.,", "2021), where it is crucial to play against increasingly stronger opponents (Silver et al., 2018) and", "avoid being exploited by other agents (Vinyals et al., 2019). Here, algorithms such as self-play (Silver", "et al., 2018; Tesauro, 1995) and fictitious self-play (Brown, 1951; Heinrich et al., 2015) have proven", "especially effective. Similarly, in single-agent RL, autocurricula methods based on Unsupervised", "Environment Design (UED, Dennis et al., 2020) have proven effective in producing agents robust to a", "wide distribution of environments (Wang et al., 2019; 2020; Jiang et al., 2021a; Parker-Holder et al.", "2022).UED seeks to adapt distributions over environments to maximise some metrics of interest.", "Minimax-regret UED seeks to maximise the regret of the learning agent, viewing this process as a", "game between a teacher that proposes challenging environments and a student that learns to solve", "them. At a Nash equilibrium of such games, the student policy provably reaches a minimax-regret", "policy over the set of possible environments, thereby providing a strong robustness guarantee", "However, prior works in UED focus on single-agent RL and do not address the dependency between", "theenvironmentandthestrategies of otheragentswithin it.Inmulti-agent domains,thebehaviourof", "other agents plays a critical role in modulating the complexity and diversity of the challenges faced by", "a learning agent. For example, an empty environment that has no blocks to hide behind might be most", "'Videos of MAEsTRO agents are available at maest ro.samvelyan. com", "1" ]

{ "arxiv_id": "2303.03376", "language": "en", "page_idx": 0 }

[ [ "0.1732", "0.03284", "0.4795", "0.0322", "0.4795", "0.04672", "0.1732", "0.04736" ], [ "0.5317", "0.1016", "0.5605", "0.1016", "0.5605", "0.10986", "0.5317", "0.10986" ], [ "0.58", "0.1048", "0.5933", "0.1048", "0.5933", "0.1092", "0.58", "0.1092" ], [ "0.4968", "0.1225", "0.5215", "0.1225", "0.5215", "0.127", "0.4968", "0.127" ], [ "0.556", "0.11743", "0.584", "0.11743", "0.584", "0.125", "0.556", "0.125" ], [ "0.3677", "0.125", "0.38", "0.125", "0.38", "0.1288", "0.3677", "0.1288" ], [ "0.4265", "0.1559", "0.4575", "0.1559", "0.4575", "0.1642", "0.4265", "0.1642" ], [ "0.492", "0.1559", "0.5146", "0.1559", "0.5146", "0.1635", "0.492", "0.1635" ], [ "0.585", "0.1559", "0.615", "0.1559", "0.615", "0.1635", "0.585", "0.1635" ], [ "0.6543", "0.1547", "0.7075", "0.1547", "0.7075", "0.1666", "0.6543", "0.1666" ], [ "0.492", "0.1642", "0.517", "0.1642", "0.517", "0.1718", "0.492", "0.1718" ], [ "0.2646", "0.1578", "0.3284", "0.1578", "0.3284", "0.1692", "0.2646", "0.1692" ], [ "0.569", "0.1654", "0.5923", "0.1654", "0.5923", "0.1692", "0.569", "0.1692" ], [ "0.66", "0.1666", "0.6665", "0.1666", "0.6665", "0.1704", "0.66", "0.1704" ], [ "0.6665", "0.1666", "0.6763", "0.1666", "0.6763", "0.1704", "0.6665", "0.1704" ], [ "0.4182", "0.1674", "0.4429", "0.1724", "0.4395", "0.1812", "0.415", "0.1755" ], [ "0.2695", "0.1686", "0.3218", "0.1686", "0.3218", "0.1787", "0.2695", "0.1787" ], [ "0.2793", "0.1932", "0.33", "0.1932", "0.33", "0.2039", "0.2793", "0.2039" ], [ "0.2793", "0.2026", "0.3342", "0.2026", "0.3342", "0.2134", "0.2793", "0.2134" ], [ "0.2778", "0.2311", "0.3423", "0.2311", "0.3423", "0.2424", "0.2778", "0.2424" ], [ "0.2769", "0.2405", "0.3145", "0.2405", "0.3145", "0.2507", "0.2769", "0.2507" ], [ "0.4697", "0.2405", "0.4944", "0.2405", "0.4944", "0.2474", "0.4697", "0.2474" ], [ "0.1765", "0.2715", "0.822", "0.2715", "0.822", "0.2822", "0.1765", "0.2822" ], [ "0.1748", "0.2827", "0.824", "0.2827", "0.824", "0.2974", "0.1748", "0.2974" ], [ "0.1732", "0.2974", "0.826", "0.2966", "0.826", "0.3113", "0.1732", "0.3125" ], [ "0.1748", "0.3125", "0.8037", "0.3125", "0.8037", "0.3257", "0.1748", "0.3257" ], [ "0.1748", "0.3423", "0.824", "0.3423", "0.824", "0.3567", "0.1748", "0.3567" ], [ "0.1748", "0.3567", "0.824", "0.3567", "0.824", "0.3706", "0.1748", "0.3706" ], [ "0.1748", "0.3706", "0.824", "0.3706", "0.824", "0.3838", "0.1748", "0.3838" ], [ "0.1732", "0.3833", "0.826", "0.3838", "0.826", "0.3984", "0.1732", "0.3977" ], [ "0.1748", "0.3984", "0.8203", "0.3984", "0.8203", "0.4124", "0.1748", "0.4124" ], [ "0.1748", "0.4192", "0.824", "0.4192", "0.824", "0.433", "0.1748", "0.433" ], [ "0.1732", "0.4343", "0.8228", "0.4343", "0.8228", "0.4475", "0.1732", "0.4475" ], [ "0.1748", "0.4475", "0.824", "0.4475", "0.824", "0.461", "0.1748", "0.461" ], [ "0.1732", "0.4602", "0.824", "0.461", "0.824", "0.476", "0.1732", "0.4749" ], [ "0.1748", "0.476", "0.8228", "0.476", "0.8228", "0.4893", "0.1748", "0.4893" ], [ "0.1748", "0.4893", "0.824", "0.4893", "0.824", "0.5024", "0.1748", "0.5024" ], [ "0.1748", "0.5024", "0.824", "0.5024", "0.824", "0.5166", "0.1748", "0.5166" ], [ "0.1748", "0.5166", "0.824", "0.5166", "0.824", "0.53", "0.1748", "0.53" ], [ "0.1748", "0.531", "0.826", "0.531", "0.826", "0.5444", "0.1748", "0.5444" ], [ "0.1748", "0.5444", "0.796", "0.5444", "0.796", "0.558", "0.1748", "0.558" ], [ "0.1748", "0.565", "0.824", "0.565", "0.824", "0.579", "0.1748", "0.579" ], [ "0.1748", "0.58", "0.826", "0.58", "0.826", "0.5933", "0.1748", "0.5933" ], [ "0.1748", "0.5933", "0.824", "0.5933", "0.824", "0.607", "0.1748", "0.607" ], [ "0.1748", "0.607", "0.824", "0.607", "0.824", "0.6206", "0.1748", "0.6206" ], [ "0.1748", "0.622", "0.824", "0.622", "0.824", "0.6353", "0.1748", "0.6353" ], [ "0.1748", "0.6353", "0.824", "0.6353", "0.824", "0.6484", "0.1748", "0.6484" ], [ "0.1732", "0.6484", "0.8228", "0.6484", "0.8228", "0.662", "0.1732", "0.662" ], [ "0.1748", "0.6636", "0.824", "0.6636", "0.824", "0.677", "0.1748", "0.677" ], [ "0.1748", "0.677", "0.826", "0.677", "0.826", "0.69", "0.1748", "0.69" ], [ "0.1732", "0.6895", "0.675", "0.6895", "0.675", "0.7026", "0.1732", "0.7026" ], [ "0.1732", "0.7104", "0.824", "0.7104", "0.824", "0.7236", "0.1732", "0.7236" ], [ "0.1748", "0.726", "0.826", "0.726", "0.826", "0.7393", "0.1748", "0.7393" ], [ "0.1732", "0.7393", "0.826", "0.7393", "0.826", "0.7524", "0.1732", "0.7524" ], [ "0.1732", "0.754", "0.8267", "0.7524", "0.8267", "0.7666", "0.1732", "0.7676" ], [ "0.1732", "0.7666", "0.8228", "0.7666", "0.8228", "0.78", "0.1732", "0.78" ], [ "0.1748", "0.781", "0.824", "0.781", "0.824", "0.7944", "0.1748", "0.7944" ], [ "0.1748", "0.7954", "0.824", "0.7954", "0.824", "0.8096", "0.1748", "0.8096" ], [ "0.1748", "0.8096", "0.576", "0.8096", "0.576", "0.8228", "0.1748", "0.8228" ], [ "0.1772", "0.845", "0.2018", "0.845", "0.2018", "0.854", "0.1772", "0.854" ], [ "0.2018", "0.8433", "0.574", "0.8433", "0.574", "0.857", "0.2018", "0.857" ], [ "0.1765", "0.8706", "0.822", "0.8706", "0.822", "0.881", "0.1765", "0.881" ], [ "0.1748", "0.8833", "0.824", "0.8833", "0.824", "0.8965", "0.1748", "0.8965" ], [ "0.1748", "0.898", "0.824", "0.898", "0.824", "0.911", "0.1748", "0.911" ], [ "0.1748", "0.911", "0.826", "0.911", "0.826", "0.925", "0.1748", "0.925" ], [ "0.4927", "0.9487", "0.5063", "0.9487", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2023", "Stude", "der", "valua", ".·", "en", "Sample", "Updal", "iodically", "Prioritise high", "s buffe", "Environment", "etore", "90", "re", "ovalua", "Generator", "Co-Player", "Population", "Environment", "Buffers", "0", "Figurel:AdiagramofMAEsTRo.MAEsTROmaintains apopulationof co-players,eachhaving", "an individual buffer of high-regret environments. When new environments are sampled, the student's", "regret is calculated with respect to the corresponding co-player and added to the co-player's buffer.", "MAESTRO continually provides high-regret environment/co-player pairs for training the student.", "challenging when playing against opponent policies that attack head-on, whereas environments that", "are full of winding hallways might be difficult when playing against defensive policies. Robust RL", "agents should be expected to interact successfully with a wide assortment of other rational agents in", "their environment (Yang et al., 2021; Mahajan et al., 2022). Therefore, to become widely applicable,", "UED must be extended to include multi-agent dynamics as part of the environment design process.", "We formalise this novel problem as an Underspecified Partially-Observable Stochastic Game (UP", "OSG), which generalises UED to multi-agent settings. We then introduce Multi-Agent Environment", "Design Strategist for Open-Ended Learning (MAESTRO), the first approach to train generally capable", "agents in two-player UPOSGs such that they are robust to changes in the environment and co-player", "policies. MAEsTRO is a replay-guided approach that explicitly considers the dependence between", "agents and environments by jointly sampling over environment/co-player pairs using a regret-based", "curriculum and population learning (see Figure 1). In partially observable two-player zero-sum", "games, we show that at equilibrium, the MAEsTRO student policy reaches a Bayes-Nash Equilibrium", "with respect to a regret-maximising distribution over environments.Furthermore,infullyobservable", "settings, it attains a Nash-Equilibrium policy in every environment against every rational agent.", "We assess the curricula induced by MAEsTRO and a variety of strong baselines in two competitive", "two-player games, namely a sparse-reward grid-based LaserTag environment with discrete actions", "(Lanctot et al., 2017) and a dense-reward pixel-based MultiCarRacing environment with continuous", "actions (Schwarting et al., 2021). In both cases, MAESTRO produces more robust agents than baseline", "autocurriculum methods on out-of-distribution (OOD) human-designed environment instances against", "unseen co-players. Furthermore, we show that MAEsTRo agents, trained only on randomised", "environments and having never seen the target task, can significantly outperform specialist agents", "trained directly on the target environment. Moreover, in analysing how the student's regret varies", "across environments and co-players, we find that a joint curriculum, as produced by MAEsTRo, is", "indeed required for finding the highest regret levels, as necessitated by UED.", "In summary, we make the following core contributions: (i) we provide the first formalism for multi-", "agent learning in underspecified environments, (ii) we introduce MAEsTRo, a novel approach to", "jointly learn autocurricula over environment/co-player pairs, implicitly modelling their dependence,", "(i) we prove MAEsTRo inherits the theoretical property from the single-agent setting of imple-", "menting a minimax-regret policy at equilibrium, which corresponds to a Bayesian Nash or Nash", "equilibrium in certain settings, and (iv) by rigorously analysing the curriculum induced by MAESTRO", "and evaluating MAEsTRO agents against strong baselines, we empirically demonstrate the importance", "of the joint curriculum over the environments and co-players.", "2", "PROBLEM STATEMENT ANDPRELIMINARIES", "In single-agentdomains,theproblemofUnsupervisedEnvironmentDesign(UED)iscastinthe", "framework of an underspecified POMDP (Dennis et al., 2020), which explicitly augments a standard", "POMDP with a set of free parameters controlling aspects of the environment subject to the design", "process. We extend this formalism to the multi-agent setting using stochastic games (Shapley, 1953).", "2" ]

