Datasets:

librarian-bots

/

dataset_abstracts

like 2

{ "abstract": "Pathology diagnosis based on EEG signals and decoding brain activity holds\nimmense importance in understanding neurological disorders. With the\nadvancement of artificial intelligence methods and machine learning techniques,\nthe potential for accurate data-driven diagnoses and effective treatments has\ngrown significantly. However, applying machine learning algorithms to\nreal-world datasets presents diverse challenges at multiple levels. The\nscarcity of labelled data, especially in low regime scenarios with limited\navailability of real patient cohorts due to high costs of recruitment,\nunderscores the vital deployment of scaling and transfer learning techniques.\nIn this study, we explore a real-world pathology classification task to\nhighlight the effectiveness of data and model scaling and cross-dataset\nknowledge transfer. As such, we observe varying performance improvements\nthrough data scaling, indicating the need for careful evaluation and labelling.\nAdditionally, we identify the challenges of possible negative transfer and\nemphasize the significance of some key components to overcome distribution\nshifts and potential spurious correlations and achieve positive transfer. We\nsee improvement in the performance of the target model on the target (NMT)\ndatasets by using the knowledge from the source dataset (TUAB) when a low\namount of labelled data was available. Our findings indicate a small and\ngeneric model (e.g. ShallowNet) performs well on a single dataset, however, a\nlarger model (e.g. TCN) performs better on transfer and learning from a larger\nand diverse dataset.", "title": "Amplifying Pathological Detection in EEG Signaling Pathways through Cross-Dataset Transfer Learning", "url": "http://arxiv.org/abs/2309.10910v1" }

{ "text_length": 1725 }

{ "abstract": "This paper describes NEREL-BIO -- an annotation scheme and corpus of PubMed\nabstracts in Russian and smaller number of abstracts in English. NEREL-BIO\nextends the general domain dataset NEREL by introducing domain-specific entity\ntypes. NEREL-BIO annotation scheme covers both general and biomedical domains\nmaking it suitable for domain transfer experiments. NEREL-BIO provides\nannotation for nested named entities as an extension of the scheme employed for\nNEREL. Nested named entities may cross entity boundaries to connect to shorter\nentities nested within longer entities, making them harder to detect.\n NEREL-BIO contains annotations for 700+ Russian and 100+ English abstracts.\nAll English PubMed annotations have corresponding Russian counterparts. Thus,\nNEREL-BIO comprises the following specific features: annotation of nested named\nentities, it can be used as a benchmark for cross-domain (NEREL -> NEREL-BIO)\nand cross-language (English -> Russian) transfer. We experiment with both\ntransformer-based sequence models and machine reading comprehension (MRC)\nmodels and report their results.\n The dataset is freely available at https://github.com/nerel-ds/NEREL-BIO.", "title": "NEREL-BIO: A Dataset of Biomedical Abstracts Annotated with Nested Named Entities", "url": "http://arxiv.org/abs/2210.11913v1" }

{ "text_length": 1294 }

{ "abstract": "Crime is an unlawful act that carries legal repercussions. Bangladesh has a\nhigh crime rate due to poverty, population growth, and many other\nsocio-economic issues. For law enforcement agencies, understanding crime\npatterns is essential for preventing future criminal activity. For this\npurpose, these agencies need structured crime database. This paper introduces a\nnovel crime dataset that contains temporal, geographic, weather, and\ndemographic data about 6574 crime incidents of Bangladesh. We manually gather\ncrime news articles of a seven year time span from a daily newspaper archive.\nWe extract basic features from these raw text. Using these basic features, we\nthen consult standard service-providers of geo-location and weather data in\norder to garner these information related to the collected crime incidents.\nFurthermore, we collect demographic information from Bangladesh National Census\ndata. All these information are combined that results in a standard machine\nlearning dataset. Together, 36 features are engineered for the crime prediction\ntask. Five supervised machine learning classification algorithms are then\nevaluated on this newly built dataset and satisfactory results are achieved. We\nalso conduct exploratory analysis on various aspects the dataset. This dataset\nis expected to serve as the foundation for crime incidence prediction systems\nfor Bangladesh and other countries. The findings of this study will help law\nenforcement agencies to forecast and contain crime as well as to ensure optimal\nresource allocation for crime patrol and prevention.", "title": "Crime Prediction using Machine Learning with a Novel Crime Dataset", "url": "http://arxiv.org/abs/2211.01551v1" }

{ "text_length": 1679 }

{ "abstract": "Despite the progress we have recorded in the last few years in multilingual\nnatural language processing, evaluation is typically limited to a small set of\nlanguages with available datasets which excludes a large number of low-resource\nlanguages. In this paper, we created SIB-200 -- a large-scale open-sourced\nbenchmark dataset for topic classification in 200 languages and dialects to\naddress the lack of evaluation dataset for Natural Language Understanding\n(NLU). For many of the languages covered in SIB-200, this is the first publicly\navailable evaluation dataset for NLU. The dataset is based on Flores-200\nmachine translation corpus. We annotated the English portion of the dataset and\nextended the sentence-level annotation to the remaining 203 languages covered\nin the corpus. Despite the simplicity of this task, our evaluation in\nfull-supervised setting, cross-lingual transfer setting and prompting of large\nlanguage model setting show that there is still a large gap between the\nperformance of high-resource and low-resource languages when multilingual\nevaluation is scaled to numerous world languages. We found that languages\nunseen during the pre-training of multilingual language models,\nunder-represented language families (like Nilotic and Altantic-Congo), and\nlanguages from the regions of Africa, Americas, Oceania and South East Asia,\noften have the lowest performance on our topic classification dataset. We hope\nour dataset will encourage a more inclusive evaluation of multilingual language\nmodels on a more diverse set of languages. https://github.com/dadelani/sib-200", "title": "SIB-200: A Simple, Inclusive, and Big Evaluation Dataset for Topic Classification in 200+ Languages and Dialects", "url": "http://arxiv.org/abs/2309.07445v1" }

{ "text_length": 1740 }

{ "abstract": "From the last decade, researchers in the field of machine learning (ML) and\nassistive developmental robotics (ADR) have taken an interest in artificial\nempathy (AE) as a possible future paradigm for human-robot interaction (HRI).\nHumans learn empathy since birth, therefore, it is challenging to instill this\nsense in robots and intelligent machines. Nevertheless, by training over a vast\namount of data and time, imitating empathy, to a certain extent, can be\npossible for robots. Training techniques for AE, along with findings from the\nfield of empathetic AI research, are ever-evolving. The standard workflow for\nartificial empathy consists of three stages: 1) Emotion Recognition (ER) using\nthe retrieved features from video or textual data, 2) analyzing the perceived\nemotion or degree of empathy to choose the best course of action, and 3)\ncarrying out a response action. Recent studies that show AE being used with\nvirtual agents or robots often include Deep Learning (DL) techniques. For\ninstance, models like VGGFace are used to conduct ER. Semi-supervised models\nlike Autoencoders generate the corresponding emotional states and behavioral\nresponses. However, there has not been any study that presents an independent\napproach for evaluating AE, or the degree to which a reaction was empathetic.\nThis paper aims to investigate and evaluate existing works for measuring and\nevaluating empathy, as well as the datasets that have been collected and used\nso far. Our goal is to highlight and facilitate the use of state-of-the-art\nmethods in the area of AE by comparing their performance. This will aid\nresearchers in the area of AE in selecting their approaches with precision.", "title": "Artificial Empathy Classification: A Survey of Deep Learning Techniques, Datasets, and Evaluation Scales", "url": "http://arxiv.org/abs/2310.00010v1" }

{ "text_length": 1824 }

{ "abstract": "Defect prediction has been a popular research topic where machine learning\n(ML) and deep learning (DL) have found numerous applications. However, these\nML/DL-based defect prediction models are often limited by the quality and size\nof their datasets. In this paper, we present Defectors, a large dataset for\njust-in-time and line-level defect prediction. Defectors consists of $\\approx$\n213K source code files ($\\approx$ 93K defective and $\\approx$ 120K defect-free)\nthat span across 24 popular Python projects. These projects come from 18\ndifferent domains, including machine learning, automation, and\ninternet-of-things. Such a scale and diversity make Defectors a suitable\ndataset for training ML/DL models, especially transformer models that require\nlarge and diverse datasets. We also foresee several application areas of our\ndataset including defect prediction and defect explanation.\n Dataset link: https://doi.org/10.5281/zenodo.7708984", "title": "Defectors: A Large, Diverse Python Dataset for Defect Prediction", "url": "http://arxiv.org/abs/2303.04738v4" }

{ "text_length": 1043 }

{ "abstract": "Understanding how events are semantically related to each other is the\nessence of reading comprehension. Recent event-centric reading comprehension\ndatasets focus mostly on event arguments or temporal relations. While these\ntasks partially evaluate machines' ability of narrative understanding,\nhuman-like reading comprehension requires the capability to process event-based\ninformation beyond arguments and temporal reasoning. For example, to understand\ncausality between events, we need to infer motivation or purpose; to establish\nevent hierarchy, we need to understand the composition of events. To facilitate\nthese tasks, we introduce ESTER, a comprehensive machine reading comprehension\n(MRC) dataset for Event Semantic Relation Reasoning. The dataset leverages\nnatural language queries to reason about the five most common event semantic\nrelations, provides more than 6K questions and captures 10.1K event relation\npairs. Experimental results show that the current SOTA systems achieve 22.1%,\n63.3%, and 83.5% for token-based exact-match, F1, and event-based HIT@1 scores,\nwhich are all significantly below human performances (36.0%, 79.6%, 100%\nrespectively), highlighting our dataset as a challenging benchmark.", "title": "ESTER: A Machine Reading Comprehension Dataset for Event Semantic Relation Reasoning", "url": "http://arxiv.org/abs/2104.08350v2" }

{ "text_length": 1339 }

{ "abstract": "An increasing number of reports raise concerns about the risk that machine\nlearning algorithms could amplify health disparities due to biases embedded in\nthe training data. Seyyed-Kalantari et al. find that models trained on three\nchest X-ray datasets yield disparities in false-positive rates (FPR) across\nsubgroups on the 'no-finding' label (indicating the absence of disease). The\nmodels consistently yield higher FPR on subgroups known to be historically\nunderserved, and the study concludes that the models exhibit and potentially\neven amplify systematic underdiagnosis. We argue that the experimental setup in\nthe study is insufficient to study algorithmic underdiagnosis. In the absence\nof specific knowledge (or assumptions) about the extent and nature of the\ndataset bias, it is difficult to investigate model bias. Importantly, their use\nof test data exhibiting the same bias as the training data (due to random\nsplitting) severely complicates the interpretation of the reported disparities.", "title": "Potential sources of dataset bias complicate investigation of underdiagnosis by machine learning algorithms", "url": "http://arxiv.org/abs/2201.07856v2" }

{ "text_length": 1143 }

{ "abstract": "We introduce TTSWING, a novel dataset designed for table tennis swing\nanalysis. This dataset comprises comprehensive swing information obtained\nthrough 9-axis sensors integrated into custom-made racket grips, accompanied by\nanonymized demographic data of the players. We detail the data collection and\nannotation procedures. Furthermore, we conduct pilot studies utilizing diverse\nmachine learning models for swing analysis. TTSWING holds tremendous potential\nto facilitate innovative research in table tennis analysis and is a valuable\nresource for the scientific community. We release the dataset and experimental\ncodes at https://github.com/DEPhantom/TTSWING.", "title": "TTSWING: a Dataset for Table Tennis Swing Analysis", "url": "http://arxiv.org/abs/2306.17550v1" }

{ "text_length": 747 }

{ "abstract": "The Second Diagnosis of COVID-19 using Acoustics (DiCOVA) Challenge aimed at\naccelerating the research in acoustics based detection of COVID-19, a topic at\nthe intersection of acoustics, signal processing, machine learning, and\nhealthcare. This paper presents the details of the challenge, which was an open\ncall for researchers to analyze a dataset of audio recordings consisting of\nbreathing, cough and speech signals. This data was collected from individuals\nwith and without COVID-19 infection, and the task in the challenge was a\ntwo-class classification. The development set audio recordings were collected\nfrom 965 (172 COVID-19 positive) individuals, while the evaluation set\ncontained data from 471 individuals (71 COVID-19 positive). The challenge\nfeatured four tracks, one associated with each sound category of cough, speech\nand breathing, and a fourth fusion track. A baseline system was also released\nto benchmark the participants. In this paper, we present an overview of the\nchallenge, the rationale for the data collection and the baseline system.\nFurther, a performance analysis for the systems submitted by the $16$\nparticipating teams in the leaderboard is also presented.", "title": "The Second DiCOVA Challenge: Dataset and performance analysis for COVID-19 diagnosis using acoustics", "url": "http://arxiv.org/abs/2110.01177v3" }