{ "arxiv_id": "2303.03376", "language": "en", "page_idx": 1 }

[ [ "0.1748", "0.03284", "0.4805", "0.03284", "0.4805", "0.04736", "0.1748", "0.04736" ], [ "0.1896", "0.108", "0.8164", "0.10986", "0.8164", "0.154", "0.1896", "0.1515" ], [ "0.353", "0.2203", "0.4844", "0.2203", "0.4844", "0.2343", "0.353", "0.2343" ], [ "0.527", "0.2203", "0.6406", "0.2203", "0.6406", "0.2343", "0.527", "0.2343" ], [ "0.6733", "0.2203", "0.818", "0.2203", "0.818", "0.2343", "0.6733", "0.2343" ], [ "0.1748", "0.2437", "0.824", "0.2437", "0.824", "0.2568", "0.1748", "0.2568" ], [ "0.2009", "0.2229", "0.3145", "0.2229", "0.3145", "0.2329", "0.2009", "0.2329" ], [ "0.1765", "0.2583", "0.824", "0.2583", "0.824", "0.2683", "0.1765", "0.2683" ], [ "0.1765", "0.2708", "0.8228", "0.2708", "0.8228", "0.281", "0.1765", "0.281" ], [ "0.1732", "0.2822", "0.824", "0.2822", "0.824", "0.2954", "0.1732", "0.2954" ], [ "0.1765", "0.2954", "0.824", "0.2954", "0.824", "0.3062", "0.1765", "0.3062" ], [ "0.1724", "0.305", "0.8013", "0.3062", "0.8013", "0.3208", "0.1724", "0.3193" ], [ "0.1732", "0.3257", "0.8105", "0.3257", "0.8105", "0.3403", "0.1732", "0.3403" ], [ "0.808", "0.332", "0.8164", "0.332", "0.8164", "0.336", "0.808", "0.336" ], [ "0.1765", "0.3416", "0.824", "0.3416", "0.824", "0.3547", "0.1765", "0.3547" ], [ "0.1765", "0.3547", "0.824", "0.3547", "0.824", "0.3687", "0.1765", "0.3687" ], [ "0.1765", "0.3706", "0.822", "0.3706", "0.822", "0.3806", "0.1765", "0.3806" ], [ "0.1732", "0.3818", "0.8228", "0.3833", "0.8228", "0.3965", "0.1732", "0.3958" ], [ "0.1732", "0.3965", "0.8228", "0.3958", "0.8228", "0.4104", "0.1732", "0.411" ], [ "0.1748", "0.411", "0.8228", "0.411", "0.8228", "0.4248", "0.1748", "0.4248" ], [ "0.1732", "0.4236", "0.8228", "0.4248", "0.8228", "0.4382", "0.1732", "0.4375" ], [ "0.1748", "0.4382", "0.824", "0.4382", "0.824", "0.452", "0.1748", "0.452" ], [ "0.1765", "0.4539", "0.822", "0.4539", "0.822", "0.464", "0.1765", "0.464" ], [ "0.1748", "0.4666", "0.581", "0.4666", "0.581", "0.4797", "0.1748", "0.4797" ], [ "0.6045", "0.4673", "0.8228", "0.4673", "0.8228", "0.481", "0.6045", "0.481" ], [ "0.1748", "0.4797", "0.824", "0.4797", "0.824", "0.4944", "0.1748", "0.4944" ], [ "0.1765", "0.4968", "0.3652", "0.4968", "0.3652", "0.5103", "0.1765", "0.5103" ], [ "0.357", "0.4968", "0.8228", "0.4956", "0.8228", "0.5103", "0.357", "0.511" ], [ "0.1765", "0.512", "0.8228", "0.512", "0.8228", "0.523", "0.1765", "0.523" ], [ "0.1748", "0.5244", "0.826", "0.5244", "0.826", "0.5386", "0.1748", "0.5386" ], [ "0.1748", "0.5386", "0.826", "0.5386", "0.826", "0.553", "0.1748", "0.553" ], [ "0.1748", "0.553", "0.824", "0.553", "0.824", "0.5664", "0.1748", "0.5664" ], [ "0.1732", "0.565", "0.8228", "0.565", "0.8228", "0.579", "0.1732", "0.579" ], [ "0.1732", "0.579", "0.824", "0.579", "0.824", "0.5923", "0.1732", "0.5923" ], [ "0.1748", "0.5947", "0.738", "0.5947", "0.738", "0.608", "0.1748", "0.608" ], [ "0.1732", "0.6187", "0.5073", "0.617", "0.5073", "0.632", "0.1732", "0.6333" ], [ "0.1765", "0.6426", "0.8228", "0.6426", "0.8228", "0.656", "0.1765", "0.656" ], [ "0.1748", "0.657", "0.826", "0.657", "0.826", "0.6704", "0.1748", "0.6704" ], [ "0.1748", "0.6704", "0.826", "0.6704", "0.826", "0.6846", "0.1748", "0.6846" ], [ "0.1748", "0.6846", "0.824", "0.6846", "0.824", "0.6978", "0.1748", "0.6978" ], [ "0.1748", "0.6997", "0.822", "0.6997", "0.822", "0.714", "0.1748", "0.714" ], [ "0.1748", "0.714", "0.607", "0.714", "0.607", "0.728", "0.1748", "0.728" ], [ "0.1732", "0.7344", "0.8228", "0.7344", "0.8228", "0.7476", "0.1732", "0.7476" ], [ "0.1748", "0.75", "0.3857", "0.75", "0.3857", "0.763", "0.1748", "0.763" ], [ "0.3938", "0.75", "0.8228", "0.75", "0.8228", "0.763", "0.3938", "0.763" ], [ "0.1732", "0.763", "0.826", "0.763", "0.826", "0.7764", "0.1732", "0.7764" ], [ "0.1765", "0.781", "0.2507", "0.781", "0.2507", "0.7915", "0.1765", "0.7915" ], [ "0.2418", "0.7764", "0.824", "0.7744", "0.824", "0.7925", "0.2418", "0.7944" ], [ "0.1748", "0.7925", "0.826", "0.7925", "0.826", "0.806", "0.1748", "0.806" ], [ "0.1748", "0.806", "0.54", "0.806", "0.54", "0.8193", "0.1748", "0.8193" ], [ "0.1748", "0.827", "0.826", "0.827", "0.826", "0.8403", "0.1748", "0.8403" ], [ "0.1732", "0.8413", "0.824", "0.8413", "0.824", "0.855", "0.1732", "0.855" ], [ "0.1748", "0.856", "0.826", "0.856", "0.826", "0.869", "0.1748", "0.869" ], [ "0.1748", "0.869", "0.824", "0.869", "0.824", "0.8833", "0.1748", "0.8833" ], [ "0.1748", "0.8833", "0.826", "0.8833", "0.826", "0.8965", "0.1748", "0.8965" ], [ "0.1732", "0.896", "0.826", "0.8965", "0.826", "0.911", "0.1732", "0.91" ], [ "0.1732", "0.909", "0.8203", "0.91", "0.8203", "0.925", "0.1732", "0.924" ], [ "0.4927", "0.9487", "0.5063", "0.9487", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2023", "鞋", "(b) Middle of training", "(c) End of training", "(d) Zero-shot evaluation", "Figure 2: Emergent complexity of autocurricula induced by MAEsTRo.Examples of partially observable", "(a)Startoftraining", "environmentsprovidedtotheMAEsTRo studentagentatthe(a)start,(b)middle,and (c)end of training.Levels", "becomemorecomplexovertime.LaserTaglevels(toprow)increaseinwall densityandactiveengagement", "between the student and opponent. MultiCarRacing tracks (bottom row) become increasingly more challenging", "with many sharp turns.(d)Exampleheld-out human-designed LaserTag levels andFormula1benchmark tracks", "(Jiang et al., 202la) used for OOD evaluation. For the full list of evaluation environments see Appendix B.", "An Underspecified Partially Observable Stochastic Game (UPOSG) is given by the tuple M", "-", "(n, A, O, , S, T,I, R, ). A, O, and S denote the action, observation, and state spaces, respectively.", " is the set of the environment's free parameters, such as possible positions of walls in a maze or", "levelsofagame.Theseparameterscanbedistinctateverytimestepandarealsoincorporatedintothe", "transition function T : S × A × -→ (S), where A = An is the joint action of all agents. Each", "agent draws individual observations according to the observation function I : S × N → O and obtains", "reward according to the reward function R : S × A × N → R, where N = {1, ..., n]. The discount", "factor is denoted by .Each configuration of theparameter e defines a specific instantiation", "of the environment Me, which is often referred to as a level (Jiang et al., 202la; Parker-Holder", "et al..2022).In this work.we will simplyreferto a specificinstance as an environment when cleal", "from the context. We use -i to notate all agents except i: π.", "(i,..,,1,元i+1,..,TN).For", "an agent i with a stochastic policy i : O × A → [0, 1], we define the value in Me with co-players", "元-ias V(元i, 元_) =E[", "t=otril where rt are the rewards achieved by agent i when following", "policy;inMg.Thegoalistosequentiallyprovidevaluesforandco-players-itoagent2", "during training so that the policy ; is robust to any possible environment and co-player policies,", "i.e., i = argmin, maxo,元-,(V(π*, _i) - V(πi, i)), where π* is the optimal policy on ", "with co-players π_z. UPOSGs are general by nature and allow for cooperative, competitive, and", "mixed scenarios. Furthermore, can represent different game layouts, changes in observations, and", "environment dynamics. When n = 1, UPOSGs are identical to UPOMDPs for single-agent tasks. In", "the remainder of this work, we concentrate on competitive settings with n = 2 agents.", "2.1 UED APPROACHES IN SINGLE-AGENTRL", "A curriculum over the environment parameters can arise from a teacher maximising a utility function", "Ut(π, O) based on the student's policy π. The most naive form of UED is domain randomisation (DR,", "Jakobi, 1997; Sadeghi & Levine, 2017), whereby environments are sampled uniformly at random,", "corresponding to a constant utility U (π, 0) = C. Recent UED approaches use regret as the objective", "for maximisation (Dennis et al., 2020; Gur et al., 2021): UR(π, 0) = max*eI[REGRET° (π, *)) =", "max*e(Ve(*) - Ve()), where * is the optimal policy on 0.", "Empirically,regret-based objectives produce curricula of increasing complexity that re", "sult in more robust policies.", "Moreover, if the learning process reaches a Nash equilib", "rium,the student provably follows a minimax-regret policy (Dennis et al., 2020):E", "argminse", "en(maxe,BEe,I[REGRET'(πA, B)}), where II and @ are the strategy sets of the", "student and the teacher, respectively. While this is an appealing property, the optimal policy * is", "unknown in practice and must be approximated instead.", "Prioritized Level Replay (PLR, Jiang et al., 202la;b) continually curates an environment buffer con-", "taining the environments generated under domain randomisation with the highest learning potential,", "e.g., as measured by estimated regret. PLR alternates between evaluating new environments for learn-", "ing potential and performing prioritised training of the agent on the environments with the highest", "learning potential so far found. By training the student only on curated high-regret environments,", "PLR provably results in a minimax-regret student policy at Nash equilibrium. MAEsTRO makes use", "of PLR-style curation of environment buffers to discover high-regret environments during training.", "3" ]

{ "arxiv_id": "2303.03376", "language": "en", "page_idx": 2 }

[ [ "0.1732", "0.03284", "0.4805", "0.0322", "0.4805", "0.04672", "0.1732", "0.04736" ], [ "0.1748", "0.1048", "0.6807", "0.1048", "0.6807", "0.1187", "0.1748", "0.1187" ], [ "0.1748", "0.132", "0.8267", "0.132", "0.8267", "0.1459", "0.1748", "0.1459" ], [ "0.1732", "0.1433", "0.826", "0.1445", "0.826", "0.159", "0.1732", "0.1578" ], [ "0.1748", "0.159", "0.824", "0.159", "0.824", "0.1724", "0.1748", "0.1724" ], [ "0.1732", "0.1718", "0.8228", "0.1718", "0.8228", "0.1863", "0.1732", "0.1863" ], [ "0.1724", "0.1875", "0.7534", "0.1849", "0.7534", "0.1995", "0.1724", "0.202" ], [ "0.1708", "0.2052", "0.826", "0.2058", "0.826", "0.2235", "0.1708", "0.2229" ], [ "0.1732", "0.2229", "0.824", "0.2229", "0.824", "0.2361", "0.1732", "0.2361" ], [ "0.1732", "0.2361", "0.6235", "0.2361", "0.6235", "0.25", "0.1732", "0.25" ], [ "0.1732", "0.2556", "0.824", "0.2568", "0.824", "0.2715", "0.1732", "0.2708" ], [ "0.1732", "0.2715", "0.8228", "0.2715", "0.8228", "0.286", "0.1732", "0.286" ], [ "0.1691", "0.2854", "0.8286", "0.2778", "0.8286", "0.3062", "0.1691", "0.3132" ], [ "0.1978", "0.2842", "0.2793", "0.2822", "0.2793", "0.2935", "0.1978", "0.2954" ], [ "0.1732", "0.3062", "0.648", "0.3062", "0.648", "0.3193", "0.1732", "0.3193" ], [ "0.1748", "0.339", "0.1929", "0.339", "0.1929", "0.3528", "0.1748", "0.3528" ], [ "0.1978", "0.339", "0.2834", "0.339", "0.2834", "0.3542", "0.1978", "0.3542" ], [ "0.1748", "0.363", "0.2051", "0.3655", "0.2035", "0.3774", "0.1732", "0.3757" ], [ "0.2091", "0.3643", "0.4192", "0.3643", "0.4192", "0.3774", "0.2091", "0.3774" ], [ "0.1732", "0.3865", "0.8267", "0.385", "0.8267", "0.3997", "0.1732", "0.401" ], [ "0.1748", "0.401", "0.5557", "0.401", "0.5557", "0.414", "0.1748", "0.414" ], [ "0.5537", "0.3997", "0.8267", "0.3997", "0.8267", "0.4136", "0.5537", "0.4136" ], [ "0.619", "0.414", "0.714", "0.414", "0.714", "0.4248", "0.619", "0.4248" ], [ "0.728", "0.4136", "0.826", "0.4136", "0.826", "0.4248", "0.728", "0.4248" ], [ "0.556", "0.4167", "0.574", "0.4124", "0.5786", "0.4192", "0.5605", "0.4248" ], [ "0.1748", "0.4153", "0.556", "0.4153", "0.556", "0.4287", "0.1748", "0.4287" ], [ "0.1748", "0.4287", "0.5522", "0.4287", "0.5522", "0.4426", "0.1748", "0.4426" ], [ "0.5605", "0.4236", "0.7974", "0.4236", "0.7974", "0.4375", "0.5605", "0.4375" ], [ "0.5645", "0.4382", "0.661", "0.4382", "0.661", "0.4495", "0.5645", "0.4495" ], [ "0.6587", "0.4382", "0.824", "0.4382", "0.824", "0.449", "0.6587", "0.449" ], [ "0.562", "0.4495", "0.8267", "0.4495", "0.8267", "0.4634", "0.562", "0.4634" ], [ "0.1732", "0.4426", "0.5522", "0.4426", "0.5522", "0.4558", "0.1732", "0.4558" ], [ "0.1765", "0.4583", "0.5483", "0.4583", "0.5483", "0.4685", "0.1765", "0.4685" ], [ "0.626", "0.4697", "0.6445", "0.4697", "0.6445", "0.4817", "0.626", "0.4817" ], [ "0.66", "0.4697", "0.6846", "0.4697", "0.6846", "0.481", "0.66", "0.481" ], [ "0.7295", "0.4697", "0.7563", "0.4697", "0.7563", "0.481", "0.7295", "0.481" ], [ "0.6978", "0.4717", "0.7183", "0.4717", "0.7183", "0.4797", "0.6978", "0.4797" ], [ "0.1732", "0.4702", "0.55", "0.4702", "0.55", "0.4841", "0.1732", "0.4841" ], [ "0.7754", "0.473", "0.796", "0.473", "0.796", "0.481", "0.7754", "0.481" ], [ "0.7935", "0.4734", "0.8105", "0.4734", "0.8105", "0.481", "0.7935", "0.481" ], [ "0.626", "0.4817", "0.6543", "0.4817", "0.6543", "0.4944", "0.626", "0.4944" ], [ "0.66", "0.4817", "0.693", "0.4817", "0.693", "0.4944", "0.66", "0.4944" ], [ "0.697", "0.4817", "0.726", "0.4817", "0.726", "0.4944", "0.697", "0.4944" ], [ "0.7344", "0.4817", "0.763", "0.4817", "0.763", "0.4944", "0.7344", "0.4944" ], [ "0.7715", "0.481", "0.8164", "0.481", "0.8164", "0.4956", "0.7715", "0.4956" ], [ "0.581", "0.4841", "0.6045", "0.4841", "0.6045", "0.4968", "0.581", "0.4968" ], [ "0.1748", "0.4841", "0.5513", "0.4841", "0.5513", "0.4976", "0.1748", "0.4976" ], [ "0.3096", "0.4956", "0.5513", "0.4988", "0.5513", "0.513", "0.3096", "0.5103" ], [ "0.626", "0.4968", "0.652", "0.4968", "0.652", "0.5083", "0.626", "0.5083" ], [ "0.6626", "0.4968", "0.693", "0.4968", "0.693", "0.5083", "0.6626", "0.5083" ], [ "0.699", "0.4968", "0.726", "0.4968", "0.726", "0.5083", "0.699", "0.5083" ], [ "0.7344", "0.4968", "0.761", "0.4968", "0.761", "0.5083", "0.7344", "0.5083" ], [ "0.774", "0.4968", "0.814", "0.4968", "0.814", "0.5083", "0.774", "0.5083" ], [ "0.58", "0.4976", "0.6045", "0.4976", "0.6045", "0.509", "0.58", "0.509" ], [ "0.1748", "0.4988", "0.3186", "0.4988", "0.3186", "0.512", "0.1748", "0.512" ], [ "0.626", "0.509", "0.653", "0.509", "0.653", "0.521", "0.626", "0.521" ], [ "0.6626", "0.509", "0.6978", "0.509", "0.6978", "0.521", "0.6626", "0.521" ], [ "0.694", "0.509", "0.726", "0.509", "0.726", "0.521", "0.694", "0.521" ], [ "0.732", "0.509", "0.761", "0.509", "0.761", "0.521", "0.732", "0.521" ], [ "0.7817", "0.509", "0.812", "0.509", "0.812", "0.521", "0.7817", "0.521" ], [ "0.584", "0.511", "0.601", "0.511", "0.601", "0.521", "0.584", "0.521" ], [ "0.1748", "0.512", "0.55", "0.512", "0.55", "0.526", "0.1748", "0.526" ], [ "0.572", "0.5234", "0.615", "0.5234", "0.615", "0.534", "0.572", "0.534" ], [ "0.6274", "0.5234", "0.653", "0.5234", "0.653", "0.532", "0.6274", "0.532" ], [ "0.66", "0.523", "0.726", "0.523", "0.726", "0.5327", "0.66", "0.5327" ], [ "0.7344", "0.5234", "0.7607", "0.5234", "0.7607", "0.532", "0.7344", "0.532" ], [ "0.1748", "0.539", "0.824", "0.539", "0.824", "0.5537", "0.1748", "0.5537" ], [ "0.1748", "0.526", "0.5513", "0.526", "0.5513", "0.5405", "0.1748", "0.5405" ], [ "0.1732", "0.553", "0.824", "0.553", "0.824", "0.5664", "0.1732", "0.5664" ], [ "0.1724", "0.5674", "0.826", "0.5674", "0.826", "0.5806", "0.1724", "0.5806" ], [ "0.1765", "0.5996", "0.2141", "0.5996", "0.2141", "0.6113", "0.1765", "0.6113" ], [ "0.2075", "0.5996", "0.79", "0.5996", "0.79", "0.613", "0.2075", "0.613" ], [ "0.1748", "0.6255", "0.4478", "0.6255", "0.4478", "0.64", "0.1748", "0.64" ], [ "0.4558", "0.631", "0.625", "0.631", "0.625", "0.6445", "0.4558", "0.6445" ], [ "0.1748", "0.64", "0.4468", "0.64", "0.4468", "0.654", "0.1748", "0.654" ], [ "0.1748", "0.654", "0.4478", "0.654", "0.4478", "0.6675", "0.1748", "0.6675" ], [ "0.4575", "0.6484", "0.675", "0.648", "0.675", "0.662", "0.4575", "0.6636" ], [ "0.464", "0.664", "0.4795", "0.664", "0.4795", "0.6724", "0.464", "0.6724" ], [ "0.4722", "0.661", "0.785", "0.66", "0.785", "0.675", "0.4722", "0.6753" ], [ "0.1748", "0.6685", "0.4468", "0.6685", "0.4468", "0.6816", "0.1748", "0.6816" ], [ "0.4617", "0.6753", "0.7983", "0.6753", "0.7983", "0.6895", "0.4617", "0.6895" ], [ "0.1748", "0.6816", "0.4453", "0.6816", "0.4453", "0.696", "0.1748", "0.696" ], [ "0.4617", "0.688", "0.6094", "0.688", "0.6094", "0.702", "0.4617", "0.702" ], [ "0.1732", "0.6963", "0.4028", "0.6963", "0.4028", "0.7104", "0.1732", "0.7104" ], [ "0.4983", "0.7007", "0.6313", "0.7007", "0.6313", "0.7153", "0.4983", "0.7153" ], [ "0.5254", "0.714", "0.5728", "0.714", "0.5728", "0.726", "0.5254", "0.726" ], [ "0.653", "0.714", "0.8228", "0.714", "0.8228", "0.728", "0.653", "0.728" ], [ "0.524", "0.726", "0.665", "0.726", "0.665", "0.7407", "0.524", "0.7407" ], [ "0.1748", "0.7173", "0.4453", "0.7173", "0.4453", "0.731", "0.1748", "0.731" ], [ "0.1748", "0.731", "0.4478", "0.731", "0.4478", "0.744", "0.1748", "0.744" ], [ "0.7075", "0.728", "0.8228", "0.7266", "0.8228", "0.738", "0.7075", "0.7393" ], [ "0.523", "0.7373", "0.6284", "0.738", "0.6284", "0.7524", "0.523", "0.752" ], [ "0.1765", "0.747", "0.4453", "0.747", "0.4453", "0.757", "0.1765", "0.757" ], [ "0.5483", "0.7524", "0.615", "0.7524", "0.615", "0.7646", "0.5483", "0.7646" ], [ "0.6357", "0.754", "0.8228", "0.754", "0.8228", "0.7646", "0.6357", "0.7646" ], [ "0.5483", "0.7666", "0.773", "0.7666", "0.773", "0.78", "0.5483", "0.78" ], [ "0.1724", "0.7583", "0.4453", "0.7583", "0.4453", "0.7715", "0.1724", "0.7715" ], [ "0.1732", "0.773", "0.4468", "0.773", "0.4468", "0.786", "0.1732", "0.786" ], [ "0.46", "0.771", "0.468", "0.771", "0.468", "0.776", "0.46", "0.776" ], [ "0.5483", "0.779", "0.7256", "0.779", "0.7256", "0.7925", "0.5483", "0.7925" ], [ "0.1724", "0.785", "0.4453", "0.785", "0.4453", "0.799", "0.1724", "0.799" ], [ "0.4592", "0.799", "0.4656", "0.7944", "0.4697", "0.8", "0.4631", "0.8037" ], [ "0.524", "0.7935", "0.5513", "0.7935", "0.5513", "0.8047", "0.524", "0.8047" ], [ "0.1732", "0.8", "0.4468", "0.799", "0.4468", "0.813", "0.1732", "0.8145" ], [ "0.5483", "0.806", "0.591", "0.806", "0.591", "0.8164", "0.5483", "0.8164" ], [ "0.6274", "0.8066", "0.822", "0.8066", "0.822", "0.8174", "0.6274", "0.8174" ], [ "0.1748", "0.8145", "0.4468", "0.8145", "0.4468", "0.8276", "0.1748", "0.8276" ], [ "0.46", "0.8096", "0.4697", "0.8096", "0.4697", "0.8164", "0.46", "0.8164" ], [ "0.5474", "0.8174", "0.773", "0.8184", "0.773", "0.8335", "0.5474", "0.832" ], [ "0.1748", "0.8276", "0.4453", "0.8276", "0.4453", "0.8413", "0.1748", "0.8413" ], [ "0.1732", "0.8413", "0.4468", "0.8403", "0.4468", "0.855", "0.1732", "0.856" ], [ "0.524", "0.836", "0.7715", "0.836", "0.7715", "0.8496", "0.524", "0.8496" ], [ "0.4631", "0.854", "0.4714", "0.86", "0.464", "0.8643", "0.4575", "0.8594" ], [ "0.524", "0.851", "0.7344", "0.851", "0.7344", "0.8643", "0.524", "0.8643" ], [ "0.5", "0.8657", "0.687", "0.8657", "0.687", "0.88", "0.5", "0.88" ], [ "0.1732", "0.856", "0.4468", "0.856", "0.4468", "0.869", "0.1732", "0.869" ], [ "0.1748", "0.869", "0.4468", "0.869", "0.4468", "0.8833", "0.1748", "0.8833" ], [ "0.7236", "0.867", "0.8228", "0.867", "0.8228", "0.8774", "0.7236", "0.8774" ], [ "0.1732", "0.8853", "0.4429", "0.8853", "0.4429", "0.896", "0.1732", "0.896" ], [ "0.1732", "0.898", "0.826", "0.898", "0.826", "0.911", "0.1732", "0.911" ], [ "0.1748", "0.911", "0.361", "0.911", "0.361", "0.925", "0.1748", "0.925" ], [ "0.4927", "0.9487", "0.505", "0.9487", "0.505", "0.961", "0.4927", "0.961" ] ]