{ "text_length": 1327 }

{ "abstract": "The DiCOVA challenge aims at accelerating research in diagnosing COVID-19\nusing acoustics (DiCOVA), a topic at the intersection of speech and audio\nprocessing, respiratory health diagnosis, and machine learning. This challenge\nis an open call for researchers to analyze a dataset of sound recordings\ncollected from COVID-19 infected and non-COVID-19 individuals for a two-class\nclassification. These recordings were collected via crowdsourcing from multiple\ncountries, through a website application. The challenge features two tracks,\none focusing on cough sounds, and the other on using a collection of breath,\nsustained vowel phonation, and number counting speech recordings. In this\npaper, we introduce the challenge and provide a detailed description of the\ntask, and present a baseline system for the task.", "title": "DiCOVA Challenge: Dataset, task, and baseline system for COVID-19 diagnosis using acoustics", "url": "http://arxiv.org/abs/2103.09148v3" }

{ "text_length": 937 }

{ "abstract": "This article presents GrowliFlower, a georeferenced, image-based UAV time\nseries dataset of two monitored cauliflower fields of size 0.39 and 0.60 ha\nacquired in 2020 and 2021. The dataset contains RGB and multispectral\northophotos from which about 14,000 individual plant coordinates are derived\nand provided. The coordinates enable the dataset users the extraction of\ncomplete and incomplete time series of image patches showing individual plants.\nThe dataset contains collected phenotypic traits of 740 plants, including the\ndevelopmental stage as well as plant and cauliflower size. As the harvestable\nproduct is completely covered by leaves, plant IDs and coordinates are provided\nto extract image pairs of plants pre and post defoliation, to facilitate\nestimations of cauliflower head size. Moreover, the dataset contains\npixel-accurate leaf and plant instance segmentations, as well as stem\nannotations to address tasks like classification, detection, segmentation,\ninstance segmentation, and similar computer vision tasks. The dataset aims to\nfoster the development and evaluation of machine learning approaches. It\nspecifically focuses on the analysis of growth and development of cauliflower\nand the derivation of phenotypic traits to foster the development of automation\nin agriculture. Two baseline results of instance segmentation at plant and leaf\nlevel based on the labeled instance segmentation data are presented. The entire\ndata set is publicly available.", "title": "GrowliFlower: An image time series dataset for GROWth analysis of cauLIFLOWER", "url": "http://arxiv.org/abs/2204.00294v1" }

{ "text_length": 1585 }

{ "abstract": "The article presents a new multi-label comprehensive image dataset from\nflexible endoscopy, colonoscopy and capsule endoscopy, named ERS. The\ncollection has been labeled according to the full medical specification of\n'Minimum Standard Terminology 3.0' (MST 3.0), describing all possible findings\nin the gastrointestinal tract (104 possible labels), extended with an\nadditional 19 labels useful in common machine learning applications.\n The dataset contains around 6000 precisely and 115,000 approximately labeled\nframes from endoscopy videos, 3600 precise and 22,600 approximate segmentation\nmasks, and 1.23 million unlabeled frames from flexible and capsule endoscopy\nvideos. The labeled data cover almost entirely the MST 3.0 standard. The data\ncame from 1520 videos of 1135 patients.\n Additionally, this paper proposes and describes four exemplary experiments in\ngastrointestinal image classification task performed using the created dataset.\nThe obtained results indicate the high usefulness and flexibility of the\ndataset in training and testing machine learning algorithms in the field of\nendoscopic data analysis.", "title": "ERS: a novel comprehensive endoscopy image dataset for machine learning, compliant with the MST 3.0 specification", "url": "http://arxiv.org/abs/2201.08746v1" }

{ "text_length": 1270 }

{ "abstract": "Scarcity of labeled histopathology data limits the applicability of deep\nlearning methods to under-profiled cancer types and labels. Transfer learning\nallows researchers to overcome the limitations of small datasets by\npre-training machine learning models on larger datasets similar to the small\ntarget dataset. However, similarity between datasets is often determined\nheuristically. In this paper, we propose a principled notion of distance\nbetween histopathology datasets based on a hierarchical generalization of\noptimal transport distances. Our method does not require any training, is\nagnostic to model type, and preserves much of the hierarchical structure in\nhistopathology datasets imposed by tiling. We apply our method to H&E stained\nslides from The Cancer Genome Atlas from six different cancer types. We show\nthat our method outperforms a baseline distance in a cancer-type prediction\ntask. Our results also show that our optimal transport distance predicts\ndifficulty of transferability in a tumor vs.normal prediction setting.", "title": "Hierarchical Optimal Transport for Comparing Histopathology Datasets", "url": "http://arxiv.org/abs/2204.08324v2" }

{ "text_length": 1143 }

{ "abstract": "The application of deep learning to non-stationary temporal datasets can lead\nto overfitted models that underperform under regime changes. In this work, we\npropose a modular machine learning pipeline for ranking predictions on temporal\npanel datasets which is robust under regime changes. The modularity of the\npipeline allows the use of different models, including Gradient Boosting\nDecision Trees (GBDTs) and Neural Networks, with and without feature\nengineering. We evaluate our framework on financial data for stock portfolio\nprediction, and find that GBDT models with dropout display high performance,\nrobustness and generalisability with reduced complexity and computational cost.\nWe then demonstrate how online learning techniques, which require no retraining\nof models, can be used post-prediction to enhance the results. First, we show\nthat dynamic feature projection improves robustness by reducing drawdown in\nregime changes. Second, we demonstrate that dynamical model ensembling based on\nselection of models with good recent performance leads to improved Sharpe and\nCalmar ratios of out-of-sample predictions. We also evaluate the robustness of\nour pipeline across different data splits and random seeds with good\nreproducibility.", "title": "Online learning techniques for prediction of temporal tabular datasets with regime changes", "url": "http://arxiv.org/abs/2301.00790v4" }

{ "text_length": 1368 }

{ "abstract": "Vision-and-language tasks have increasingly drawn more attention as a means\nto evaluate human-like reasoning in machine learning models. A popular task in\nthe field is visual question answering (VQA), which aims to answer questions\nabout images. However, VQA models have been shown to exploit language bias by\nlearning the statistical correlations between questions and answers without\nlooking into the image content: e.g., questions about the color of a banana are\nanswered with yellow, even if the banana in the image is green. If societal\nbias (e.g., sexism, racism, ableism, etc.) is present in the training data,\nthis problem may be causing VQA models to learn harmful stereotypes. For this\nreason, we investigate gender and racial bias in five VQA datasets. In our\nanalysis, we find that the distribution of answers is highly different between\nquestions about women and men, as well as the existence of detrimental\ngender-stereotypical samples. Likewise, we identify that specific race-related\nattributes are underrepresented, whereas potentially discriminatory samples\nappear in the analyzed datasets. Our findings suggest that there are dangers\nassociated to using VQA datasets without considering and dealing with the\npotentially harmful stereotypes. We conclude the paper by proposing solutions\nto alleviate the problem before, during, and after the dataset collection\nprocess.", "title": "Gender and Racial Bias in Visual Question Answering Datasets", "url": "http://arxiv.org/abs/2205.08148v3" }

{ "text_length": 1482 }

{ "abstract": "Classifying samples in incomplete datasets is a common aim for machine\nlearning practitioners, but is non-trivial. Missing data is found in most\nreal-world datasets and these missing values are typically imputed using\nestablished methods, followed by classification of the now complete, imputed,\nsamples. The focus of the machine learning researcher is then to optimise the\ndownstream classification performance. In this study, we highlight that it is\nimperative to consider the quality of the imputation. We demonstrate how the\ncommonly used measures for assessing quality are flawed and propose a new class\nof discrepancy scores which focus on how well the method recreates the overall\ndistribution of the data. To conclude, we highlight the compromised\ninterpretability of classifier models trained using poorly imputed data.", "title": "Classification of datasets with imputed missing values: does imputation quality matter?", "url": "http://arxiv.org/abs/2206.08478v1" }

{ "text_length": 950 }

{ "abstract": "Tactile sensing is an essential modality for smart robots as it enables them\nto interact flexibly with physical objects in their environment. Recent\nadvancements in electronic skins have led to the development of data-driven\nmachine learning methods that exploit this important sensory modality. However,\ncurrent datasets used to train such algorithms are limited to standard\nsynchronous tactile sensors. There is a dearth of neuromorphic event-based\ntactile datasets, principally due to the scarcity of large-scale event-based\ntactile sensors. Having such datasets is crucial for the development and\nevaluation of new algorithms that process spatio-temporal event-based data. For\nexample, evaluating spiking neural networks on conventional frame-based\ndatasets is considered sub-optimal. Here, we debut a novel neuromorphic Spiking\nTactile MNIST (ST-MNIST) dataset, which comprises handwritten digits obtained\nby human participants writing on a neuromorphic tactile sensor array. We also\ndescribe an initial effort to evaluate our ST-MNIST dataset using existing\nartificial and spiking neural network models. The classification accuracies\nprovided herein can serve as performance benchmarks for future work. We\nanticipate that our ST-MNIST dataset will be of interest and useful to the\nneuromorphic and robotics research communities.", "title": "ST-MNIST -- The Spiking Tactile MNIST Neuromorphic Dataset", "url": "http://arxiv.org/abs/2005.04319v1" }

{ "text_length": 1427 }

{ "abstract": "Task-oriented dialogue research has mainly focused on a few popular languages\nlike English and Chinese, due to the high dataset creation cost for a new\nlanguage. To reduce the cost, we apply manual editing to automatically\ntranslated data. We create a new multilingual benchmark, X-RiSAWOZ, by\ntranslating the Chinese RiSAWOZ to 4 languages: English, French, Hindi, Korean;\nand a code-mixed English-Hindi language. X-RiSAWOZ has more than 18,000\nhuman-verified dialogue utterances for each language, and unlike most\nmultilingual prior work, is an end-to-end dataset for building\nfully-functioning agents.\n The many difficulties we encountered in creating X-RiSAWOZ led us to develop\na toolset to accelerate the post-editing of a new language dataset after\ntranslation. This toolset improves machine translation with a hybrid entity\nalignment technique that combines neural with dictionary-based methods, along\nwith many automated and semi-automated validation checks.\n We establish strong baselines for X-RiSAWOZ by training dialogue agents in\nthe zero- and few-shot settings where limited gold data is available in the\ntarget language. Our results suggest that our translation and post-editing\nmethodology and toolset can be used to create new high-quality multilingual\ndialogue agents cost-effectively. Our dataset, code, and toolkit are released\nopen-source.", "title": "X-RiSAWOZ: High-Quality End-to-End Multilingual Dialogue Datasets and Few-shot Agents", "url": "http://arxiv.org/abs/2306.17674v1" }

{ "text_length": 1483 }

{ "abstract": "Climate change is global, yet its concrete impacts can strongly vary between\ndifferent locations in the same region. Seasonal weather forecasts currently\noperate at the mesoscale (> 1 km). For more targeted mitigation and adaptation,\nmodelling impacts to < 100 m is needed. Yet, the relationship between driving\nvariables and Earth's surface at such local scales remains unresolved by\ncurrent physical models. Large Earth observation datasets now enable us to\ncreate machine learning models capable of translating coarse weather\ninformation into high-resolution Earth surface forecasts. Here, we define\nhigh-resolution Earth surface forecasting as video prediction of satellite\nimagery conditional on mesoscale weather forecasts. Video prediction has been\ntackled with deep learning models. Developing such models requires\nanalysis-ready datasets. We introduce EarthNet2021, a new, curated dataset\ncontaining target spatio-temporal Sentinel 2 satellite imagery at 20 m\nresolution, matched with high-resolution topography and mesoscale (1.28 km)\nweather variables. With over 32000 samples it is suitable for training deep\nneural networks. Comparing multiple Earth surface forecasts is not trivial.\nHence, we define the EarthNetScore, a novel ranking criterion for models\nforecasting Earth surface reflectance. For model intercomparison we frame\nEarthNet2021 as a challenge with four tracks based on different test sets.\nThese allow evaluation of model validity and robustness as well as model\napplicability to extreme events and the complete annual vegetation cycle. In\naddition to forecasting directly observable weather impacts through\nsatellite-derived vegetation indices, capable Earth surface models will enable\ndownstream applications such as crop yield prediction, forest health\nassessments, coastline management, or biodiversity monitoring. Find data, code,\nand how to participate at www.earthnet.tech .", "title": "EarthNet2021: A novel large-scale dataset and challenge for forecasting localized climate impacts", "url": "http://arxiv.org/abs/2012.06246v1" }

{ "text_length": 2042 }

{ "abstract": "Current research on hate speech analysis is typically oriented towards\nmonolingual and single classification tasks. In this paper, we present a new\nmultilingual hate speech analysis dataset for English, Hindi, Arabic, French,\nGerman and Spanish languages for multiple domains across hate speech - Abuse,\nRacism, Sexism, Religious Hate and Extremism. To the best of our knowledge,\nthis paper is the first to address the problem of identifying various types of\nhate speech in these five wide domains in these six languages. In this work, we\ndescribe how we created the dataset, created annotations at high level and low\nlevel for different domains and how we use it to test the current\nstate-of-the-art multilingual and multitask learning approaches. We evaluate\nour dataset in various monolingual, cross-lingual and machine translation\nclassification settings and compare it against open source English datasets\nthat we aggregated and merged for this task. Then we discuss how this approach\ncan be used to create large scale hate-speech datasets and how to leverage our\nannotations in order to improve hate speech detection and classification in\ngeneral.", "title": "LAHM : Large Annotated Dataset for Multi-Domain and Multilingual Hate Speech Identification", "url": "http://arxiv.org/abs/2304.00913v1" }

{ "text_length": 1279 }

{ "abstract": "Histopathology is critical for the diagnosis of many diseases, including\ncancer. These protocols typically require pathologists to manually evaluate\nslides under a microscope, which is time-consuming and subjective, leading to\ninterest in machine learning to automate analysis. However, computational\ntechniques are limited by batch effects, where technical factors like\ndifferences in preparation protocol or scanners can alter the appearance of\nslides, causing models trained on one institution to fail when generalizing to\nothers. Here, we propose a domain adaptation method that improves the\ngeneralization of histopathological models to data from unseen institutions,\nwithout the need for labels or retraining in these new settings. Our approach\nintroduces an optimal transport (OT) loss, that extends adversarial methods\nthat penalize models if images from different institutions can be distinguished\nin their representation space. Unlike previous methods, which operate on single\nsamples, our loss accounts for distributional differences between batches of\nimages. We show that on the Camelyon17 dataset, while both methods can adapt to\nglobal differences in color distribution, only our OT loss can reliably\nclassify a cancer phenotype unseen during training. Together, our results\nsuggest that OT improves generalization on rare but critical phenotypes that\nmay only make up a small fraction of the total tiles and variation in a slide.", "title": "Domain adaptation using optimal transport for invariant learning using histopathology datasets", "url": "http://arxiv.org/abs/2303.02241v1" }

{ "text_length": 1574 }

{ "abstract": "The approach used not only challenges some of the fundamental mathematical\ntechniques used so far in early experiments of the same trend but also\nintroduces new scopes and new horizons for interesting results. The physics\ngoverning spectrograms have been optimized in the project along with exploring\nhow it handles the intense requirements of the problem at hand. Major\ncontributions and developments brought under the light, through this project\ninvolve using better mathematical techniques and problem-specific machine\nlearning methods. Improvised data analysis and data augmentation for audio\ndatasets like frequency masking and random frequency-time stretching are used\nin the project and hence are explained in this paper. In the used methodology,\nthe audio transforms principle were also tried and explored, and indeed the\ninsights gained were used constructively in the later stages of the project.\nUsing a deep learning principle is surely one of them. Also, in this paper, the\npotential scopes and upcoming research openings in both short and long term\ntunnel of time has been presented. Although much of the results gained are\ndomain-specific as of now, they are surely potent enough to produce novel\nsolutions in various different domains of diverse backgrounds.", "title": "Acoustic scene classification using auditory datasets", "url": "http://arxiv.org/abs/2112.13450v2" }

{ "text_length": 1362 }

{ "abstract": "We present multiplexed gradient descent (MGD), a gradient descent framework\ndesigned to easily train analog or digital neural networks in hardware. MGD\nutilizes zero-order optimization techniques for online training of hardware\nneural networks. We demonstrate its ability to train neural networks on modern\nmachine learning datasets, including CIFAR-10 and Fashion-MNIST, and compare\nits performance to backpropagation. Assuming realistic timescales and hardware\nparameters, our results indicate that these optimization techniques can train a\nnetwork on emerging hardware platforms orders of magnitude faster than the\nwall-clock time of training via backpropagation on a standard GPU, even in the\npresence of imperfect weight updates or device-to-device variations in the\nhardware. We additionally describe how it can be applied to existing hardware\nas part of chip-in-the-loop training, or integrated directly at the hardware\nlevel. Crucially, the MGD framework is highly flexible, and its gradient\ndescent process can be optimized to compensate for specific hardware\nlimitations such as slow parameter-update speeds or limited input bandwidth.", "title": "Multiplexed gradient descent: Fast online training of modern datasets on hardware neural networks without backpropagation", "url": "http://arxiv.org/abs/2303.03986v1" }

{ "text_length": 1301 }

{ "abstract": "Over 97 million people speak Vietnamese as their native language in the\nworld. However, there are few research studies on machine reading comprehension\n(MRC) for Vietnamese, the task of understanding a text and answering questions\nrelated to it. Due to the lack of benchmark datasets for Vietnamese, we present\nthe Vietnamese Question Answering Dataset (UIT-ViQuAD), a new dataset for the\nlow-resource language as Vietnamese to evaluate MRC models. This dataset\ncomprises over 23,000 human-generated question-answer pairs based on 5,109\npassages of 174 Vietnamese articles from Wikipedia. In particular, we propose a\nnew process of dataset creation for Vietnamese MRC. Our in-depth analyses\nillustrate that our dataset requires abilities beyond simple reasoning like\nword matching and demands single-sentence and multiple-sentence inferences.\nBesides, we conduct experiments on state-of-the-art MRC methods for English and\nChinese as the first experimental models on UIT-ViQuAD. We also estimate human\nperformance on the dataset and compare it to the experimental results of\npowerful machine learning models. As a result, the substantial differences\nbetween human performance and the best model performance on the dataset\nindicate that improvements can be made on UIT-ViQuAD in future research. Our\ndataset is freely available on our website to encourage the research community\nto overcome challenges in Vietnamese MRC.", "title": "A Vietnamese Dataset for Evaluating Machine Reading Comprehension", "url": "http://arxiv.org/abs/2009.14725v3" }

{ "text_length": 1519 }

{ "abstract": "A smart home energy dataset that records miscellaneous energy consumption\ndata is publicly offered. The proposed energy activity dataset (EAD) has a high\ndata type diversity in contrast to existing load monitoring datasets. In EAD, a\nsimple data point is labeled with the appliance, brand, and event information,\nwhereas a complex data point has an extra application label. Several\ndiscoveries have been made on the energy consumption patterns of many\nappliances. Load curves of the appliances are measured when different events\nand applications are triggered and utilized. A revised\nlongest-common-subsequence (LCS) similarity measurement algorithm is proposed\nto calculate energy dataset similarities. Thus, the data quality prior\ninformation becomes available before training machine learning models. In\naddition, a subsample convolutional neural network (SCNN) is put forward. It\nserves as a non-intrusive optical character recognition (OCR) approach to\nobtain energy data directly from monitors of power meters. The link for the EAD\ndataset is:\nhttps://drive.google.com/drive/folders/1zn0V6Q8eXXSKxKgcs8ZRValL5VEn3anD", "title": "An Energy Activity Dataset for Smart Homes", "url": "http://arxiv.org/abs/2208.13416v2" }

{ "text_length": 1199 }

{ "abstract": "Lifelong learning (LL) aims to improve a predictive model as the data source\nevolves continuously. Most work in this learning paradigm has focused on\nresolving the problem of 'catastrophic forgetting,' which refers to a notorious\ndilemma between improving model accuracy over new data and retaining accuracy\nover previous data. Yet, it is also known that machine learning (ML) models can\nbe vulnerable in the sense that tiny, adversarial input perturbations can\ndeceive the models into producing erroneous predictions. This motivates the\nresearch objective of this paper - specification of a new LL framework that can\nsalvage model robustness (against adversarial attacks) from catastrophic\nforgetting. Specifically, we propose a new memory-replay LL strategy that\nleverages modern bi-level optimization techniques to determine the 'coreset' of\nthe current data (i.e., a small amount of data to be memorized) for ease of\npreserving adversarial robustness over time. We term the resulting LL framework\n'Data-Efficient Robustness-Preserving LL' (DERPLL). The effectiveness of DERPLL\nis evaluated for class-incremental image classification using ResNet-18 over\nthe CIFAR-10 dataset. Experimental results show that DERPLL outperforms the\nconventional coreset-guided LL baseline and achieves a substantial improvement\nin both standard accuracy and robust accuracy.", "title": "Robustness-preserving Lifelong Learning via Dataset Condensation", "url": "http://arxiv.org/abs/2303.04183v1" }

{ "text_length": 1458 }

{ "abstract": "The MS MARCO ranking dataset has been widely used for training deep learning\nmodels for IR tasks, achieving considerable effectiveness on diverse zero-shot\nscenarios. However, this type of resource is scarce in languages other than\nEnglish. In this work, we present mMARCO, a multilingual version of the MS\nMARCO passage ranking dataset comprising 13 languages that was created using\nmachine translation. We evaluated mMARCO by finetuning monolingual and\nmultilingual reranking models, as well as a multilingual dense retrieval model\non this dataset. We also evaluated models finetuned using the mMARCO dataset in\na zero-shot scenario on Mr. TyDi dataset, demonstrating that multilingual\nmodels finetuned on our translated dataset achieve superior effectiveness to\nmodels finetuned on the original English version alone. Our experiments also\nshow that a distilled multilingual reranker is competitive with non-distilled\nmodels while having 5.4 times fewer parameters. Lastly, we show a positive\ncorrelation between translation quality and retrieval effectiveness, providing\nevidence that improvements in translation methods might lead to improvements in\nmultilingual information retrieval. The translated datasets and finetuned\nmodels are available at https://github.com/unicamp-dl/mMARCO.", "title": "mMARCO: A Multilingual Version of the MS MARCO Passage Ranking Dataset", "url": "http://arxiv.org/abs/2108.13897v5" }

{ "text_length": 1394 }

{ "abstract": "This paper considers the problem of enhancing user privacy in common machine\nlearning development tasks, such as data annotation and inspection, by\nsubstituting the real data with samples form a generative adversarial network.\nWe propose employing Bayesian differential privacy as the means to achieve a\nrigorous theoretical guarantee while providing a better privacy-utility\ntrade-off. We demonstrate experimentally that our approach produces\nhigher-fidelity samples, compared to prior work, allowing to (1) detect more\nsubtle data errors and biases, and (2) reduce the need for real data labelling\nby achieving high accuracy when training directly on artificial samples.", "title": "Generating Higher-Fidelity Synthetic Datasets with Privacy Guarantees", "url": "http://arxiv.org/abs/2003.00997v1" }

{ "text_length": 776 }

{ "abstract": "Current state of the art acoustic models can easily comprise more than 100\nmillion parameters. This growing complexity demands larger training datasets to\nmaintain a decent generalization of the final decision function. An ideal\ndataset is not necessarily large in size, but large with respect to the amount\nof unique speakers, utilized hardware and varying recording conditions. This\nenables a machine learning model to explore as much of the domain-specific\ninput space as possible during parameter estimation. This work introduces\nCommon Phone, a gender-balanced, multilingual corpus recorded from more than\n11.000 contributors via Mozilla's Common Voice project. It comprises around 116\nhours of speech enriched with automatically generated phonetic segmentation. A\nWav2Vec 2.0 acoustic model was trained with the Common Phone to perform\nphonetic symbol recognition and validate the quality of the generated phonetic\nannotation. The architecture achieved a PER of 18.1 % on the entire test set,\ncomputed with all 101 unique phonetic symbols, showing slight differences\nbetween the individual languages. We conclude that Common Phone provides\nsufficient variability and reliable phonetic annotation to help bridging the\ngap between research and application of acoustic models.", "title": "Common Phone: A Multilingual Dataset for Robust Acoustic Modelling", "url": "http://arxiv.org/abs/2201.05912v2" }

{ "text_length": 1380 }

{ "abstract": "Multimodal large-scale datasets for outdoor scenes are mostly designed for\nurban driving problems. The scenes are highly structured and semantically\ndifferent from scenarios seen in nature-centered scenes such as gardens or\nparks. To promote machine learning methods for nature-oriented applications,\nsuch as agriculture and gardening, we propose the multimodal synthetic dataset\nfor Enclosed garDEN scenes (EDEN). The dataset features more than 300K images\ncaptured from more than 100 garden models. Each image is annotated with various\nlow/high-level vision modalities, including semantic segmentation, depth,\nsurface normals, intrinsic colors, and optical flow. Experimental results on\nthe state-of-the-art methods for semantic segmentation and monocular depth\nprediction, two important tasks in computer vision, show positive impact of\npre-training deep networks on our dataset for unstructured natural scenes. The\ndataset and related materials will be available at\nhttps://lhoangan.github.io/eden.", "title": "EDEN: Multimodal Synthetic Dataset of Enclosed GarDEN Scenes", "url": "http://arxiv.org/abs/2011.04389v2" }