[ "Published as a conference paper at ICLR 2023", "2.2 CO-PLAYER AUTOCURRICULA IN COMPETITIVE MULTI-AGENT RL", "The choice of co-players plays a crucial role for training agents in multi-agent domains (Leibo et al.,", "2019). A simple but effective approach is self-play (SP, Silver et al., 2018), whereby the agent always", "plays with copies of itself. In competitive two-player games, SP produces a curriculum in which", "opponents match each other in skill level. However, sP may cause cycles in the strategy space when", "agents forget how to play against previous versions of their policy (Garnelo et al., 2021).", "Fictitious Self-Play (FSP, Heinrich et al., 2015) overcomes the emergence of such cycles by training", "an agent against a uniform mixture of all previous policies. However, FsP can result in wasting a", "large number of interactions against significantly weaker opponents.", "Prioritized Fictitious Self-Play (PFSP, Vinyals et al., 2019) mitigates this potential inefficiency", "by matching agent A with a frozen opponent B from the set of candidates C with probability", "e whedins theet rulum oroppons.orle rd()", "f(P[A beats B)", "(1 - )P, where p E R+, forces PFSP to focus on the hardest opponents.", "3", "METHOD", "3.1", "AN ILLUSTRATIVE EXAMPLE", "To highlight the importance of curricula over the joint Table l: An illustrative two-player game.", "space of environments and co-players, we provide an il-", "Rows correspond to co-player policies and", "A,TB,TC).", "Columns indi-", "1", "lustrative example of a simple two-player game in Ta-", "ble 1. Here, the goal of the regret-maximising teacher", "cate different environments and e", "[01, 02, 03, 04].", "Thepayoffmatrix repre-", "sents the regret of the student on pair (o, ).", "is to select an environment/co-player pair for the stu-", "dent.Ignoring the co-player and selecting the highest", "01", "02", "04", "03", "regret environment leads to choosing 3. Similarly, ig", "E", "ve", "0.6", "0.1", "0.4", "0.2", "0.325", "TA", "noring environments and selecting the highest regret co-", "This yields a suboptimal pair (93,", "0.1", "0.5", "0.4", "0.3", "0.325", "TB", "player leads to πc.", "0.2", "0.4", "0.4", "0.4", "0.35", "TC", "πc) with REGRET(03,πc) =0.4, whereas a teacher", "E元~11", "0.3", "0.330.4", "0.3", "REGRET(01, ^A) = 0.6. Thus, naively treating the environment and co-player as independent can", "over the joint space yields the optimal pair (01, A) with", "yield a sub-optimal curriculum. Such overspecialisation toward a subset of environmental challenges", "at the expense of overall robustness commonly emerges in multi-agent settings (Garnelo et al., 2021)", "3.2", "MULTI-AGENT ENVIRONMENT DESIGN STRATEGIST FOR OPEN-ENDED LEARNING", "In this section, we describe a new multi-", "Algorithm 1: MAESTRO", "agent UED approach called Multi-Agent", "Environment Design Strategist for Open-", "Input:Environment generator", "21", "Initialise: Student policy , co-player population ", "Ended Learning (MAESTRO) which in-", "3 Initialise: Environment buffers V’ E β, A(') := 0.", "duces a regret-based autocurricula jointly", "4 fori ={1,2,...} do", "over environments and co-players.", "for many episodes do", "T~B", " Sample co-player via Eq 1", "Sample replay decision", "MAESTRo is a replay-guided approach", "(Jiang et al., 202la) that relies on an en-", "> see Section 3.2.2", "if replaying then", "vironmentgeneratorto continuouslycre-", "0~A(元)", "Sampleareplayenvironment", "Collect trajectory of π using (0, ')", "ate new environment instances..Rather", "than storing a single environment buffer,", "10", "Update with rewards R(↑)", "as done by PLR, MAESTRO maintains a", "", "else", "population of policies with each pol-", "0~0", " Samplearandomenvironment", "icy assigned its own environment buffer.", "13", "Collect trajectory T of using (,)", "In each environment 0 produced by the", "generator, MAESTRO evaluates the stu-", "Compute regret score S = Regret(0, )", "10", "Update A(π') with using score S", "BBU(),A():= 0", "dent's performance against a non-uniform", "mixture of co-player policies in B. High-", "frozen weights", "regret environment/co-player pairs are", "then stored, along with the student agent's regret estimate for that pair, in the corresponding co-", "player's environment buffer.", "4" ]

{ "arxiv_id": "2303.03376", "language": "en", "page_idx": 3 }

[ [ "0.1732", "0.03284", "0.4805", "0.0322", "0.4805", "0.04672", "0.1732", "0.04736" ], [ "0.1765", "0.1048", "0.8228", "0.1048", "0.8228", "0.1193", "0.1765", "0.1193" ], [ "0.1765", "0.1193", "0.8228", "0.1193", "0.8228", "0.1332", "0.1765", "0.1332" ], [ "0.1748", "0.1332", "0.8228", "0.1332", "0.8228", "0.1465", "0.1748", "0.1465" ], [ "0.1748", "0.1465", "0.824", "0.1465", "0.824", "0.161", "0.1748", "0.161" ], [ "0.1732", "0.161", "0.7485", "0.161", "0.7485", "0.1749", "0.1732", "0.1749" ], [ "0.1765", "0.1901", "0.2214", "0.1901", "0.2214", "0.202", "0.1765", "0.202" ], [ "0.2173", "0.1895", "0.556", "0.1895", "0.556", "0.2026", "0.2173", "0.2026" ], [ "0.1748", "0.2134", "0.826", "0.2134", "0.826", "0.2279", "0.1748", "0.2279" ], [ "0.1732", "0.2279", "0.826", "0.2279", "0.826", "0.2412", "0.1732", "0.2412" ], [ "0.1732", "0.2424", "0.824", "0.2424", "0.824", "0.2556", "0.1732", "0.2556" ], [ "0.1748", "0.2556", "0.826", "0.2556", "0.826", "0.2695", "0.1748", "0.2695" ], [ "0.1748", "0.2695", "0.824", "0.2695", "0.824", "0.2827", "0.1748", "0.2827" ], [ "0.1748", "0.2827", "0.8228", "0.2827", "0.8228", "0.2966", "0.1748", "0.2966" ], [ "0.1748", "0.2966", "0.824", "0.2966", "0.824", "0.3113", "0.1748", "0.3113" ], [ "0.1748", "0.3113", "0.824", "0.3113", "0.824", "0.3245", "0.1748", "0.3245" ], [ "0.1748", "0.3257", "0.3718", "0.3257", "0.3718", "0.336", "0.1748", "0.336" ], [ "0.1748", "0.3528", "0.5786", "0.3528", "0.5786", "0.3662", "0.1748", "0.3662" ], [ "0.1748", "0.3774", "0.824", "0.3774", "0.824", "0.3914", "0.1748", "0.3914" ], [ "0.1732", "0.3926", "0.824", "0.3926", "0.824", "0.406", "0.1732", "0.406" ], [ "0.1732", "0.406", "0.601", "0.4048", "0.601", "0.4192", "0.1732", "0.4204" ], [ "0.3228", "0.4224", "0.6724", "0.4248", "0.6724", "0.4458", "0.3228", "0.443" ], [ "0.8013", "0.4268", "0.8267", "0.4268", "0.8267", "0.4443", "0.8013", "0.4443" ], [ "0.464", "0.4426", "0.4958", "0.4426", "0.4958", "0.4507", "0.464", "0.4507" ], [ "0.5063", "0.44", "0.562", "0.44", "0.562", "0.4507", "0.5063", "0.4507" ], [ "0.1748", "0.4539", "0.8228", "0.4539", "0.8228", "0.4685", "0.1748", "0.4685" ], [ "0.1732", "0.4685", "0.8228", "0.4685", "0.8228", "0.4817", "0.1732", "0.4817" ], [ "0.1732", "0.4817", "0.824", "0.481", "0.824", "0.4956", "0.1732", "0.4968" ], [ "0.1732", "0.4968", "0.826", "0.4968", "0.826", "0.511", "0.1732", "0.511" ], [ "0.1708", "0.512", "0.6772", "0.507", "0.6772", "0.529", "0.1708", "0.534" ], [ "0.6665", "0.512", "0.824", "0.5146", "0.824", "0.529", "0.6665", "0.5264" ], [ "0.1748", "0.531", "0.8228", "0.531", "0.8228", "0.5444", "0.1748", "0.5444" ], [ "0.1748", "0.5444", "0.8228", "0.5444", "0.8228", "0.558", "0.1748", "0.558" ], [ "0.1732", "0.5596", "0.8228", "0.5596", "0.8228", "0.5728", "0.1732", "0.5728" ], [ "0.1748", "0.5728", "0.8228", "0.5728", "0.8228", "0.586", "0.1748", "0.586" ], [ "0.1765", "0.5884", "0.8228", "0.5884", "0.8228", "0.5986", "0.1765", "0.5986" ], [ "0.1724", "0.5996", "0.8228", "0.5996", "0.8228", "0.613", "0.1724", "0.613" ], [ "0.1732", "0.6143", "0.584", "0.6143", "0.584", "0.6274", "0.1732", "0.6274" ], [ "0.1732", "0.643", "0.505", "0.643", "0.505", "0.657", "0.1732", "0.657" ], [ "0.1765", "0.672", "0.822", "0.672", "0.822", "0.6816", "0.1765", "0.6816" ], [ "0.1765", "0.686", "0.822", "0.686", "0.822", "0.6963", "0.1765", "0.6963" ], [ "0.1748", "0.6987", "0.8228", "0.6987", "0.8228", "0.712", "0.1748", "0.712" ], [ "0.1765", "0.712", "0.8228", "0.712", "0.8228", "0.726", "0.1765", "0.726" ], [ "0.1748", "0.726", "0.824", "0.726", "0.824", "0.7407", "0.1748", "0.7407" ], [ "0.1748", "0.741", "0.8228", "0.741", "0.8228", "0.756", "0.1748", "0.756" ], [ "0.1732", "0.755", "0.826", "0.756", "0.826", "0.771", "0.1732", "0.7695" ], [ "0.1732", "0.768", "0.822", "0.771", "0.822", "0.785", "0.1732", "0.7827" ], [ "0.3423", "0.7886", "0.6504", "0.7886", "0.6504", "0.803", "0.3423", "0.803" ], [ "0.8022", "0.789", "0.824", "0.789", "0.824", "0.8037", "0.8022", "0.8037" ], [ "0.3865", "0.8037", "0.4885", "0.8037", "0.4885", "0.8145", "0.3865", "0.8145" ], [ "0.1732", "0.8193", "0.824", "0.8193", "0.824", "0.8335", "0.1732", "0.8335" ], [ "0.1748", "0.834", "0.8228", "0.834", "0.8228", "0.8477", "0.1748", "0.8477" ], [ "0.1748", "0.8467", "0.824", "0.8467", "0.824", "0.86", "0.1748", "0.86" ], [ "0.1748", "0.8613", "0.826", "0.8613", "0.826", "0.875", "0.1748", "0.875" ], [ "0.1748", "0.875", "0.8267", "0.875", "0.8267", "0.888", "0.1748", "0.888" ], [ "0.1732", "0.888", "0.818", "0.888", "0.818", "0.903", "0.1732", "0.903" ], [ "0.1936", "0.909", "0.8228", "0.91", "0.8228", "0.925", "0.1936", "0.924" ], [ "0.492", "0.9463", "0.5073", "0.9463", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2023", "MAESTRO maintains a dynamic population of co-players of various skill levels throughout training", "(Czarnecki et al., 2020). Figure 1 presents the overall training paradigm of MAEsTRO and Algorithm I", "provides its pseudocode. We note that when applied to a fixed singleton environment, MAESTRO", "becomes a variation of PFSP (Vinyals et al., 2019), where the mixture of policies from the population", "is computed based on the agent's regret estimate, rather than the probability of winning.", "3.2.1", "MAINTAININGAPOPULATIONOFCO-PLAYERS", "A key issue of using replay-guided autocurricula for multi-agent settings is nonstationarity. Specif-", "ically, using PLR with SP results in inaccurate regret estimates over environment/co-player pairs,", "as the co-player policies evolve over training. MAEsTRO overcomes this issue by maintaining a", "population of past policies (Lanctot et al., 2017). This approach confers several key benefits. First,", "re-encountering past agents helps avoid cycles in strategy space (Balduzzi et al., 2019; Garnelo", "et al., 2021). Second, MAEsTRO maintains accurate regret estimates in the face of nonstationarity", "by employing a separate environment buffer for each policy in the population. Third, MAESTRO", "always optimises a single policy throughout training, rather than a set of distinct policies which can", "becomputationallyexpensive", "3.2.2 CURATING THE ENVIRONMENT/CO-PLAYER PAIRS", "To optimise for the global regret over the joint environment/co-player space, a MAEsTRo student is", "trained to best respond to a non-uniform mixture of policies from a population by prioritising training", "against co-players with high-regret environments in their buffers:", "Co-PLAYERHR E arg maxt,max, Regret(0, n'),", "(1)", "T'EB", "0EA)", "where B is the co-player population, A(π') is the environment buffer of agent ', and Regret is", "the estimated regret of student for the pair (0, '). To ensure that the student learns to best respond", "to high-regret co-players as well as to the entire population B, we enforce all members of to be", "assigned a minimum probability . For instance, if Equation (1) returns a single highest-regret co-", "player, then the resulting prioritised distribution assigns a weight of N-AW-", "-1) to the highest-regret", "co-player and weight of to the remaining co-players.For each co-player ' E B, MAEsTRo", "maintains a PLR environment buffer A(') with the top-K high-regret levels. Once the co-player", "is sampled, we make a replay decision: with probability p, we use a bandit to sample a training", "environment from A('),2 and with probability 1 -p, we sample a new environment for evaluation", "fromtheenvironmentgenerator.SimilartoRobustPLR(Jiangetal..2o2la),weonlyupdatethe", "student policy on environments sampled from the environment buffer. This provides MAEsTRo with", "strong robustness guarantees, which we discuss in Section 3.3.", "3.3ROBUSTNESSGUARANTEESOFMAESTRO", "WeanalysetheexpectedbehaviourofMAEsTROifthesystemreachesequilibrium:doesMAESTRO", "producearegret-maximisingdistributionover environmentsand co-players and isthepolicy optimal", "with respect to these distributions? We cast this problem in terms of Bayesian Nash equilibrium", "(BNE) behaviour in individual environments. BNE is an extension of Nash equilibrium (NE) where", "each co-player j has an unknown type parameter oj, which affects the dynamics of the game and", "is only known to that player. The distribution over these parameters N is assumed to be common", "knowledge. Equilibria are then defined as the set of policies, conditioned on their unknown type,", "each being a best response to the policies of the other players. That is, for policy j of any player j,", "; Eargmaxf E [U(;(0),元-,(0-j)))", "(2)", "元jEIljONEON", "In MAEsTRO, we can assume each co-player is effectively omniscient, as each is co-evolved for", "maximal performance against the student in the environments it is paired with in its high-regret", "environment buffer. In contrast, the student has conferred no special advantages and has access to only", "the standard observations. We formalise this setting as a -i-knowing game. This game corresponds", "to the POSG with the same set of players, action space, rewards, and states as the original UPOSG,", "but with sampled at the first time step and provided to co-players -i as part of their observation.", "2Sampling is based on environment's regret score,staleness,and otherfeaturesfollowing (Jiang etal.,202b).", "5" ]

{ "arxiv_id": "2303.03376", "language": "en", "page_idx": 4 }

[ [ "0.1993", "0.125", "0.7974", "0.127", "0.7974", "0.1477", "0.1993", "0.1459" ], [ "0.3203", "0.2008", "0.4143", "0.2026", "0.4143", "0.2172", "0.3203", "0.2153" ], [ "0.4346", "0.2026", "0.5645", "0.2026", "0.5645", "0.2166", "0.4346", "0.2166" ], [ "0.589", "0.2039", "0.6807", "0.2039", "0.6807", "0.2153", "0.589", "0.2153" ], [ "0.2385", "0.2166", "0.761", "0.2166", "0.761", "0.2311", "0.2385", "0.2311" ], [ "0.4192", "0.207", "0.4307", "0.2052", "0.4314", "0.2102", "0.4224", "0.2122" ], [ "0.3506", "0.2299", "0.652", "0.2299", "0.652", "0.2443", "0.3506", "0.2443" ], [ "0.02287", "0.262", "0.05963", "0.262", "0.05963", "0.4072", "0.02287", "0.4072" ], [ "0.4617", "0.279", "0.54", "0.279", "0.54", "0.2966", "0.4617", "0.2966" ], [ "0.2345", "0.3113", "0.7656", "0.3113", "0.7656", "0.3245", "0.2345", "0.3245" ], [ "0.2345", "0.3245", "0.7646", "0.3245", "0.7646", "0.3384", "0.2345", "0.3384" ], [ "0.2345", "0.3384", "0.7656", "0.3384", "0.7656", "0.3528", "0.2345", "0.3528" ], [ "0.232", "0.351", "0.7656", "0.3516", "0.7656", "0.3662", "0.232", "0.3655" ], [ "0.2336", "0.3662", "0.7656", "0.3662", "0.7656", "0.3801", "0.2336", "0.3801" ], [ "0.2336", "0.3801", "0.7656", "0.3801", "0.7656", "0.3933", "0.2336", "0.3933" ], [ "0.02287", "0.3958", "0.0621", "0.3965", "0.05963", "0.4944", "0.0196", "0.4937" ], [ "0.2345", "0.3958", "0.7646", "0.3958", "0.7646", "0.406", "0.2345", "0.406" ], [ "0.2336", "0.4077", "0.7656", "0.4077", "0.7656", "0.4216", "0.2336", "0.4216" ], [ "0.2345", "0.4216", "0.7656", "0.4216", "0.7656", "0.435", "0.2345", "0.435" ], [ "0.2345", "0.435", "0.767", "0.435", "0.767", "0.449", "0.2345", "0.449" ], [ "0.2336", "0.449", "0.7656", "0.449", "0.7656", "0.4622", "0.2336", "0.4622" ], [ "0.2361", "0.464", "0.7646", "0.464", "0.7646", "0.4749", "0.2361", "0.4749" ], [ "0.2345", "0.4766", "0.7656", "0.4766", "0.7656", "0.4905", "0.2345", "0.4905" ], [ "0.02287", "0.4849", "0.05963", "0.4849", "0.05963", "0.712", "0.02287", "0.712" ], [ "0.2345", "0.4924", "0.7646", "0.4924", "0.7646", "0.5024", "0.2345", "0.5024" ], [ "0.2345", "0.504", "0.7656", "0.504", "0.7656", "0.5176", "0.2345", "0.5176" ], [ "0.2336", "0.518", "0.7646", "0.518", "0.7646", "0.532", "0.2336", "0.532" ], [ "0.232", "0.531", "0.6006", "0.531", "0.6006", "0.5444", "0.232", "0.5444" ], [ "0.1748", "0.571", "0.2035", "0.571", "0.2035", "0.582", "0.1748", "0.582" ], [ "0.1969", "0.5693", "0.312", "0.5693", "0.312", "0.584", "0.1969", "0.584" ], [ "0.1748", "0.5996", "0.824", "0.5996", "0.824", "0.6143", "0.1748", "0.6143" ], [ "0.1732", "0.6143", "0.826", "0.613", "0.826", "0.6274", "0.1732", "0.629" ], [ "0.1748", "0.629", "0.824", "0.629", "0.824", "0.6426", "0.1748", "0.6426" ], [ "0.1748", "0.6426", "0.824", "0.6426", "0.824", "0.656", "0.1748", "0.656" ], [ "0.1748", "0.656", "0.824", "0.656", "0.824", "0.6694", "0.1748", "0.6694" ], [ "0.1724", "0.6685", "0.814", "0.6694", "0.814", "0.6846", "0.1724", "0.683" ], [ "0.1748", "0.69", "0.824", "0.69", "0.824", "0.704", "0.1748", "0.704" ], [ "0.1765", "0.7056", "0.8228", "0.7056", "0.8228", "0.7163", "0.1765", "0.7163" ], [ "0.1748", "0.7183", "0.824", "0.7183", "0.824", "0.732", "0.1748", "0.732" ], [ "0.1748", "0.732", "0.826", "0.732", "0.826", "0.7456", "0.1748", "0.7456" ], [ "0.1748", "0.7456", "0.824", "0.7456", "0.824", "0.759", "0.1748", "0.759" ], [ "0.1748", "0.7603", "0.826", "0.7603", "0.826", "0.7734", "0.1748", "0.7734" ], [ "0.1748", "0.7734", "0.824", "0.7734", "0.824", "0.787", "0.1748", "0.787" ], [ "0.1732", "0.786", "0.8228", "0.786", "0.8228", "0.8", "0.1732", "0.8" ], [ "0.1748", "0.8003", "0.824", "0.8003", "0.824", "0.8145", "0.1748", "0.8145" ], [ "0.1724", "0.8145", "0.824", "0.813", "0.824", "0.8276", "0.1724", "0.829" ], [ "0.1765", "0.83", "0.7983", "0.83", "0.7983", "0.8403", "0.1765", "0.8403" ], [ "0.1732", "0.8486", "0.826", "0.8486", "0.826", "0.8623", "0.1732", "0.8623" ], [ "0.1748", "0.863", "0.824", "0.863", "0.824", "0.877", "0.1748", "0.877" ], [ "0.1732", "0.877", "0.826", "0.877", "0.826", "0.89", "0.1732", "0.89" ], [ "0.1748", "0.89", "0.824", "0.89", "0.824", "0.904", "0.1748", "0.904" ], [ "0.1732", "0.9224", "0.7183", "0.924", "0.7183", "0.938", "0.1732", "0.9375" ] ]

[ "Local Similarity-Aware Deep Feature Embedding", "Chen Huang", "Chen ChangeLoy", "XiaoouTang", "Department of Information Engineering, The Chinese University of Hong Kong", "", "{chuang,ccloy,xtang}@ie.cuhk.edu.hk", "911", "Abstract", "Existing deep embedding methods in vision tasks are capable of learning a compact", "Euclidean space from images, where Euclidean distances correspond to a similarity", "metric. To make learning more effective and efficient, hard sample mining is", "usually employed, with samples identified through computing the Euclidean feature", "distance. However, the global Euclidean distance cannot faithfully characterize", "the true feature similarity in a complex visual feature space, where the intraclass", "['s]", "distance inahigh-density regionmaybelargerthanthe interclassdistance in", "low-density regions. In this paper, we introduce a Position-Dependent Deep Metric", "(PDDM) unit, which is capable of learning a similarity metric adaptive to local", "feature structure. The metric can be used to select genuinely hard samples in a", "local neighborhood to guide the deep embedding learning in an online and robust", "manner.Thenewlayeris appealing in thatitis pluggabletoany convolutional", "networks and is trained end-to-end. Our local similarity-aware feature embedding", "1A0680911", "notonlydemonstratesfasterconvergenceandboostedperformanceontwocomplex", "image retrieval datasets, its large margin nature also leads to superior generalization", "results under the large and open set scenarios of transfer learning and zero-shot", "learning on ImageNet 2010 and ImageNet-10K datasets.", "1.", "Introduction", "Deep embedding methods aim at learning a compact feature embedding f(α) e Rd from image r", "using a deep convolutional neural network (CNN). They have been increasingly adopted in a variety of", "vision tasks such as product visual search [1, 14, 29, 33] and face verification [13, 27]. The embedding", "objective is usually in a Euclidean sense:the Euclidean distance Di,j=llf(ci)-f()ll2 between", "two feature vectors should preserve their semantic relationship encoded pairwise (by contrastive", "loss [1]), in triplets [27, 33] or even higher order relationships (e.g., by lifted structured loss [29].", "It is widely observed that an effective data sampling strategy is crucial to ensure the quality and", "learningefficiencyofdeepembedding,asthereareoftenmanymoreeasyexamplesthanthose", "meaningful hard examples. Selecting overly easy samples can in practice lead to slow convergence", "and poor performance since many of them satisfy the constraint well and give nearly zero loss,", "without exerting any effect on parameter update during the back-propagation [3].,Hence hard", "example mining [7] becomes an indispensable step in state-of-the-art deep embedding methods.", "These methods usually choose hard samples by computing the convenient Euclidean distance in the", "embedding space. For instance, in [27, 29], hard negatives with small Euclidean distances are found", "online in a mini-batch. An exception is [33] where an online reservoir importance sampling scheme", "is proposed to sample discriminative triplets by relevance scores. Nevertheless, these scores are", "computedofflinewithdifferenthand-craftedfeaturesanddistancemetrics,whichissuboptimal", "We question the effectiveness of using a single and global Euclidean distance metric for finding hard", "samples, especially for real-world vision tasks that exhibit complex feature variations due to pose,", "lighting, and appearance. As shown in a fine-grained bird image retrieval example in Figure 1(a),", "the diversity of feature patterns learned for each class throughout the feature space can easily lead", "30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain." ]