{ "text_length": 1097 }

{ "abstract": "Machine learning potentials are an important tool for molecular simulation,\nbut their development is held back by a shortage of high quality datasets to\ntrain them on. We describe the SPICE dataset, a new quantum chemistry dataset\nfor training potentials relevant to simulating drug-like small molecules\ninteracting with proteins. It contains over 1.1 million conformations for a\ndiverse set of small molecules, dimers, dipeptides, and solvated amino acids.\nIt includes 15 elements, charged and uncharged molecules, and a wide range of\ncovalent and non-covalent interactions. It provides both forces and energies\ncalculated at the {\\omega}B97M-D3(BJ)/def2-TZVPPD level of theory, along with\nother useful quantities such as multipole moments and bond orders. We train a\nset of machine learning potentials on it and demonstrate that they can achieve\nchemical accuracy across a broad region of chemical space. It can serve as a\nvaluable resource for the creation of transferable, ready to use potential\nfunctions for use in molecular simulations.", "title": "SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials", "url": "http://arxiv.org/abs/2209.10702v2" }

{ "text_length": 1171 }

{ "abstract": "Entity linking (EL) is the task of linking a textual mention to its\ncorresponding entry in a knowledge base, and is critical for many\nknowledge-intensive NLP applications. When applied to tables in scientific\npapers, EL is a step toward large-scale scientific knowledge bases that could\nenable advanced scientific question answering and analytics. We present the\nfirst dataset for EL in scientific tables. EL for scientific tables is\nespecially challenging because scientific knowledge bases can be very\nincomplete, and disambiguating table mentions typically requires understanding\nthe papers's tet in addition to the table. Our dataset, S2abEL, focuses on EL\nin machine learning results tables and includes hand-labeled cell types,\nattributed sources, and entity links from the PaperswithCode taxonomy for 8,429\ncells from 732 tables. We introduce a neural baseline method designed for EL on\nscientific tables containing many out-of-knowledge-base mentions, and show that\nit significantly outperforms a state-of-the-art generic table EL method. The\nbest baselines fall below human performance, and our analysis highlights\navenues for improvement.", "title": "S2abEL: A Dataset for Entity Linking from Scientific Tables", "url": "http://arxiv.org/abs/2305.00366v1" }

{ "text_length": 1242 }

{ "abstract": "Forest biomass is a key influence for future climate, and the world urgently\nneeds highly scalable financing schemes, such as carbon offsetting\ncertifications, to protect and restore forests. Current manual forest carbon\nstock inventory methods of measuring single trees by hand are time, labour, and\ncost-intensive and have been shown to be subjective. They can lead to\nsubstantial overestimation of the carbon stock and ultimately distrust in\nforest financing. The potential for impact and scale of leveraging advancements\nin machine learning and remote sensing technologies is promising but needs to\nbe of high quality in order to replace the current forest stock protocols for\ncertifications.\n In this paper, we present ReforesTree, a benchmark dataset of forest carbon\nstock in six agro-forestry carbon offsetting sites in Ecuador. Furthermore, we\nshow that a deep learning-based end-to-end model using individual tree\ndetection from low cost RGB-only drone imagery is accurately estimating forest\ncarbon stock within official carbon offsetting certification standards.\nAdditionally, our baseline CNN model outperforms state-of-the-art\nsatellite-based forest biomass and carbon stock estimates for this type of\nsmall-scale, tropical agro-forestry sites. We present this dataset to encourage\nmachine learning research in this area to increase accountability and\ntransparency of monitoring, verification and reporting (MVR) in carbon\noffsetting projects, as well as scaling global reforestation financing through\naccurate remote sensing.", "title": "ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep Learning and Aerial Imagery", "url": "http://arxiv.org/abs/2201.11192v1" }

{ "text_length": 1680 }

{ "abstract": "Healthcare is an important aspect of human life. Use of technologies in\nhealthcare has increased manifolds after the pandemic. Internet of Things based\nsystems and devices proposed in literature can help elders, children and adults\nfacing/experiencing health problems. This paper exhaustively reviews\nthirty-nine wearable based datasets which can be used for evaluating the system\nto recognize Activities of Daily Living and Falls. A comparative analysis on\nthe SisFall dataset using five machine learning methods i.e., Logistic\nRegression, Linear Discriminant Analysis, K-Nearest Neighbor, Decision Tree and\nNaive Bayes is performed in python. The dataset is modified in two ways, in\nfirst all the attributes present in dataset are used as it is and labelled in\nbinary form. In second, magnitude of three axes(x,y,z) for three sensors value\nare computed and then used in experiment with label attribute. The experiments\nare performed on one subject, ten subjects and all the subjects and compared in\nterms of accuracy, precision and recall. The results obtained from this study\nproves that KNN outperforms other machine learning methods in terms of\naccuracy, precision and recall. It is also concluded that personalization of\ndata improves accuracy.", "title": "Analyzing Wearables Dataset to Predict ADLs and Falls: A Pilot Study", "url": "http://arxiv.org/abs/2209.04785v1" }

{ "text_length": 1353 }

{ "abstract": "Multi-modal Machine Translation (MMT) enables the use of visual information\nto enhance the quality of translations. The visual information can serve as a\nvaluable piece of context information to decrease the ambiguity of input\nsentences. Despite the increasing popularity of such a technique, good and\nsizeable datasets are scarce, limiting the full extent of their potential.\nHausa, a Chadic language, is a member of the Afro-Asiatic language family. It\nis estimated that about 100 to 150 million people speak the language, with more\nthan 80 million indigenous speakers. This is more than any of the other Chadic\nlanguages. Despite a large number of speakers, the Hausa language is considered\nlow-resource in natural language processing (NLP). This is due to the absence\nof sufficient resources to implement most NLP tasks. While some datasets exist,\nthey are either scarce, machine-generated, or in the religious domain.\nTherefore, there is a need to create training and evaluation data for\nimplementing machine learning tasks and bridging the research gap in the\nlanguage. This work presents the Hausa Visual Genome (HaVG), a dataset that\ncontains the description of an image or a section within the image in Hausa and\nits equivalent in English. To prepare the dataset, we started by translating\nthe English description of the images in the Hindi Visual Genome (HVG) into\nHausa automatically. Afterward, the synthetic Hausa data was carefully\npost-edited considering the respective images. The dataset comprises 32,923\nimages and their descriptions that are divided into training, development,\ntest, and challenge test set. The Hausa Visual Genome is the first dataset of\nits kind and can be used for Hausa-English machine translation, multi-modal\nresearch, and image description, among various other natural language\nprocessing and generation tasks.", "title": "Hausa Visual Genome: A Dataset for Multi-Modal English to Hausa Machine Translation", "url": "http://arxiv.org/abs/2205.01133v2" }

{ "text_length": 1971 }

{ "abstract": "After more than two years since the beginning of the COVID-19 pandemic, the\npressure of this crisis continues to devastate globally. The use of chest X-ray\n(CXR) imaging as a complementary screening strategy to RT-PCR testing is not\nonly prevailing but has greatly increased due to its routine clinical use for\nrespiratory complaints. Thus far, many visual perception models have been\nproposed for COVID-19 screening based on CXR imaging. Nevertheless, the\naccuracy and the generalization capacity of these models are very much\ndependent on the diversity and the size of the dataset they were trained on.\nMotivated by this, we introduce COVIDx CXR-3, a large-scale benchmark dataset\nof CXR images for supporting COVID-19 computer vision research. COVIDx CXR-3 is\ncomposed of 30,386 CXR images from a multinational cohort of 17,026 patients\nfrom at least 51 countries, making it, to the best of our knowledge, the most\nextensive, most diverse COVID-19 CXR dataset in open access form. Here, we\nprovide comprehensive details on the various aspects of the proposed dataset\nincluding patient demographics, imaging views, and infection types. The hope is\nthat COVIDx CXR-3 can assist scientists in advancing machine learning research\nagainst both the COVID-19 pandemic and related diseases.", "title": "COVIDx CXR-3: A Large-Scale, Open-Source Benchmark Dataset of Chest X-ray Images for Computer-Aided COVID-19 Diagnostics", "url": "http://arxiv.org/abs/2206.03671v3" }

{ "text_length": 1440 }

{ "abstract": "This study presents machine learning models that forecast and categorize lost\ncirculation severity preemptively using a large class imbalanced drilling\ndataset. We demonstrate reproducible core techniques involved in tackling a\nlarge drilling engineering challenge utilizing easily interpretable machine\nlearning approaches.\n We utilized a 65,000+ records data with class imbalance problem from Azadegan\noilfield formations in Iran. Eleven of the dataset's seventeen parameters are\nchosen to be used in the classification of five lost circulation events. To\ngenerate classification models, we used six basic machine learning algorithms\nand four ensemble learning methods. Linear Discriminant Analysis (LDA),\nLogistic Regression (LR), Support Vector Machines (SVM), Classification and\nRegression Trees (CART), k-Nearest Neighbors (KNN), and Gaussian Naive Bayes\n(GNB) are the six fundamental techniques. We also used bagging and boosting\nensemble learning techniques in the investigation of solutions for improved\npredicting performance. The performance of these algorithms is measured using\nfour metrics: accuracy, precision, recall, and F1-score. The F1-score weighted\nto represent the data imbalance is chosen as the preferred evaluation\ncriterion.\n The CART model was found to be the best in class for identifying drilling\nfluid circulation loss events with an average weighted F1-score of 0.9904 and\nstandard deviation of 0.0015. Upon application of ensemble learning techniques,\na Random Forest ensemble of decision trees showed the best predictive\nperformance. It identified and classified lost circulation events with a\nperfect weighted F1-score of 1.0. Using Permutation Feature Importance (PFI),\nthe measured depth was found to be the most influential factor in accurately\nrecognizing lost circulation events while drilling.", "title": "A Case Study on the Classification of Lost Circulation Events During Drilling using Machine Learning Techniques on an Imbalanced Large Dataset", "url": "http://arxiv.org/abs/2209.01607v2" }

{ "text_length": 2012 }

{ "abstract": "The FAIR Guiding Principles aim to improve the findability, accessibility,\ninteroperability, and reusability of digital content by making them both human\nand machine actionable. However, these principles have not yet been broadly\nadopted in the domain of machine learning-based program analyses and\noptimizations for High-Performance Computing (HPC). In this paper, we design a\nmethodology to make HPC datasets and machine learning models FAIR after\ninvestigating existing FAIRness assessment and improvement techniques. Our\nmethodology includes a comprehensive, quantitative assessment for elected data,\nfollowed by concrete, actionable suggestions to improve FAIRness with respect\nto common issues related to persistent identifiers, rich metadata descriptions,\nlicense and provenance information. Moreover, we select a representative\ntraining dataset to evaluate our methodology. The experiment shows the\nmethodology can effectively improve the dataset and model's FAIRness from an\ninitial score of 19.1% to the final score of 83.0%.", "title": "Making Machine Learning Datasets and Models FAIR for HPC: A Methodology and Case Study", "url": "http://arxiv.org/abs/2211.02092v1" }

{ "text_length": 1156 }

{ "abstract": "Language is a method by which individuals express their thoughts. Each\nlanguage has its own set of alphabetic and numeric characters. People can\ncommunicate with one another through either oral or written communication.\nHowever, each language has a sign language counterpart. Individuals who are\ndeaf and/or mute communicate through sign language. The Bangla language also\nhas a sign language, which is called BDSL. The dataset is about Bangla hand\nsign images. The collection contains 49 individual Bangla alphabet images in\nsign language. BDSL49 is a dataset that consists of 29,490 images with 49\nlabels. Images of 14 different adult individuals, each with a distinct\nbackground and appearance, have been recorded during data collection. Several\nstrategies have been used to eliminate noise from datasets during preparation.\nThis dataset is available to researchers for free. They can develop automated\nsystems using machine learning, computer vision, and deep learning techniques.\nIn addition, two models were used in this dataset. The first is for detection,\nwhile the second is for recognition.", "title": "BDSL 49: A Comprehensive Dataset of Bangla Sign Language", "url": "http://arxiv.org/abs/2208.06827v1" }

{ "text_length": 1191 }

{ "abstract": "The detection of hate speech in political discourse is a critical issue, and\nthis becomes even more challenging in low-resource languages. To address this\nissue, we introduce a new dataset named IEHate, which contains 11,457 manually\nannotated Hindi tweets related to the Indian Assembly Election Campaign from\nNovember 1, 2021, to March 9, 2022. We performed a detailed analysis of the\ndataset, focusing on the prevalence of hate speech in political communication\nand the different forms of hateful language used. Additionally, we benchmark\nthe dataset using a range of machine learning, deep learning, and\ntransformer-based algorithms. Our experiments reveal that the performance of\nthese models can be further improved, highlighting the need for more advanced\ntechniques for hate speech detection in low-resource languages. In particular,\nthe relatively higher score of human evaluation over algorithms emphasizes the\nimportance of utilizing both human and automated approaches for effective hate\nspeech moderation. Our IEHate dataset can serve as a valuable resource for\nresearchers and practitioners working on developing and evaluating hate speech\ndetection techniques in low-resource languages. Overall, our work underscores\nthe importance of addressing the challenges of identifying and mitigating hate\nspeech in political discourse, particularly in the context of low-resource\nlanguages. The dataset and resources for this work are made available at\nhttps://github.com/Farhan-jafri/Indian-Election.", "title": "Uncovering Political Hate Speech During Indian Election Campaign: A New Low-Resource Dataset and Baselines", "url": "http://arxiv.org/abs/2306.14764v2" }