{ "arxiv_id": "1610.08904", "language": "en", "page_idx": 0 }

[ [ "0.4346", "0.089", "0.5024", "0.0903", "0.5024", "0.1016", "0.4346", "0.101" ], [ "0.546", "0.0903", "0.6406", "0.0903", "0.6406", "0.09784", "0.546", "0.09784" ], [ "0.6807", "0.0884", "0.822", "0.0884", "0.822", "0.0985", "0.6807", "0.0985" ], [ "0.2532", "0.09155", "0.3245", "0.09155", "0.3245", "0.1016", "0.2532", "0.1016" ], [ "0.1772", "0.09155", "0.1993", "0.09155", "0.1993", "0.1042", "0.1772", "0.1042" ], [ "0.3284", "0.09216", "0.4126", "0.09216", "0.4126", "0.09973", "0.3284", "0.09973" ], [ "0.505", "0.09155", "0.528", "0.09155", "0.528", "0.1029", "0.505", "0.1029" ], [ "0.5522", "0.09973", "0.6357", "0.09973", "0.6357", "0.1067", "0.5522", "0.1067" ], [ "0.6978", "0.09973", "0.8057", "0.09973", "0.8057", "0.1067", "0.6978", "0.1067" ], [ "0.455", "0.11615", "0.4739", "0.11615", "0.4739", "0.1206", "0.455", "0.1206" ], [ "0.6055", "0.113", "0.66", "0.113", "0.66", "0.1237", "0.6055", "0.1237" ], [ "0.6113", "0.125", "0.6357", "0.125", "0.6357", "0.132", "0.6113", "0.132" ], [ "0.665", "0.125", "0.6807", "0.125", "0.6807", "0.132", "0.665", "0.132" ], [ "0.1896", "0.1976", "0.1993", "0.1976", "0.1993", "0.2008", "0.1896", "0.2008" ], [ "0.584", "0.2153", "0.6313", "0.2153", "0.6313", "0.2229", "0.584", "0.2229" ], [ "0.511", "0.2267", "0.5254", "0.2267", "0.5254", "0.2361", "0.511", "0.2361" ], [ "0.589", "0.2299", "0.615", "0.2299", "0.615", "0.238", "0.589", "0.238" ], [ "0.6284", "0.2374", "0.634", "0.2374", "0.634", "0.2424", "0.6284", "0.2424" ], [ "0.6665", "0.2507", "0.697", "0.2507", "0.697", "0.2588", "0.6665", "0.2588" ], [ "0.4182", "0.2727", "0.492", "0.2727", "0.492", "0.2827", "0.4182", "0.2827" ], [ "0.1805", "0.2886", "0.1969", "0.2886", "0.1969", "0.2954", "0.1805", "0.2954" ], [ "0.1978", "0.286", "0.2361", "0.286", "0.2361", "0.2966", "0.1978", "0.2966" ], [ "0.2507", "0.2886", "0.3096", "0.2886", "0.3096", "0.2954", "0.2507", "0.2954" ], [ "0.524", "0.2974", "0.6323", "0.2998", "0.6323", "0.31", "0.524", "0.308" ], [ "0.71", "0.2974", "0.8057", "0.2986", "0.8057", "0.3093", "0.71", "0.308" ], [ "0.3857", "0.3093", "0.4192", "0.3093", "0.4192", "0.3164", "0.3857", "0.3164" ], [ "0.384", "0.3018", "0.4004", "0.3018", "0.4004", "0.3062", "0.384", "0.3062" ], [ "0.638", "0.3113", "0.722", "0.3113", "0.722", "0.318", "0.638", "0.318" ], [ "0.1748", "0.3352", "0.8228", "0.3352", "0.8228", "0.3484", "0.1748", "0.3484" ], [ "0.1748", "0.3499", "0.8228", "0.3499", "0.8228", "0.363", "0.1748", "0.363" ], [ "0.1732", "0.363", "0.824", "0.3623", "0.824", "0.377", "0.1732", "0.3774" ], [ "0.1765", "0.379", "0.8228", "0.379", "0.8228", "0.3894", "0.1765", "0.3894" ], [ "0.1748", "0.3914", "0.8228", "0.3914", "0.8228", "0.4048", "0.1748", "0.4048" ], [ "0.1748", "0.4048", "0.826", "0.4048", "0.826", "0.4185", "0.1748", "0.4185" ], [ "0.1748", "0.4185", "0.7485", "0.4185", "0.7485", "0.432", "0.1748", "0.432" ], [ "0.1748", "0.4702", "0.8228", "0.4702", "0.8228", "0.4841", "0.1748", "0.4841" ], [ "0.1765", "0.4849", "0.8228", "0.4849", "0.8228", "0.4988", "0.1765", "0.4988" ], [ "0.1748", "0.4988", "0.8228", "0.4988", "0.8228", "0.512", "0.1748", "0.512" ], [ "0.1732", "0.511", "0.824", "0.512", "0.824", "0.5264", "0.1732", "0.526" ], [ "0.1748", "0.5264", "0.826", "0.5264", "0.826", "0.5405", "0.1748", "0.5405" ], [ "0.1748", "0.5405", "0.6763", "0.5405", "0.6763", "0.5537", "0.1748", "0.5537" ], [ "0.1748", "0.56", "0.824", "0.56", "0.824", "0.5737", "0.1748", "0.5737" ], [ "0.1748", "0.5747", "0.824", "0.5747", "0.824", "0.5884", "0.1748", "0.5884" ], [ "0.1748", "0.5884", "0.824", "0.5884", "0.824", "0.6016", "0.1748", "0.6016" ], [ "0.1765", "0.604", "0.8228", "0.604", "0.8228", "0.6143", "0.1765", "0.6143" ], [ "0.1748", "0.616", "0.8228", "0.616", "0.8228", "0.63", "0.1748", "0.63" ], [ "0.1765", "0.631", "0.8228", "0.631", "0.8228", "0.6416", "0.1765", "0.6416" ], [ "0.1748", "0.643", "0.824", "0.643", "0.824", "0.657", "0.1748", "0.657" ], [ "0.1748", "0.657", "0.8228", "0.657", "0.8228", "0.6704", "0.1748", "0.6704" ], [ "0.1748", "0.6704", "0.824", "0.6704", "0.824", "0.6846", "0.1748", "0.6846" ], [ "0.1732", "0.6846", "0.822", "0.6846", "0.822", "0.6978", "0.1732", "0.6978" ], [ "0.1748", "0.6987", "0.826", "0.6987", "0.826", "0.712", "0.1748", "0.712" ], [ "0.1765", "0.712", "0.824", "0.712", "0.824", "0.726", "0.1765", "0.726" ], [ "0.1748", "0.726", "0.824", "0.726", "0.824", "0.7393", "0.1748", "0.7393" ], [ "0.1748", "0.7407", "0.824", "0.7407", "0.824", "0.754", "0.1748", "0.754" ], [ "0.1748", "0.754", "0.824", "0.754", "0.824", "0.7676", "0.1748", "0.7676" ], [ "0.1732", "0.7676", "0.6396", "0.7676", "0.6396", "0.781", "0.1732", "0.781" ], [ "0.1732", "0.786", "0.8228", "0.786", "0.8228", "0.8", "0.1732", "0.8" ], [ "0.1748", "0.802", "0.8228", "0.802", "0.8228", "0.815", "0.1748", "0.815" ], [ "0.1765", "0.8174", "0.8228", "0.8174", "0.8228", "0.8276", "0.1765", "0.8276" ], [ "0.1748", "0.83", "0.824", "0.83", "0.824", "0.8433", "0.1748", "0.8433" ], [ "0.1748", "0.8433", "0.824", "0.8433", "0.824", "0.857", "0.1748", "0.857" ], [ "0.1765", "0.8594", "0.8228", "0.8594", "0.8228", "0.869", "0.1765", "0.869" ], [ "0.1765", "0.8726", "0.8228", "0.8726", "0.8228", "0.8833", "0.1765", "0.8833" ], [ "0.1748", "0.8853", "0.824", "0.8853", "0.824", "0.8984", "0.1748", "0.8984" ], [ "0.1732", "0.8984", "0.6274", "0.8984", "0.6274", "0.912", "0.1732", "0.912" ], [ "0.4927", "0.936", "0.5063", "0.936", "0.5063", "0.9478", "0.4927", "0.9478" ] ]

[ "Horned Puffin", "DeepEuclideanmetri", "Position-Dependent Deep Metric", "Myrtle Warbler", "(a)", "NashvilleWarbler", "(b)", "Tripletloss[27.33]", "Doubleheaderhingeloss", "", "Positive sim", "Nega", "larit", "", "Mini-batch", "(c)", "(x)", "", "PDDM", "YellowWarbler", "Eles", "antTern", "ElegantTern", "Feature embedding spac", "Similarity score space", "Howthre", "Cam", "guadrupletmining", "Figure 1: (a) 2-D feature embedding (by t-SNE [19]) of the CUB-200-2011 [32] test set. The", "intraclass distance can be larger than the interclass distance under the global Euclidean metric, which", "can mislead the hard sample mining and consequently deep embedding learning.We propose a PDDM", "unitthat incorporates theabsoluteposition(i.e.,featuremean denotedbytheredtriangle)to adapt", "metric to the local feature structure. (b) Overlapped similarity distribution by the Euclidean metric (the", "similarity scores are transformed from distances by a Sigmoid-like function) vs. the well-separated", "distribution by PDDM. (c) PDDM-guided hard sample mining and embedding learning.", "to a larger intraclass Euclidean distance than the interclass distance. Such a heterogeneous feature", "distribution yields a highly overlapped similarity score distribution for the positive and negative", "pairs, as shown in the left chart of Figure 1(b). We observed similar phenomenon for the global", "Mahalanobis metric [11, 12, 21, 35, 36] in our experiments. It is not difficult to see that using a single", "and global metric would easily mislead the hard sample mining. To circumvent this issue, Cui et", "al. [3] resorted to human intervention for harvesting genuinely hard samples.", "Mitigating the aforementioned issue demands an improved metric that is adaptive to the local feature", "structure. In this study we wish to learn a local-adaptive similarity metric online, which will be", "exploited to search for high-quality hard samples in local neighborhood to facilitate a more effective", "deepembeddinglearning.Ourkeychallengeslieintheformulationofanewlayerandlossfunction", "that jointly consider the similarity metric learning, hard samples selection, and deep embedding", "learning.Existingstudies[13,14,27,29,33onlyconsiderthetwolatterobjectivesbutnottogether", "with the first. To this end, we propose a new Position-Dependent Deep Metric (PDDM) unit for", "similarity metric learning. It is readily pluggable to train end-to-end with an existing deep embedding", "learning CNN. We formulate the PDDM such that it learns locally adaptive metric (unlike the", "global Euclidean metric), through a non-linear regression on both the absolute feature difference", "and feature mean (which encodes absolute position) of a data pair. As depicted in the right chart of", "Figure 1(b), the proposed metric yields a similarity score distribution that is more distinguishable", "than the conventional Euclidean metric. As shown in Figure 1(c), hard samples are mined from", "the resulting similarity score space and used to optimize the feature embedding space in a seamless", "manner. The similarity metric learning in PDDM and embedding learning in the associated CNN are", "jointly optimized using a novel large-margin double-header hinge loss.", "Image retrieval experiments on two challenging real-world vision datasets, CUB-200-2011 [32] and", "CARS196 [15], show that our local similarity-aware feature embedding significantly outperforms", "state-of-the-artdeepembeddingmethodsthatcomewithouttheonlinemetriclearningandassociated", "hard sample mining scheme.Moreover, the proposed approach incurs a far lower computational", "cost and encourages faster convergence than those structured embedding methods (e.g., [29]), which", "needtocomputeafullyconnecteddensematrixof pairwisedistancesinamini-batch.Wefurther", "demonstrateourlearned embeddingisgeneralizabletonew classes inlarge openset scenarios.This", "is validated in the transfer learning and zero-shot learning (using the ImageNet hierarchy as auxiliary", "knowledge) tasks on ImageNet 2010 and ImageNet-10K [5] datasets.", "2" ]