{ "text_length": 1648 }

{ "abstract": "As restricted quantum computers are slowly becoming a reality, the search for\nmeaningful first applications intensifies. In this domain, one of the more\ninvestigated approaches is the use of a special type of quantum circuit - a\nso-called quantum neural network -- to serve as a basis for a machine learning\nmodel. Roughly speaking, as the name suggests, a quantum neural network can\nplay a similar role to a neural network. However, specifically for applications\nin machine learning contexts, very little is known about suitable circuit\narchitectures, or model hyperparameters one should use to achieve good learning\nperformance. In this work, we apply the functional ANOVA framework to quantum\nneural networks to analyze which of the hyperparameters were most influential\nfor their predictive performance. We analyze one of the most typically used\nquantum neural network architectures. We then apply this to $7$ open-source\ndatasets from the OpenML-CC18 classification benchmark whose number of features\nis small enough to fit on quantum hardware with less than $20$ qubits. Three\nmain levels of importance were detected from the ranking of hyperparameters\nobtained with functional ANOVA. Our experiment both confirmed expected patterns\nand revealed new insights. For instance, setting well the learning rate is\ndeemed the most critical hyperparameter in terms of marginal contribution on\nall datasets, whereas the particular choice of entangling gates used is\nconsidered the least important except on one dataset. This work introduces new\nmethodologies to study quantum machine learning models and provides new\ninsights toward quantum model selection.", "title": "Hyperparameter Importance of Quantum Neural Networks Across Small Datasets", "url": "http://arxiv.org/abs/2206.09992v1" }

{ "text_length": 1763 }

{ "abstract": "This paper provides an analysis of character-level machine translation models\nused in pivot-based translation when applied to sparse and noisy datasets, such\nas crowdsourced movie subtitles. In our experiments, we find that such\ncharacter-level models cut the number of untranslated words by over 40% and are\nespecially competitive (improvements of 2-3 BLEU points) in the case of limited\ntraining data. We explore the impact of character alignment, phrase table\nfiltering, bitext size and the choice of pivot language on translation quality.\nWe further compare cascaded translation models to the use of synthetic training\ndata via multiple pivots, and we find that the latter works significantly\nbetter. Finally, we demonstrate that neither word-nor character-BLEU correlate\nperfectly with human judgments, due to BLEU's sensitivity to length.", "title": "Analyzing the Use of Character-Level Translation with Sparse and Noisy Datasets", "url": "http://arxiv.org/abs/2109.13723v1" }

{ "text_length": 958 }

{ "abstract": "The area affected by the earthquake is vast and often difficult to entirely\ncover, and the earthquake itself is a sudden event that causes multiple defects\nsimultaneously, that cannot be effectively traced using traditional, manual\nmethods. This article presents an innovative approach to the problem of\ndetecting damage after sudden events by using an interconnected set of deep\nmachine learning models organized in a single pipeline and allowing for easy\nmodification and swapping models seamlessly. Models in the pipeline were\ntrained with a synthetic dataset and were adapted to be further evaluated and\nused with unmanned aerial vehicles (UAVs) in real-world conditions. Thanks to\nthe methods presented in the article, it is possible to obtain high accuracy in\ndetecting buildings defects, segmenting constructions into their components and\nestimating their technical condition based on a single drone flight.", "title": "Computer Vision based inspection on post-earthquake with UAV synthetic dataset", "url": "http://arxiv.org/abs/2210.05282v1" }

{ "text_length": 1027 }

{ "abstract": "In this work, we assess several deep learning strategies for hyperspectral\npansharpening. First, we present a new dataset with a greater extent than any\nother in the state of the art. This dataset, collected using the ASI PRISMA\nsatellite, covers about 262200 km2, and its heterogeneity is granted by\nrandomly sampling the Earth's soil. Second, we adapted several state of the art\napproaches based on deep learning to fit PRISMA hyperspectral data and then\nassessed, quantitatively and qualitatively, the performance in this new\nscenario. The investigation has included two settings: Reduced Resolution (RR)\nto evaluate the techniques in a supervised environment and Full Resolution (FR)\nfor a real-world evaluation. The main purpose is the evaluation of the\nreconstruction fidelity of the considered methods. In both scenarios, for the\nsake of completeness, we also included machine-learning-free approaches. From\nthis extensive analysis has emerged that data-driven neural network methods\noutperform machine-learning-free approaches and adapt better to the task of\nhyperspectral pansharpening, both in RR and FR protocols.", "title": "Deep Learning Hyperspectral Pansharpening on large scale PRISMA dataset", "url": "http://arxiv.org/abs/2307.11666v2" }

{ "text_length": 1230 }

{ "abstract": "Safety-critical applications require machine learning models that output\naccurate and calibrated probabilities. While uncalibrated deep networks are\nknown to make over-confident predictions, it is unclear how model confidence is\nimpacted by the variations in the data, such as label noise or class size. In\nthis paper, we investigate the role of the dataset quality by studying the\nimpact of dataset size and the label noise on the model confidence. We\ntheoretically explain and experimentally demonstrate that, surprisingly, label\nnoise in the training data leads to under-confident networks, while reduced\ndataset size leads to over-confident models. We then study the impact of\ndataset heterogeneity, where data quality varies across classes, on model\nconfidence. We demonstrate that this leads to heterogenous confidence/accuracy\nbehavior in the test data and is poorly handled by the standard calibration\nalgorithms. To overcome this, we propose an intuitive heterogenous calibration\ntechnique and show that the proposed approach leads to improved calibration\nmetrics (both average and worst-case errors) on the CIFAR datasets.", "title": "On the Role of Dataset Quality and Heterogeneity in Model Confidence", "url": "http://arxiv.org/abs/2002.09831v1" }

{ "text_length": 1235 }

{ "abstract": "The development of a materials synthesis route is usually based on heuristics\nand experience. A possible new approach would be to apply data-driven\napproaches to learn the patterns of synthesis from past experience and use them\nto predict the syntheses of novel materials. However, this route is impeded by\nthe lack of a large-scale database of synthesis formulations. In this work, we\napplied advanced machine learning and natural language processing techniques to\nconstruct a dataset of 35,675 solution-based synthesis \"recipes\" extracted from\nthe scientific literature. Each recipe contains essential synthesis information\nincluding the precursors and target materials, their quantities, and the\nsynthesis actions and corresponding attributes. Every recipe is also augmented\nwith the reaction formula. Through this work, we are making freely available\nthe first large dataset of solution-based inorganic materials synthesis\nrecipes.", "title": "Dataset of Solution-based Inorganic Materials Synthesis Recipes Extracted from the Scientific Literature", "url": "http://arxiv.org/abs/2111.10874v1" }

{ "text_length": 1074 }

{ "abstract": "Application of interpretable machine learning techniques on medical datasets\nfacilitate early and fast diagnoses, along with getting deeper insight into the\ndata. Furthermore, the transparency of these models increase trust among\napplication domain experts. Medical datasets face common issues such as\nheterogeneous measurements, imbalanced classes with limited sample size, and\nmissing data, which hinder the straightforward application of machine learning\ntechniques. In this paper we present a family of prototype-based (PB)\ninterpretable models which are capable of handling these issues. The models\nintroduced in this contribution show comparable or superior performance to\nalternative techniques applicable in such situations. However, unlike ensemble\nbased models, which have to compromise on easy interpretation, the PB models\nhere do not. Moreover we propose a strategy of harnessing the power of\nensembles while maintaining the intrinsic interpretability of the PB models, by\naveraging the model parameter manifolds. All the models were evaluated on a\nsynthetic (publicly available dataset) in addition to detailed analyses of two\nreal-world medical datasets (one publicly available). Results indicated that\nthe models and strategies we introduced addressed the challenges of real-world\nmedical data, while remaining computationally inexpensive and transparent, as\nwell as similar or superior in performance compared to their alternatives.", "title": "Interpretable Models Capable of Handling Systematic Missingness in Imbalanced Classes and Heterogeneous Datasets", "url": "http://arxiv.org/abs/2206.02056v1" }

{ "text_length": 1596 }

{ "abstract": "The localization speed and accuracy in the indoor scenario can greatly impact\nthe Quality of Experience of the user. While many individual machine learning\nmodels can achieve comparable positioning performance, their prediction\nmechanisms offer different complexity to the system. In this work, we propose a\nfingerprinting positioning method for multi-building and multi-floor\ndeployments, composed of a cascade of three models for building classification,\nfloor classification, and 2D localization regression. We conduct an exhaustive\nsearch for the optimally performing one in each step of the cascade while\nvalidating on 14 different openly available datasets. As a result, we bring\nforward the best-performing combination of models in terms of overall\npositioning accuracy and processing speed and evaluate on independent sets of\nsamples. We reduce the mean prediction time by 71% while achieving comparable\npositioning performance across all considered datasets. Moreover, in case of\nvoluminous training dataset, the prediction time is reduced down to 1% of the\nbenchmark's.", "title": "Towards Accelerated Localization Performance Across Indoor Positioning Datasets", "url": "http://arxiv.org/abs/2204.10788v1" }

{ "text_length": 1193 }

{ "abstract": "This article describes techniques employed in the production of a synthetic\ndataset of driver telematics emulated from a similar real insurance dataset.\nThe synthetic dataset generated has 100,000 policies that included observations\nabout driver's claims experience together with associated classical risk\nvariables and telematics-related variables. This work is aimed to produce a\nresource that can be used to advance models to assess risks for usage-based\ninsurance. It follows a three-stage process using machine learning algorithms.\nThe first stage is simulating values for the number of claims as multiple\nbinary classifications applying feedforward neural networks. The second stage\nis simulating values for aggregated amount of claims as regression using\nfeedforward neural networks, with number of claims included in the set of\nfeature variables. In the final stage, a synthetic portfolio of the space of\nfeature variables is generated applying an extended $\\texttt{SMOTE}$ algorithm.\nThe resulting dataset is evaluated by comparing the synthetic and real datasets\nwhen Poisson and gamma regression models are fitted to the respective data.\nOther visualization and data summarization produce remarkable similar\nstatistics between the two datasets. We hope that researchers interested in\nobtaining telematics datasets to calibrate models or learning algorithms will\nfind our work valuable.", "title": "Synthetic Dataset Generation of Driver Telematics", "url": "http://arxiv.org/abs/2102.00252v1" }

{ "text_length": 1480 }

{ "abstract": "Recent advances in machine learning have significantly improved the\nunderstanding of source code data and achieved good performance on a number of\ndownstream tasks. Open source repositories like GitHub enable this process with\nrich unlabeled code data. However, the lack of high quality labeled data has\nlargely hindered the progress of several code related tasks, such as program\ntranslation, summarization, synthesis, and code search. This paper introduces\nXLCoST, Cross-Lingual Code SnippeT dataset, a new benchmark dataset for\ncross-lingual code intelligence. Our dataset contains fine-grained parallel\ndata from 8 languages (7 commonly used programming languages and English), and\nsupports 10 cross-lingual code tasks. To the best of our knowledge, it is the\nlargest parallel dataset for source code both in terms of size and the number\nof languages. We also provide the performance of several state-of-the-art\nbaseline models for each task. We believe this new dataset can be a valuable\nasset for the research community and facilitate the development and validation\nof new methods for cross-lingual code intelligence.", "title": "XLCoST: A Benchmark Dataset for Cross-lingual Code Intelligence", "url": "http://arxiv.org/abs/2206.08474v1" }

{ "text_length": 1221 }

{ "abstract": "We provide here a dataset for tasks related to natural language understanding\nand natural language inference. The dataset contains logical puzzles in natural\nlanguage from three domains: comparing puzzles, knighs and knaves, and zebra\npuzzles. Each puzzle is associated with the entire set of atomic questions that\ncan be generated based on the relations and individuals occurring in the text.\nFor each question we provide the correct answer: entailment, contradiction or\nambiguity. The answer's correctness is verified against theorem provers. Good\npuzzles have two properties: (i) each piece of information is necessary and\n(ii) no unnecessary information is provided. These properties make puzzles\ninteresting candidates for machine comprehension tasks.", "title": "A Puzzle-Based Dataset for Natural Language Inference", "url": "http://arxiv.org/abs/2112.05742v1" }