{ "arxiv_id": "1610.08904", "language": "en", "page_idx": 1 }

[ [ "0.5195", "0.09155", "0.5537", "0.09155", "0.5537", "0.0985", "0.5195", "0.0985" ], [ "0.3186", "0.0985", "0.343", "0.0985", "0.343", "0.1067", "0.3186", "0.1067" ], [ "0.3635", "0.0985", "0.392", "0.0985", "0.392", "0.1067", "0.3635", "0.1067" ], [ "0.4355", "0.0985", "0.4436", "0.0985", "0.4436", "0.1029", "0.4355", "0.1029" ], [ "0.1772", "0.108", "0.2296", "0.108", "0.2296", "0.11554", "0.1772", "0.11554" ], [ "0.5146", "0.108", "0.539", "0.10986", "0.5376", "0.11743", "0.513", "0.11554" ], [ "0.714", "0.1048", "0.74", "0.1048", "0.74", "0.11237", "0.714", "0.11237" ], [ "0.2418", "0.11426", "0.2607", "0.11426", "0.2607", "0.1187", "0.2418", "0.1187" ], [ "0.2615", "0.11237", "0.3064", "0.11237", "0.3064", "0.1225", "0.2615", "0.1225" ], [ "0.3945", "0.11426", "0.4631", "0.113", "0.4631", "0.132", "0.3945", "0.1332" ], [ "0.183", "0.1206", "0.2214", "0.1206", "0.2214", "0.1338", "0.183", "0.1338" ], [ "0.367", "0.1218", "0.392", "0.1218", "0.392", "0.13", "0.367", "0.13" ], [ "0.614", "0.1237", "0.63", "0.1237", "0.63", "0.1276", "0.614", "0.1276" ], [ "0.6724", "0.125", "0.6895", "0.125", "0.6895", "0.1288", "0.6724", "0.1288" ], [ "0.615", "0.132", "0.6396", "0.132", "0.6396", "0.1433", "0.615", "0.1433" ], [ "0.3555", "0.1445", "0.3938", "0.1445", "0.3938", "0.1547", "0.3555", "0.1547" ], [ "0.2279", "0.1276", "0.308", "0.13", "0.308", "0.1445", "0.2279", "0.1427" ], [ "0.3203", "0.1483", "0.353", "0.1483", "0.353", "0.1528", "0.3203", "0.1528" ], [ "0.4314", "0.1459", "0.4495", "0.1459", "0.4495", "0.154", "0.4314", "0.154" ], [ "0.71", "0.159", "0.722", "0.159", "0.722", "0.1635", "0.71", "0.1635" ], [ "0.2507", "0.1497", "0.298", "0.1528", "0.2974", "0.1642", "0.25", "0.161" ], [ "0.2632", "0.1755", "0.2942", "0.1755", "0.2942", "0.1843", "0.2632", "0.1843" ], [ "0.3162", "0.1704", "0.3938", "0.1686", "0.3938", "0.1787", "0.3162", "0.1812" ], [ "0.619", "0.178", "0.638", "0.1831", "0.634", "0.1913", "0.615", "0.1863" ], [ "0.3464", "0.1901", "0.4248", "0.1901", "0.4248", "0.2008", "0.3464", "0.2008" ], [ "0.4412", "0.1875", "0.505", "0.1875", "0.505", "0.1976", "0.4412", "0.1976" ], [ "0.4592", "0.1989", "0.4927", "0.1989", "0.4927", "0.2058", "0.4592", "0.2058" ], [ "0.3406", "0.2122", "0.3635", "0.2122", "0.3635", "0.2235", "0.3406", "0.2235" ], [ "0.3586", "0.2122", "0.511", "0.2122", "0.511", "0.2229", "0.3586", "0.2229" ], [ "0.6665", "0.2102", "0.7573", "0.2084", "0.7573", "0.2229", "0.6665", "0.2247" ], [ "0.1732", "0.2311", "0.824", "0.2329", "0.824", "0.2468", "0.1732", "0.2443" ], [ "0.1732", "0.2468", "0.312", "0.2468", "0.312", "0.26", "0.1732", "0.26" ], [ "0.1765", "0.2766", "0.1896", "0.2766", "0.1896", "0.2886", "0.1765", "0.2886" ], [ "0.1978", "0.276", "0.3186", "0.276", "0.3186", "0.2903", "0.1978", "0.2903" ], [ "0.1748", "0.3093", "0.824", "0.3093", "0.824", "0.3225", "0.1748", "0.3225" ], [ "0.1748", "0.3225", "0.824", "0.3225", "0.824", "0.336", "0.1748", "0.336" ], [ "0.1765", "0.3384", "0.824", "0.3384", "0.824", "0.3484", "0.1765", "0.3484" ], [ "0.1732", "0.351", "0.824", "0.3499", "0.824", "0.3643", "0.1732", "0.3655" ], [ "0.1748", "0.3643", "0.8228", "0.3643", "0.8228", "0.3774", "0.1748", "0.3774" ], [ "0.1748", "0.379", "0.822", "0.379", "0.822", "0.3894", "0.1748", "0.3894" ], [ "0.1748", "0.3926", "0.824", "0.3926", "0.824", "0.406", "0.1748", "0.406" ], [ "0.1748", "0.4048", "0.826", "0.4048", "0.826", "0.4185", "0.1748", "0.4185" ], [ "0.1765", "0.4204", "0.8228", "0.4204", "0.8228", "0.4312", "0.1765", "0.4312" ], [ "0.1748", "0.433", "0.824", "0.433", "0.824", "0.4463", "0.1748", "0.4463" ], [ "0.1732", "0.4475", "0.824", "0.4475", "0.824", "0.461", "0.1732", "0.461" ], [ "0.1748", "0.461", "0.5376", "0.461", "0.5376", "0.4749", "0.1748", "0.4749" ], [ "0.1748", "0.4797", "0.824", "0.4797", "0.824", "0.4944", "0.1748", "0.4944" ], [ "0.1748", "0.4956", "0.824", "0.4956", "0.824", "0.509", "0.1748", "0.509" ], [ "0.1748", "0.509", "0.824", "0.509", "0.824", "0.523", "0.1748", "0.523" ], [ "0.1748", "0.523", "0.824", "0.523", "0.824", "0.536", "0.1748", "0.536" ], [ "0.1765", "0.5386", "0.8228", "0.5386", "0.8228", "0.549", "0.1765", "0.549" ], [ "0.1748", "0.5503", "0.824", "0.5503", "0.824", "0.5645", "0.1748", "0.5645" ], [ "0.1748", "0.5645", "0.824", "0.5645", "0.824", "0.5776", "0.1748", "0.5776" ], [ "0.1748", "0.5776", "0.824", "0.5776", "0.824", "0.5913", "0.1748", "0.5913" ], [ "0.1748", "0.5913", "0.826", "0.5913", "0.826", "0.605", "0.1748", "0.605" ], [ "0.1732", "0.605", "0.824", "0.604", "0.824", "0.6187", "0.1732", "0.619" ], [ "0.1765", "0.6206", "0.822", "0.6206", "0.822", "0.631", "0.1765", "0.631" ], [ "0.1748", "0.6333", "0.8228", "0.6333", "0.8228", "0.6465", "0.1748", "0.6465" ], [ "0.1732", "0.6465", "0.826", "0.6465", "0.826", "0.66", "0.1732", "0.66" ], [ "0.1732", "0.66", "0.824", "0.66", "0.824", "0.674", "0.1732", "0.674" ], [ "0.1732", "0.674", "0.8228", "0.674", "0.8228", "0.6875", "0.1732", "0.6875" ], [ "0.1724", "0.686", "0.826", "0.688", "0.826", "0.7026", "0.1724", "0.7007" ], [ "0.1748", "0.7026", "0.8037", "0.7026", "0.8037", "0.7134", "0.1748", "0.7134" ], [ "0.1708", "0.7373", "0.5376", "0.7393", "0.5376", "0.757", "0.1708", "0.755" ], [ "0.1748", "0.7734", "0.824", "0.7734", "0.824", "0.787", "0.1748", "0.787" ], [ "0.1748", "0.787", "0.824", "0.787", "0.824", "0.8003", "0.1748", "0.8003" ], [ "0.1748", "0.802", "0.824", "0.802", "0.824", "0.815", "0.1748", "0.815" ], [ "0.1732", "0.8164", "0.822", "0.8164", "0.822", "0.83", "0.1732", "0.83" ], [ "0.1748", "0.83", "0.8228", "0.83", "0.8228", "0.8447", "0.1748", "0.8447" ], [ "0.1732", "0.8433", "0.822", "0.8433", "0.822", "0.857", "0.1732", "0.857" ], [ "0.1748", "0.858", "0.8228", "0.858", "0.8228", "0.872", "0.1748", "0.872" ], [ "0.1748", "0.8706", "0.824", "0.8706", "0.824", "0.884", "0.1748", "0.884" ], [ "0.1724", "0.884", "0.826", "0.8833", "0.826", "0.898", "0.1724", "0.8984" ], [ "0.1748", "0.8984", "0.5107", "0.8984", "0.5107", "0.912", "0.1748", "0.912" ], [ "0.492", "0.9346", "0.5073", "0.9346", "0.5073", "0.9497", "0.492", "0.9497" ] ]

[ "Similarit", "Ox", "CNN", "", "Mini-batches", "PDDM", "PDDM", "Tard", "quadruple", "(20-6", "(xy)", "CNN", "Fr", "T", "", "CNN", "(i,)=arg min s,", "-", "H", "T.", "argmaxs,", "max S,", "OXCNN", "", "Feature embedding", "Double-header", "inge loss", "(a)", "Overall networkarchitecture", "(b)PDDMunit", "Figure 2:(a)The overall network architecture.All CNNs have shared architectures and parameters.", "(b) The PDDM unit.", "2", "Relatedwork", "Hard sample mining in deep learning: Hard sample mining is a popular technique used in computer", "vision for training robust classifier. The method aims at augmenting a training set progressively with", "falsepositiveexampleswiththemodellearnedsofar.Itisthecoreofmanysuccessful visionsolutions", "e.g. pedestrian detection [4, 7]. In a similar spirit, contemporary deep embedding methods [27, 29]", "choose hard samples in a mini-batch by computing the Euclidean distance in the embedding space", "For instance,Schroff etal.[27l selected onlinethe semi-hard negativesamples withrelatively smal", "Euclidean distances. Wang et al. [33] proposed an online reservoir importance sampling algorithm", "to sample triplets by relevance scores, which are computed offline with different distance metrics.", "Similarstudiesonimagedescriptorlearning[28]andunsupervisedfeaturelearning[34alsoselect", "hard samples according to the Euclidean distance-based losses in their respective CNNs. We argue in", "this paper that the global Euclidean distance is a suboptimal similarity metric for hard sample mining,", "and propose a locally adaptivemetric for bettermining.", "Metric learning: An effective similarity metric is at the core of hard sample mining. Euclidean", "distance is the simplest similarity metric, and it is widely used by current deep embedding methods", "where Euclidean feature distances directly correspond the similarity. Similarities can be encoded", "pairwise with a contrastive loss [1] or in more flexible triplets [27, 33]. Song et al. [29] extended to", "evenhigherordersimilarityconstraintsbyliftingthepairwisedistanceswithinamini-batchtothe", "dense matrix of pairwise distances. Beyond Euclidean metric, one can actually turn to the parametric", "Mahalanobis metric instead. Representative works [12, 36] minimize the Mahalanobis distance", "between positive sample pairs while maximizing the distance between negative pairs. Alternatives", "directly optimize the Mahalanobis metric for nearest neighbor classification via the method of", "Neighbourhood Component Analysis (NCA) [11], Large Margin Nearest Neighbor (LMNN) [35] or", "NearestClassMean(NCM)[21l.However.thecommondrawbackoftheMahalanobisandEuclidean", "metrics is that they are both global and are far from being ideal in the presence of heterogeneous", "feature distribution (see Figure 1(a)). An intuitive remedy would be to learn multiple metrics [9],", "which would be computationally expensive though. Xiong et al. [37] proposed a single adaptive", "metric using the absolute position information in random forest classifiers. Our approach shares the", "similar intuition, but incorporates the position information by a deep CNN in a more principled way,", "andcanjointlylearn similarity-awaredeepfeaturesinsteadof usinghand-crafted onesasin[37]", "3 Local similarity-aware deep embedding", "Let X = [(ri, yi)) be an imagery dataset, where yi is the class label of image Ci. Our goal is to jointly", "learn a deep feature embedding f(α) from image r into a feature space Rd, and a similarity metric", "Si.j = S(f(ci), f(r,)) e IR1, such that the metric can robustly select hard samples online to learn a", "discriminative, similarity-aware feature embedding. Ideally,the learned features (f(r:), f(,))from", "the set of positive pairs P = (i, j)lyi = yi} should be close to each other with a large similarity", "score Si.j, while the learned features from the set of negative pairs N = {(i, )lyi yif should be", "pushed far away with a small similarity score. Importantly, this relationship should hold independent", "of the (heterogeneous) feature distribution in Rd, where a global metric Si., can fail. To adapt Si., to", "the latent structure of feature embeddings, we propose a Position-Dependent Deep Metric (PDDM)", "unit that can be trained end-to-end, see Figure 2(b).", "3" ]