{ "text_length": 844 }

{ "abstract": "The paper describes the MetroPT data set, an outcome of a eXplainable\nPredictive Maintenance (XPM) project with an urban metro public transportation\nservice in Porto, Portugal. The data was collected in 2022 that aimed to\nevaluate machine learning methods for online anomaly detection and failure\nprediction. By capturing several analogic sensor signals (pressure,\ntemperature, current consumption), digital signals (control signals, discrete\nsignals), and GPS information (latitude, longitude, and speed), we provide a\ndataset that can be easily used to evaluate online machine learning methods.\nThis dataset contains some interesting characteristics and can be a good\nbenchmark for predictive maintenance models.", "title": "A Benchmark dataset for predictive maintenance", "url": "http://arxiv.org/abs/2207.05466v3" }

{ "text_length": 795 }

{ "abstract": "Providing explanations in the context of Visual Question Answering (VQA)\npresents a fundamental problem in machine learning. To obtain detailed insights\ninto the process of generating natural language explanations for VQA, we\nintroduce the large-scale CLEVR-X dataset that extends the CLEVR dataset with\nnatural language explanations. For each image-question pair in the CLEVR\ndataset, CLEVR-X contains multiple structured textual explanations which are\nderived from the original scene graphs. By construction, the CLEVR-X\nexplanations are correct and describe the reasoning and visual information that\nis necessary to answer a given question. We conducted a user study to confirm\nthat the ground-truth explanations in our proposed dataset are indeed complete\nand relevant. We present baseline results for generating natural language\nexplanations in the context of VQA using two state-of-the-art frameworks on the\nCLEVR-X dataset. Furthermore, we provide a detailed analysis of the explanation\ngeneration quality for different question and answer types. Additionally, we\nstudy the influence of using different numbers of ground-truth explanations on\nthe convergence of natural language generation (NLG) metrics. The CLEVR-X\ndataset is publicly available at\n\\url{https://explainableml.github.io/CLEVR-X/}.", "title": "CLEVR-X: A Visual Reasoning Dataset for Natural Language Explanations", "url": "http://arxiv.org/abs/2204.02380v1" }

{ "text_length": 1408 }

{ "abstract": "Representation learning focused on disentangling the underlying factors of\nvariation in given data has become an important area of research in machine\nlearning. However, most of the studies in this area have relied on datasets\nfrom the computer vision domain and thus, have not been readily extended to\nmusic. In this paper, we present a new symbolic music dataset that will help\nresearchers working on disentanglement problems demonstrate the efficacy of\ntheir algorithms on diverse domains. This will also provide a means for\nevaluating algorithms specifically designed for music. To this end, we create a\ndataset comprising of 2-bar monophonic melodies where each melody is the result\nof a unique combination of nine latent factors that span ordinal, categorical,\nand binary types. The dataset is large enough (approx. 1.3 million data points)\nto train and test deep networks for disentanglement learning. In addition, we\npresent benchmarking experiments using popular unsupervised disentanglement\nalgorithms on this dataset and compare the results with those obtained on an\nimage-based dataset.", "title": "dMelodies: A Music Dataset for Disentanglement Learning", "url": "http://arxiv.org/abs/2007.15067v1" }

{ "text_length": 1188 }

{ "abstract": "Predicting the probability of non-performing loans for individuals has a\nvital and beneficial role for banks to decrease credit risk and make the right\ndecisions before giving the loan. The trend to make these decisions are based\non credit study and in accordance with generally accepted standards, loan\npayment history, and demographic data of the clients. In this work, we evaluate\nhow different Machine learning models such as Random Forest, Decision tree,\nKNN, SVM, and XGBoost perform on the dataset provided by a private bank in\nEthiopia. Further, motivated by this evaluation we explore different feature\nselection methods to state the important features for the bank. Our findings\nshow that XGBoost achieves the highest F1 score on the KMeans SMOTE\nover-sampled data. We also found that the most important features are the age\nof the applicant, years of employment, and total income of the applicant rather\nthan collateral-related features in evaluating credit risk.", "title": "Machine Learning Models Evaluation and Feature Importance Analysis on NPL Dataset", "url": "http://arxiv.org/abs/2209.09638v1" }

{ "text_length": 1090 }

{ "abstract": "In many machine learning domains, datasets are characterized by highly\nimbalanced and overlapping classes. Particularly in the medical domain, a\nspecific list of symptoms can be labeled as one of various different\nconditions. Some of these conditions may be more prevalent than others by\nseveral orders of magnitude. Here we present a novel unsupervised domain\nadaptation scheme for such datasets. The scheme, based on a specific type of\nQuantification, is designed to work under both label and conditional shifts. It\nis demonstrated on datasets generated from electronic health records and\nprovides high quality results for both Quantification and Domain Adaptation in\nvery challenging scenarios. Potential benefits of using this scheme in the\ncurrent COVID-19 outbreak, for estimation of prevalence and probability of\ninfection are discussed.", "title": "Domain Adaptation in Highly Imbalanced and Overlapping Datasets", "url": "http://arxiv.org/abs/2005.03585v2" }

{ "text_length": 942 }

{ "abstract": "We present the development of a dataset for Kazakh named entity recognition.\nThe dataset was built as there is a clear need for publicly available annotated\ncorpora in Kazakh, as well as annotation guidelines containing\nstraightforward--but rigorous--rules and examples. The dataset annotation,\nbased on the IOB2 scheme, was carried out on television news text by two native\nKazakh speakers under the supervision of the first author. The resulting\ndataset contains 112,702 sentences and 136,333 annotations for 25 entity\nclasses. State-of-the-art machine learning models to automatise Kazakh named\nentity recognition were also built, with the best-performing model achieving an\nexact match F1-score of 97.22% on the test set. The annotated dataset,\nguidelines, and codes used to train the models are freely available for\ndownload under the CC BY 4.0 licence from https://github.com/IS2AI/KazNERD.", "title": "KazNERD: Kazakh Named Entity Recognition Dataset", "url": "http://arxiv.org/abs/2111.13419v2" }

{ "text_length": 979 }

{ "abstract": "Pir\\'a is a reading comprehension dataset focused on the ocean, the Brazilian\ncoast, and climate change, built from a collection of scientific abstracts and\nreports on these topics. This dataset represents a versatile language resource,\nparticularly useful for testing the ability of current machine learning models\nto acquire expert scientific knowledge. Despite its potential, a detailed set\nof baselines has not yet been developed for Pir\\'a. By creating these\nbaselines, researchers can more easily utilize Pir\\'a as a resource for testing\nmachine learning models across a wide range of question answering tasks. In\nthis paper, we define six benchmarks over the Pir\\'a dataset, covering closed\ngenerative question answering, machine reading comprehension, information\nretrieval, open question answering, answer triggering, and multiple choice\nquestion answering. As part of this effort, we have also produced a curated\nversion of the original dataset, where we fixed a number of grammar issues,\nrepetitions, and other shortcomings. Furthermore, the dataset has been extended\nin several new directions, so as to face the aforementioned benchmarks:\ntranslation of supporting texts from English into Portuguese, classification\nlabels for answerability, automatic paraphrases of questions and answers, and\nmultiple choice candidates. The results described in this paper provide several\npoints of reference for researchers interested in exploring the challenges\nprovided by the Pir\\'a dataset.", "title": "Benchmarks for Pirá 2.0, a Reading Comprehension Dataset about the Ocean, the Brazilian Coast, and Climate Change", "url": "http://arxiv.org/abs/2309.10945v1" }

{ "text_length": 1640 }

{ "abstract": "Metasurfaces have received a lot of attentions recently due to their\nversatile capability in manipulating electromagnetic wave. Advanced designs to\nsatisfy multiple objectives with non-linear constraints have motivated\nresearchers in using machine learning (ML) techniques like deep learning (DL)\nfor accelerated design of metasurfaces. For metasurfaces, it is difficult to\nmake quantitative comparisons between different ML models without having a\ncommon and yet complex dataset used in many disciplines like image\nclassification. Many studies were directed to a relatively constrained datasets\nthat are limited to specified patterns or shapes in metasurfaces. In this\npaper, we present our SUTD polarized reflection of complex metasurfaces\n(SUTD-PRCM) dataset, which contains approximately 260,000 samples of complex\nmetasurfaces created from electromagnetic simulation, and it has been used to\nbenchmark our DL models. The metasurface patterns are divided into different\nclasses to facilitate different degree of complexity, which involves\nidentifying and exploiting the relationship between the patterns and the\nelectromagnetic responses that can be compared in using different DL models.\nWith the release of this SUTD-PRCM dataset, we hope that it will be useful for\nbenchmarking existing or future DL models developed in the ML community. We\nalso propose a classification problem that is less encountered and apply neural\narchitecture search to have a preliminary understanding of potential\nmodification to the neural architecture that will improve the prediction by DL\nmodels. Our finding shows that convolution stacking is not the dominant element\nof the neural architecture anymore, which implies that low-level features are\npreferred over the traditional deep hierarchical high-level features thus\nexplains why deep convolutional neural network based models are not performing\nwell in our dataset.", "title": "SUTD-PRCM Dataset and Neural Architecture Search Approach for Complex Metasurface Design", "url": "http://arxiv.org/abs/2203.00002v1" }

{ "text_length": 2030 }

{ "abstract": "There has been increasing interest in smart factories powered by robotics\nsystems to tackle repetitive, laborious tasks. One impactful yet challenging\ntask in robotics-powered smart factory applications is robotic grasping: using\nrobotic arms to grasp objects autonomously in different settings. Robotic\ngrasping requires a variety of computer vision tasks such as object detection,\nsegmentation, grasp prediction, pick planning, etc. While significant progress\nhas been made in leveraging of machine learning for robotic grasping,\nparticularly with deep learning, a big challenge remains in the need for\nlarge-scale, high-quality RGBD datasets that cover a wide diversity of\nscenarios and permutations. To tackle this big, diverse data problem, we are\ninspired by the recent rise in the concept of metaverse, which has greatly\nclosed the gap between virtual worlds and the physical world. Metaverses allow\nus to create digital twins of real-world manufacturing scenarios and to\nvirtually create different scenarios from which large volumes of data can be\ngenerated for training models. In this paper, we present MetaGraspNet: a\nlarge-scale benchmark dataset for vision-driven robotic grasping via\nphysics-based metaverse synthesis. The proposed dataset contains 100,000 images\nand 25 different object types and is split into 5 difficulties to evaluate\nobject detection and segmentation model performance in different grasping\nscenarios. We also propose a new layout-weighted performance metric alongside\nthe dataset for evaluating object detection and segmentation performance in a\nmanner that is more appropriate for robotic grasp applications compared to\nexisting general-purpose performance metrics. Our benchmark dataset is\navailable open-source on Kaggle, with the first phase consisting of detailed\nobject detection, segmentation, layout annotations, and a layout-weighted\nperformance metric script.", "title": "MetaGraspNet_v0: A Large-Scale Benchmark Dataset for Vision-driven Robotic Grasping via Physics-based Metaverse Synthesis", "url": "http://arxiv.org/abs/2112.14663v3" }

{ "text_length": 2062 }

{ "abstract": "There have been many papers with algorithms for improving fairness of\nmachine-learning classifiers for tabular data. Unfortunately, most use only\nvery few datasets for their experimental evaluation. We introduce a suite of\nfunctions for fetching 20 fairness datasets and providing associated fairness\nmetadata. Hopefully, these will lead to more rigorous experimental evaluations\nin future fairness-aware machine learning research.", "title": "A Suite of Fairness Datasets for Tabular Classification", "url": "http://arxiv.org/abs/2308.00133v1" }

{ "text_length": 521 }

{ "abstract": "We introduce MADLAD-400, a manually audited, general domain 3T token\nmonolingual dataset based on CommonCrawl, spanning 419 languages. We discuss\nthe limitations revealed by self-auditing MADLAD-400, and the role data\nauditing had in the dataset creation process. We then train and release a\n10.7B-parameter multilingual machine translation model on 250 billion tokens\ncovering over 450 languages using publicly available data, and find that it is\ncompetitive with models that are significantly larger, and report the results\non different domains. In addition, we train a 8B-parameter language model, and\nassess the results on few-shot translation. We make the baseline models\navailable to the research community.", "title": "MADLAD-400: A Multilingual And Document-Level Large Audited Dataset", "url": "http://arxiv.org/abs/2309.04662v1" }

{ "text_length": 815 }

{ "abstract": "Subseasonal forecasting of the weather two to six weeks in advance is\ncritical for resource allocation and climate adaptation but poses many\nchallenges for the forecasting community. At this forecast horizon,\nphysics-based dynamical models have limited skill, and the targets for\nprediction depend in a complex manner on both local weather and global climate\nvariables. Recently, machine learning methods have shown promise in advancing\nthe state of the art but only at the cost of complex data curation, integrating\nexpert knowledge with aggregation across multiple relevant data sources, file\nformats, and temporal and spatial resolutions. To streamline this process and\naccelerate future development, we introduce SubseasonalClimateUSA, a curated\ndataset for training and benchmarking subseasonal forecasting models in the\nUnited States. We use this dataset to benchmark a diverse suite of subseasonal\nmodels, including operational dynamical models, classical meteorological\nbaselines, and ten state-of-the-art machine learning and deep learning-based\nmethods from the literature. Overall, our benchmarks suggest simple and\neffective ways to extend the accuracy of current operational models.\nSubseasonalClimateUSA is regularly updated and accessible via the\nhttps://github.com/microsoft/subseasonal_data/ Python package.", "title": "SubseasonalClimateUSA: A Dataset for Subseasonal Forecasting and Benchmarking", "url": "http://arxiv.org/abs/2109.10399v3" }

{ "text_length": 1436 }

{ "abstract": "The current interacting hand (IH) datasets are relatively simplistic in terms\nof background and texture, with hand joints being annotated by a machine\nannotator, which may result in inaccuracies, and the diversity of pose\ndistribution is limited. However, the variability of background, pose\ndistribution, and texture can greatly influence the generalization ability.\nTherefore, we present a large-scale synthetic dataset RenderIH for interacting\nhands with accurate and diverse pose annotations. The dataset contains 1M\nphoto-realistic images with varied backgrounds, perspectives, and hand\ntextures. To generate natural and diverse interacting poses, we propose a new\npose optimization algorithm. Additionally, for better pose estimation accuracy,\nwe introduce a transformer-based pose estimation network, TransHand, to\nleverage the correlation between interacting hands and verify the effectiveness\nof RenderIH in improving results. Our dataset is model-agnostic and can improve\nmore accuracy of any hand pose estimation method in comparison to other real or\nsynthetic datasets. Experiments have shown that pretraining on our synthetic\ndata can significantly decrease the error from 6.76mm to 5.79mm, and our\nTranshand surpasses contemporary methods. Our dataset and code are available at\nhttps://github.com/adwardlee/RenderIH.", "title": "RenderIH: A Large-scale Synthetic Dataset for 3D Interacting Hand Pose Estimation", "url": "http://arxiv.org/abs/2309.09301v3" }

{ "text_length": 1446 }

{ "abstract": "Machine learning methodologies can be adopted in cultural applications and\npropose new ways to distribute or even present the cultural content to the\npublic. For instance, speech analytics can be adopted to automatically generate\nsubtitles in theatrical plays, in order to (among other purposes) help people\nwith hearing loss. Apart from a typical speech-to-text transcription with\nAutomatic Speech Recognition (ASR), Speech Emotion Recognition (SER) can be\nused to automatically predict the underlying emotional content of speech\ndialogues in theatrical plays, and thus to provide a deeper understanding how\nthe actors utter their lines. However, real-world datasets from theatrical\nplays are not available in the literature. In this work we present GreThE, the\nGreek Theatrical Emotion dataset, a new publicly available data collection for\nspeech emotion recognition in Greek theatrical plays. The dataset contains\nutterances from various actors and plays, along with respective valence and\narousal annotations. Towards this end, multiple annotators have been asked to\nprovide their input for each speech recording and inter-annotator agreement is\ntaken into account in the final ground truth generation. In addition, we\ndiscuss the results of some indicative experiments that have been conducted\nwith machine and deep learning frameworks, using the dataset, along with some\nwidely used databases in the field of speech emotion recognition.", "title": "A Dataset for Speech Emotion Recognition in Greek Theatrical Plays", "url": "http://arxiv.org/abs/2203.15568v1" }

{ "text_length": 1543 }

{ "abstract": "The recent decade has seen an enormous rise in the popularity of deep\nlearning and neural networks. These algorithms have broken many previous\nrecords and achieved remarkable results. Their outstanding performance has\nsignificantly sped up the progress of AI, and so far various milestones have\nbeen achieved earlier than expected. However, in the case of relatively small\ndatasets, the performance of Deep Neural Networks (DNN) may suffer from reduced\naccuracy compared to other Machine Learning models. Furthermore, it is\ndifficult to construct prediction intervals or evaluate the uncertainty of\npredictions when dealing with regression tasks. In this paper, we propose an\nensemble method that attempts to estimate the uncertainty of predictions,\nincrease their accuracy and provide an interval for the expected variation.\nCompared with traditional DNNs that only provide a prediction, our proposed\nmethod can output a prediction interval by combining DNNs, extreme gradient\nboosting (XGBoost) and dissimilarity computation techniques. Albeit the simple\ndesign, this approach significantly increases accuracy on small datasets and\ndoes not introduce much complexity to the architecture of the neural network.\nThe proposed method is tested on various datasets, and a significant\nimprovement in the performance of the neural network model is seen. The model's\nprediction interval can include the ground truth value at an average rate of\n71% and 78% across training sizes of 90% and 55%, respectively. Finally, we\nhighlight other aspects and applications of the approach in experimental error\nestimation, and the application of transfer learning.", "title": "Confidence-Nets: A Step Towards better Prediction Intervals for regression Neural Networks on small datasets", "url": "http://arxiv.org/abs/2210.17092v1" }

{ "text_length": 1789 }

{ "abstract": "Super-Resolution (SR) techniques aim to enhance data resolution, enabling the\nretrieval of finer details, and improving the overall quality and fidelity of\nthe data representation. There is growing interest in applying SR methods to\ncomplex spatiotemporal systems within the Scientific Machine Learning (SciML)\ncommunity, with the hope of accelerating numerical simulations and/or improving\nforecasts in weather, climate, and related areas. However, the lack of\nstandardized benchmark datasets for comparing and validating SR methods hinders\nprogress and adoption in SciML. To address this, we introduce SuperBench, the\nfirst benchmark dataset featuring high-resolution datasets (up to\n$2048\\times2048$ dimensions), including data from fluid flows, cosmology, and\nweather. Here, we focus on validating spatial SR performance from data-centric\nand physics-preserved perspectives, as well as assessing robustness to data\ndegradation tasks. While deep learning-based SR methods (developed in the\ncomputer vision community) excel on certain tasks, despite relatively limited\nprior physics information, we identify limitations of these methods in\naccurately capturing intricate fine-scale features and preserving fundamental\nphysical properties and constraints in scientific data. These shortcomings\nhighlight the importance and subtlety of incorporating domain knowledge into ML\nmodels. We anticipate that SuperBench will significantly advance SR methods for\nscientific tasks.", "title": "SuperBench: A Super-Resolution Benchmark Dataset for Scientific Machine Learning", "url": "http://arxiv.org/abs/2306.14070v1" }

{ "text_length": 1587 }

{ "abstract": "Given the time and expense associated with bringing a drug to market,\nnumerous studies have been conducted to predict the properties of compounds\nbased on their structure using machine learning. Federated learning has been\napplied to compound datasets to increase their prediction accuracy while\nsafeguarding potentially proprietary information. However, federated learning\nis encumbered by low accuracy in not identically and independently distributed\n(non-IID) settings, i.e., data partitioning has a large label bias, and is\nconsidered unsuitable for compound datasets, which tend to have large label\nbias. To address this limitation, we utilized an alternative method of\ndistributed machine learning to chemical compound data from open sources,\ncalled data collaboration analysis (DC). We also proposed data collaboration\nanalysis using projection data (DCPd), which is an improved method that\nutilizes auxiliary PubChem data. This improves the quality of individual\nuser-side data transformations for the projection data for the creation of\nintermediate representations. The classification accuracy, i.e., area under the\ncurve in the receiver operating characteristic curve (ROC-AUC) and AUC in the\nprecision-recall curve (PR-AUC), of federated averaging (FedAvg), DC, and DCPd\nwas compared for five compound datasets. We determined that the machine\nlearning performance for non-IID settings was in the order of DCPd, DC, and\nFedAvg, although they were almost the same in identically and independently\ndistributed (IID) settings. Moreover, the results showed that compared to other\nmethods, DCPd exhibited a negligible decline in classification accuracy in\nexperiments with different degrees of label bias. Thus, DCPd can address the\nlow performance in non-IID settings, which is one of the challenges of\nfederated learning.", "title": "Data Collaboration Analysis applied to Compound Datasets and the Introduction of Projection data to Non-IID settings", "url": "http://arxiv.org/abs/2308.00280v1" }

{ "text_length": 1980 }

{ "abstract": "Deep neural networks (DNNs) have played a key role in a wide range of machine\nlearning applications. However, DNN classifiers are vulnerable to\nhuman-imperceptible adversarial perturbations, which can cause them to\nmisclassify inputs with high confidence. Thus, creating robust DNNs which can\ndefend against malicious examples is critical in applications where security\nplays a major role. In this paper, we study the effect of intrinsic dataset\nproperties on the performance of adversarial attack and defense methods,\ntesting on five popular image classification datasets - MNIST, Fashion-MNIST,\nCIFAR10/CIFAR100, and ImageNet. We find that input size and image contrast play\nkey roles in attack and defense success. Our discoveries highlight that dataset\ndesign and data preprocessing steps are important to boost the adversarial\nrobustness of DNNs. To our best knowledge, this is the first comprehensive work\nthat studies the effect of intrinsic dataset properties on adversarial machine\nlearning.", "title": "On Intrinsic Dataset Properties for Adversarial Machine Learning", "url": "http://arxiv.org/abs/2005.09170v1" }

{ "text_length": 1099 }

{ "abstract": "During the COVID-19 pandemic, computed tomography (CT) is a good way to\ndiagnose COVID-19 patients. HRCT (High-Resolution Computed Tomography) is a\nform of computed tomography that uses advanced methods to improve image\nresolution. Publicly accessible COVID-19 CT image datasets are very difficult\nto come by due to privacy concerns, which impedes the study and development of\nAI-powered COVID-19 diagnostic algorithms based on CT images. To address this\nproblem, we have introduced HRCTv1-COVID-19, a new COVID-19 high resolution\nchest CT Scan image dataset that includes not only COVID-19 cases of Ground\nGlass Opacity (GGO), Crazy Paving, and Air Space Consolidation, but also CT\nimages of cases with negative COVID-19. The HRCTv1-COVID-19 dataset, which\nincludes slice-level, and patient-level labels, has the potential to aid\nCOVID-19 research, especially for diagnosis and differentiation using\nartificial intelligence algorithms, machine learning and deep learning methods.\nThis dataset is accessible through web at: http://databiox.com and includes\n181,106 chest HRCT images from 395 patients with four labels: GGO, Crazy\nPaving, Air Space Consolidation and Negative.\n Keywords- Dataset, COVID-19, CT-Scan, Computed Tomography, Medical Imaging,\nChest Image.", "title": "A High-Resolution Chest CT-Scan Image Dataset for COVID-19 Diagnosis and Differentiation", "url": "http://arxiv.org/abs/2205.03408v1" }

{ "text_length": 1389 }

{ "abstract": "Detecting quality in large unstructured datasets requires capacities far\nbeyond the limits of human perception and communicability and, as a result,\nthere is an emerging trend towards increasingly complex analytic solutions in\ndata science to cope with this problem. This new trend towards analytic\ncomplexity represents a severe challenge for the principle of parsimony or\nOccams Razor in science. This review article combines insight from various\ndomains such as physics, computational science, data engineering, and cognitive\nscience to review the specific properties of big data. Problems for detecting\ndata quality without losing the principle of parsimony are then highlighted on\nthe basis of specific examples. Computational building block approaches for\ndata clustering can help to deal with large unstructured datasets in minimized\ncomputation time, and meaning can be extracted rapidly from large sets of\nunstructured image or video data parsimoniously through relatively simple\nunsupervised machine learning algorithms. Why we still massively lack in\nexpertise for exploiting big data wisely to extract relevant information for\nspecific tasks, recognize patterns, generate new information, or store and\nfurther process large amounts of sensor data is then reviewed; examples\nillustrating why we need subjective views and pragmatic methods to analyze big\ndata contents are brought forward. The review concludes on how cultural\ndifferences between East and West are likely to affect the course of big data\nanalytics, and the development of increasingly autonomous artificial\nintelligence aimed at coping with the big data deluge in the near future.", "title": "Occams Razor for Big Data? On Detecting Quality in Large Unstructured Datasets", "url": "http://arxiv.org/abs/2011.08663v1" }