{ "arxiv_id": "1610.08904", "language": "en", "page_idx": 2 }

[ [ "0.1748", "0.09155", "0.8228", "0.09155", "0.8228", "0.1061", "0.1748", "0.1061" ], [ "0.1748", "0.1067", "0.8228", "0.1067", "0.8228", "0.1206", "0.1748", "0.1206" ], [ "0.1732", "0.1218", "0.8228", "0.1206", "0.8228", "0.1351", "0.1732", "0.1364" ], [ "0.1748", "0.1364", "0.824", "0.1364", "0.824", "0.1497", "0.1748", "0.1497" ], [ "0.1732", "0.1497", "0.824", "0.1483", "0.824", "0.1635", "0.1732", "0.1642" ], [ "0.1732", "0.1635", "0.824", "0.1622", "0.824", "0.1768", "0.1732", "0.178" ], [ "0.1748", "0.178", "0.6445", "0.178", "0.6445", "0.1913", "0.1748", "0.1913" ], [ "0.1748", "0.2084", "0.5", "0.2084", "0.5", "0.2216", "0.1748", "0.2216" ], [ "0.1765", "0.2329", "0.822", "0.2329", "0.822", "0.2474", "0.1765", "0.2474" ], [ "0.1765", "0.2468", "0.689", "0.2468", "0.689", "0.2615", "0.1765", "0.2615" ], [ "0.681", "0.2468", "0.824", "0.2468", "0.824", "0.2615", "0.681", "0.2615" ], [ "0.1732", "0.2588", "0.826", "0.2588", "0.826", "0.2766", "0.1732", "0.2766" ], [ "0.1748", "0.2747", "0.699", "0.2747", "0.699", "0.2886", "0.1748", "0.2886" ], [ "0.25", "0.3018", "0.3628", "0.3018", "0.3628", "0.3164", "0.25", "0.3164" ], [ "0.5513", "0.3005", "0.751", "0.3018", "0.751", "0.3164", "0.5513", "0.315" ], [ "0.8022", "0.303", "0.826", "0.303", "0.826", "0.3164", "0.8022", "0.3164" ], [ "0.5156", "0.307", "0.527", "0.307", "0.527", "0.3113", "0.5156", "0.3113" ], [ "0.357", "0.3093", "0.4885", "0.305", "0.4902", "0.3257", "0.3586", "0.33" ], [ "0.3677", "0.3037", "0.423", "0.3037", "0.423", "0.3145", "0.3677", "0.3145" ], [ "0.492", "0.3208", "0.5513", "0.3193", "0.5513", "0.33", "0.492", "0.3308" ], [ "0.1748", "0.3384", "0.826", "0.3384", "0.826", "0.3528", "0.1748", "0.3528" ], [ "0.1748", "0.3623", "0.8228", "0.3623", "0.8228", "0.3757", "0.1748", "0.3757" ], [ "0.1748", "0.377", "0.822", "0.377", "0.822", "0.3901", "0.1748", "0.3901" ], [ "0.1732", "0.3894", "0.822", "0.3901", "0.822", "0.4048", "0.1732", "0.404" ], [ "0.1748", "0.4048", "0.824", "0.4048", "0.824", "0.4185", "0.1748", "0.4185" ], [ "0.1748", "0.4185", "0.824", "0.4185", "0.824", "0.432", "0.1748", "0.432" ], [ "0.1765", "0.4343", "0.8228", "0.4343", "0.8228", "0.4443", "0.1765", "0.4443" ], [ "0.1765", "0.4463", "0.824", "0.4463", "0.824", "0.4602", "0.1765", "0.4602" ], [ "0.1748", "0.459", "0.6733", "0.4583", "0.6733", "0.473", "0.1748", "0.4734" ], [ "0.1765", "0.4797", "0.8228", "0.4797", "0.8228", "0.4944", "0.1765", "0.4944" ], [ "0.1748", "0.4944", "0.8228", "0.4944", "0.8228", "0.5083", "0.1748", "0.5083" ], [ "0.1732", "0.5083", "0.822", "0.5083", "0.822", "0.5215", "0.1732", "0.5215" ], [ "0.1748", "0.5215", "0.826", "0.5215", "0.826", "0.535", "0.1748", "0.535" ], [ "0.1748", "0.535", "0.824", "0.535", "0.824", "0.549", "0.1748", "0.549" ], [ "0.1724", "0.5474", "0.826", "0.547", "0.826", "0.5645", "0.1724", "0.565" ], [ "0.1765", "0.5645", "0.824", "0.5645", "0.824", "0.5776", "0.1765", "0.5776" ], [ "0.1772", "0.579", "0.8228", "0.579", "0.8228", "0.5923", "0.1772", "0.5923" ], [ "0.1732", "0.5903", "0.824", "0.589", "0.824", "0.607", "0.1732", "0.608" ], [ "0.1724", "0.606", "0.7417", "0.604", "0.7417", "0.6206", "0.1724", "0.6226" ], [ "0.353", "0.6274", "0.6763", "0.6274", "0.6763", "0.6426", "0.353", "0.6426" ], [ "0.3381", "0.6426", "0.669", "0.6445", "0.669", "0.662", "0.3381", "0.66" ], [ "0.4902", "0.672", "0.6274", "0.672", "0.6274", "0.686", "0.4902", "0.686" ], [ "0.8022", "0.672", "0.826", "0.672", "0.826", "0.686", "0.8022", "0.686" ], [ "0.4468", "0.6724", "0.4795", "0.6724", "0.4795", "0.685", "0.4468", "0.685" ], [ "0.3708", "0.675", "0.3906", "0.6724", "0.3938", "0.6816", "0.3718", "0.6846" ], [ "0.1748", "0.702", "0.826", "0.702", "0.826", "0.7153", "0.1748", "0.7153" ], [ "0.1732", "0.714", "0.824", "0.714", "0.824", "0.729", "0.1732", "0.729" ], [ "0.1748", "0.73", "0.8145", "0.73", "0.8145", "0.7437", "0.1748", "0.7437" ], [ "0.1748", "0.7524", "0.824", "0.7524", "0.824", "0.7666", "0.1748", "0.7666" ], [ "0.1748", "0.7676", "0.824", "0.7676", "0.824", "0.781", "0.1748", "0.781" ], [ "0.1732", "0.78", "0.8228", "0.78", "0.8228", "0.7935", "0.1732", "0.7935" ], [ "0.1748", "0.7954", "0.824", "0.7954", "0.824", "0.8096", "0.1748", "0.8096" ], [ "0.1765", "0.81", "0.8228", "0.81", "0.8228", "0.821", "0.1765", "0.821" ], [ "0.1748", "0.8228", "0.8145", "0.8228", "0.8145", "0.836", "0.1748", "0.836" ], [ "0.1748", "0.8423", "0.824", "0.8423", "0.824", "0.856", "0.1748", "0.856" ], [ "0.1765", "0.8594", "0.8228", "0.8594", "0.8228", "0.869", "0.1765", "0.869" ], [ "0.1748", "0.8706", "0.824", "0.8706", "0.824", "0.884", "0.1748", "0.884" ], [ "0.1748", "0.8853", "0.8228", "0.8853", "0.8228", "0.8984", "0.1748", "0.8984" ], [ "0.1732", "0.8984", "0.824", "0.8984", "0.824", "0.912", "0.1732", "0.912" ], [ "0.492", "0.9375", "0.5063", "0.9375", "0.5063", "0.9478", "0.492", "0.9478" ] ]

[ "The overall network architecture is shown in Figure 2(a). First, we use PDDM to compute similarity", "scores for the mini-batch samples during a particular forward pass. The scores are used to select", "one hard quadruplet from the local sets of positive pairs P e P and negative pairs N e N in the", "batch. Then each sample in the hard quadruplet is separately fed into four identical CNNs with shared", "parameters W to extract d-dimensional features. Finally, a discriminative double-header hinge loss is", "applied to both the similarity score and feature embeddings. This enables us to jointly optimize the", "two that benefit each other. We will provide the details in the following.", "3.1PDDM learning and hard samplemining", "Given a feature pair (fw(c;), fw(c;)) extracted from images &; and , by an embedding function", "fw () parameterized by W, we wish to obtain an ideal similarity score yi,j =", "= 1if (i,j) E P, and", "i,j = 0 if (i, j) e N. Hence, we seek the optimal similarity metric S*(-) from an appropriate", "function space H, and also seek the optimal feature embedding parameters W*:", "(S*( ), W*) =", "l(S(fw(ci),fw(a)),yi),", "(1)", "", "saen,w|PUN)", "argmin", "(i,j)EPUN", "where l() is some loss function. We will omit the parameters W of f() in the following for brevity.", "Adapting to local feature structure. The standard Euclidean or Mahalanobis metric can be seen as a", "special form of function S(,) that is based solely on the feature difference vector u =f(ci)-f(c,)", "or its linearly transformed version. These metrics are suboptimal in a heterogeneous embedding", "space, thus could easily fail the searching of genuinely hard samples. On the contrary, the proposed", "PDDM leverages the absolute feature position to adapt the metric throughout the embedding space", "Specifically,inspiredby37l,apartfromthefeaturedifferencevectoru,weadditionallyincorporate", "the feature mean vector u = (f(α;) + f(r,)/2 to encode the absolute position. Unlike [37], we", "formulate a principled learnable similarity metric from u and in our CNN.", "Formally, as shown in Figure 2(b), we first normalize the features f(ri) and f(c:) onto the unit", "hypersphere, i.e., lf(α)/2 = 1, in order to maintain feature comparability. Such normalized features", "are used to compute their relative and absolute positions encoded in u and v, each followed by", "a fully connected layer, an elementwise ReLU nonlinear function o($) = max(O, $), and again,", "2-normalization r()=/rl2.To treat u and v differently, the fully connected layers applied to", "them are not shared, parameterized by Wu e IRdxd, bu IRd and Wu e IRdxd, bu e Rd, respectively.", "The nonlinearities ensure the model is not trivially equivalent to be the mapping from f(c:) and", "f() themselves. Then we concatenate the mapped u' and ' vectors and pass them through another", "fully connected layer parameterized by W。 e R2dxd, be IRd and the ReLU function, and finally", "map to a score Si,j = S(f(ri), f(r;)) e R1 via W, e IRdx1,bs e IR1. To summarize:", "u=lf(ri)-f(r)l,=(f(ri)+f(c) /2,", "u=r(o(Wuu+bu)),u=r(o(Wuu+bu),", ", Sij =Wsc+bs", "(2)", "+bc", "W", "In this way, we transform the seeking of the similarity metric function S(, ) into the joint learning of", "CNN parameters (Wu, W, We, W., bu, bu, be, b, for the PDDM unit, and W for feature embeddings)", "The parameters collectively define a flexible nonlinear regression function for the similarity score.", "Double-header hinge loss.To optimize all these CNN parameters, we can choose a standard", "regression loss function l(), e.g., logistic regression loss. Or alternatively, we can cast the problem as", "a binary classification one as in [37]. However, in both cases the CNN is prone to overfitting, because", "the supervisory binary similarity labels yi,j E [0, 1] tend to independently push the scores towards", "twosinglepoints.Whileinpractice,thesimilarityscoresofpositiveandnegativepairsliveona1-D", "manifold following some distribution patterns on heterogeneous data, as illustrated in Figure I(b).", "This motivates us to design a loss function l() to separate the similarity distributions, instead of in", "anindependentpointwisewaythatisnoise-sensitive.OneintuitiveoptionistoimposetheFisher", "criterion [20] on the similarity scores, i.e., maximizing the ratio between the interclass and intraclass", "scatters of scores. Similarly,it can be reduced to maximize (μp -μn)?/(Varp+ Varn) in our", "1-D case, where μ and Var are the mean and variance of each score distribution. Unfortunately, the", "4" ]

{ "arxiv_id": "1610.08904", "language": "en", "page_idx": 3 }