{ "text_length": 1770 }

{ "abstract": "Purpose: The development of machine learning models for surgical workflow and\ninstrument recognition from temporal data represents a challenging task due to\nthe complex nature of surgical workflows. In particular, the imbalanced\ndistribution of data is one of the major challenges in the domain of surgical\nworkflow recognition. In order to obtain meaningful results, careful\npartitioning of data into training, validation, and test sets, as well as the\nselection of suitable evaluation metrics are crucial. Methods: In this work, we\npresent an openly available web-based application that enables interactive\nexploration of dataset partitions. The proposed visual framework facilitates\nthe assessment of dataset splits for surgical workflow recognition, especially\nwith regard to identifying sub-optimal dataset splits. Currently, it supports\nvisualization of surgical phase and instrument annotations. Results: In order\nto validate the dedicated interactive visualizations, we use a dataset split of\nthe Cholec80 dataset. This dataset split was specifically selected to reflect a\ncase of strong data imbalance. Using our software, we were able to identify\nphases, phase transitions, and combinations of surgical instruments that were\nnot represented in one of the sets. Conclusion: In order to obtain meaningful\nresults in highly unbalanced class distributions, special care should be taken\nwith respect to the selection of an appropriate split. Interactive data\nvisualization represents a promising approach for the assessment of machine\nlearning datasets. The source code is available at\nhttps://github.com/Cardio-AI/endovis-ml", "title": "Surgical Phase and Instrument Recognition: How to identify appropriate Dataset Splits", "url": "http://arxiv.org/abs/2306.16879v1" }

{ "text_length": 1750 }

{ "abstract": "Movie-making has become one of the most costly and risky endeavors in the\nentertainment industry. Continuous change in the preference of the audience\nmakes it harder to predict what kind of movie will be financially successful at\nthe box office. So, it is no wonder that cautious, intelligent stakeholders and\nlarge production houses will always want to know the probable revenue that will\nbe generated by a movie before making an investment. Researchers have been\nworking on finding an optimal strategy to help investors in making the right\ndecisions. But the lack of a large, up-to-date dataset makes their work harder.\nIn this work, we introduce an up-to-date, richer, and larger dataset that we\nhave prepared by scraping IMDb for researchers and data analysts to work with.\nThe compiled dataset contains the summery data of 7.5 million titles and detail\ninformation of more than 200K movies. Additionally, we perform different\nstatistical analysis approaches on our dataset to find out how a movie's\nrevenue is affected by different pre-released attributes such as budget,\nruntime, release month, content rating, genre etc. In our analysis, we have\nfound that having a star cast/director has a positive impact on generated\nrevenue. We introduce a novel approach for calculating the star power of a\nmovie. Based on our analysis we select a set of attributes as features and\ntrain different machine learning algorithms to predict a movie's expected\nrevenue. Based on generated revenue, we classified the movies in 10 categories\nand achieved a one-class-away accuracy rate of almost 60% (bingo accuracy of\n30%). All the generated datasets and analysis codes are available online. We\nalso made the source codes of our scraper bots public, so that researchers\ninterested in extending this work can easily modify these bots as they need and\nprepare their own up-to-date datasets.", "title": "Presenting a Larger Up-to-date Movie Dataset and Investigating the Effects of Pre-released Attributes on Gross Revenue", "url": "http://arxiv.org/abs/2110.07039v2" }

{ "text_length": 2030 }

{ "abstract": "Generative AI workflows heavily rely on data-centric tasks - such as\nfiltering samples by annotation fields, vector distances, or scores produced by\ncustom classifiers. At the same time, computer vision datasets are quickly\napproaching petabyte volumes, rendering data wrangling difficult. In addition,\nthe iterative nature of data preparation necessitates robust dataset sharing\nand versioning mechanisms, both of which are hard to implement ad-hoc. To solve\nthese challenges, we propose a \"dataset factory\" approach that separates the\nstorage and processing of samples from metadata and enables data-centric\noperations at scale for machine learning teams and individual researchers.", "title": "Dataset Factory: A Toolchain For Generative Computer Vision Datasets", "url": "http://arxiv.org/abs/2309.11608v1" }

{ "text_length": 787 }

{ "abstract": "Synthetic datasets are beneficial for machine learning researchers due to the\npossibility of experimenting with new strategies and algorithms in the training\nand testing phases. These datasets can easily include more scenarios that might\nbe costly to research with real data or can complement and, in some cases,\nreplace real data measurements, depending on the quality of the synthetic data.\nThey can also solve the unbalanced data problem, avoid overfitting, and can be\nused in training while testing can be done with real data. In this paper, we\npresent, to the best of our knowledge, the first synthetic dataset for Unmanned\nAerial Vehicle (UAV) attacks in 5G and beyond networks based on the following\nkey observable network parameters that indicate power levels: the Received\nSignal Strength Indicator (RSSI) and the Signal to Interference-plus-Noise\nRatio (SINR). The main objective of this data is to enable deep network\ndevelopment for UAV communication security. Especially, for algorithm\ndevelopment or the analysis of time-series data applied to UAV attack\nrecognition. Our proposed dataset provides insights into network functionality\nwhen static or moving UAV attackers target authenticated UAVs in an urban\nenvironment. The dataset also considers the presence and absence of\nauthenticated terrestrial users in the network, which may decrease the deep\nnetworks ability to identify attacks. Furthermore, the data provides deeper\ncomprehension of the metrics available in the 5G physical and MAC layers for\nmachine learning and statistics research. The dataset will available at link\narchive-beta.ics.uci.edu", "title": "A Synthetic Dataset for 5G UAV Attacks Based on Observable Network Parameters", "url": "http://arxiv.org/abs/2211.09706v1" }

{ "text_length": 1732 }

{ "abstract": "Nowadays the use of Machine Learning (ML) algorithms is spreading in the\nfield of Remote Sensing, with applications ranging from detection and\nclassification of land use and monitoring to the prediction of many natural or\nanthropic phenomena of interest. One main limit of their employment is related\nto the need for a huge amount of data for training the neural network, chosen\nfor the specific application, and the resulting computational weight and time\nrequired to collect the necessary data. In this letter the architecture of an\ninnovative tool, enabling researchers to create in an automatic way suitable\ndatasets for AI (Artificial Intelligence) applications in the EO (Earth\nObservation) context, is presented. Two versions of the architecture have been\nimplemented and made available on Git-Hub, with a specific Graphical User\nInterface (GUI) for non-expert users.", "title": "Automatic Dataset Builder for Machine Learning Applications to Satellite Imagery", "url": "http://arxiv.org/abs/2008.01578v1" }

{ "text_length": 989 }

{ "abstract": "Enhancing interoperability and information exchange between domain-specific\nsoftware products for BIM is an important aspect in the Architecture,\nEngineering, Construction and Operations industry. Recent research started\ninvestigating methods from the areas of machine and deep learning for semantic\nenrichment of BIM models. However, training and evaluation of these machine\nlearning algorithms requires sufficiently large and comprehensive datasets.\nThis work presents IFCNet, a dataset of single-entity IFC files spanning a\nbroad range of IFC classes containing both geometric and semantic information.\nUsing only the geometric information of objects, the experiments show that\nthree different deep learning models are able to achieve good classification\nperformance.", "title": "IFCNet: A Benchmark Dataset for IFC Entity Classification", "url": "http://arxiv.org/abs/2106.09712v1" }

{ "text_length": 862 }

{ "abstract": "The rise of Machine Learning (ML) is gradually digitalizing and reshaping the\nfashion industry. Recent years have witnessed a number of fashion AI\napplications, for example, virtual try-ons. Textile material identification and\ncategorization play a crucial role in the fashion textile sector, including\nfashion design, retails, and recycling. At the same time, Net Zero is a global\ngoal and the fashion industry is undergoing a significant change so that\ntextile materials can be reused, repaired and recycled in a sustainable manner.\nThere is still a challenge in identifying textile materials automatically for\ngarments, as we lack a low-cost and effective technique for identifying them.\nIn light of this, we build the first fashion textile dataset, TextileNet, based\non textile material taxonomies - a fibre taxonomy and a fabric taxonomy\ngenerated in collaboration with material scientists. TextileNet can be used to\ntrain and evaluate the state-of-the-art Deep Learning models for textile\nmaterials. We hope to standardize textile related datasets through the use of\ntaxonomies. TextileNet contains 33 fibres labels and 27 fabrics labels, and has\nin total 760,949 images. We use standard Convolutional Neural Networks (CNNs)\nand Vision Transformers (ViTs) to establish baselines for this dataset. Future\napplications for this dataset range from textile classification to optimization\nof the textile supply chain and interactive design for consumers. We envision\nthat this can contribute to the development of a new AI-based fashion platform.", "title": "TextileNet: A Material Taxonomy-based Fashion Textile Dataset", "url": "http://arxiv.org/abs/2301.06160v1" }

{ "text_length": 1643 }

{ "abstract": "The effectiveness of machine learning models is significantly affected by the\nsize of the dataset and the quality of features as redundant and irrelevant\nfeatures can radically degrade the performance. This paper proposes IGRF-RFE: a\nhybrid feature selection method tasked for multi-class network anomalies using\na Multilayer perceptron (MLP) network. IGRF-RFE can be considered as a feature\nreduction technique based on both the filter feature selection method and the\nwrapper feature selection method. In our proposed method, we use the filter\nfeature selection method, which is the combination of Information Gain and\nRandom Forest Importance, to reduce the feature subset search space. Then, we\napply recursive feature elimination(RFE) as a wrapper feature selection method\nto further eliminate redundant features recursively on the reduced feature\nsubsets. Our experimental results obtained based on the UNSW-NB15 dataset\nconfirm that our proposed method can improve the accuracy of anomaly detection\nwhile reducing the feature dimension. The results show that the feature\ndimension is reduced from 42 to 23 while the multi-classification accuracy of\nMLP is improved from 82.25% to 84.24%.", "title": "IGRF-RFE: A Hybrid Feature Selection Method for MLP-based Network Intrusion Detection on UNSW-NB15 Dataset", "url": "http://arxiv.org/abs/2203.16365v2" }

{ "text_length": 1335 }

{ "abstract": "Small Object Detection (SOD) is an important machine vision topic because (i)\na variety of real-world applications require object detection for distant\nobjects and (ii) SOD is a challenging task due to the noisy, blurred, and\nless-informative image appearances of small objects. This paper proposes a new\nSOD dataset consisting of 39,070 images including 137,121 bird instances, which\nis called the Small Object Detection for Spotting Birds (SOD4SB) dataset. The\ndetail of the challenge with the SOD4SB dataset is introduced in this paper. In\ntotal, 223 participants joined this challenge. This paper briefly introduces\nthe award-winning methods. The dataset, the baseline code, and the website for\nevaluation on the public testset are publicly available.", "title": "MVA2023 Small Object Detection Challenge for Spotting Birds: Dataset, Methods, and Results", "url": "http://arxiv.org/abs/2307.09143v1" }

{ "text_length": 880 }

{ "abstract": "Camera images are ubiquitous in machine learning research. They also play a\ncentral role in the delivery of important services spanning medicine and\nenvironmental surveying. However, the application of machine learning models in\nthese domains has been limited because of robustness concerns. A primary\nfailure mode are performance drops due to differences between the training and\ndeployment data. While there are methods to prospectively validate the\nrobustness of machine learning models to such dataset drifts, existing\napproaches do not account for explicit models of the primary object of\ninterest: the data. This limits our ability to study and understand the\nrelationship between data generation and downstream machine learning model\nperformance in a physically accurate manner. In this study, we demonstrate how\nto overcome this limitation by pairing traditional machine learning with\nphysical optics to obtain explicit and differentiable data models. We\ndemonstrate how such data models can be constructed for image data and used to\ncontrol downstream machine learning model performance related to dataset drift.\nThe findings are distilled into three applications. First, drift synthesis\nenables the controlled generation of physically faithful drift test cases to\npower model selection and targeted generalization. Second, the gradient\nconnection between machine learning task model and data model allows advanced,\nprecise tolerancing of task model sensitivity to changes in the data\ngeneration. These drift forensics can be used to precisely specify the\nacceptable data environments in which a task model may be run. Third, drift\noptimization opens up the possibility to create drifts that can help the task\nmodel learn better faster, effectively optimizing the data generating process\nitself. A guide to access the open code and datasets is available at\nhttps://github.com/aiaudit-org/raw2logit.", "title": "Data Models for Dataset Drift Controls in Machine Learning With Optical Images", "url": "http://arxiv.org/abs/2211.02578v3" }

{ "text_length": 2020 }

{ "abstract": "This paper empirically investigates the influence of different data splits\nand splitting strategies on the performance of dysfluency detection systems.\nFor this, we perform experiments using wav2vec 2.0 models with a classification\nhead as well as support vector machines (SVM) in conjunction with the features\nextracted from the wav2vec 2.0 model to detect dysfluencies. We train and\nevaluate the systems with different non-speaker-exclusive and speaker-exclusive\nsplits of the Stuttering Events in Podcasts (SEP-28k) dataset to shed some\nlight on the variability of results w.r.t. to the partition method used.\nFurthermore, we show that the SEP-28k dataset is dominated by only a few\nspeakers, making it difficult to evaluate. To remedy this problem, we created\nSEP-28k-Extended (SEP-28k-E), containing semi-automatically generated speaker\nand gender information for the SEP-28k corpus, and suggest different data\nsplits, each useful for evaluating other aspects of methods for dysfluency\ndetection.", "title": "The Influence of Dataset Partitioning on Dysfluency Detection Systems", "url": "http://arxiv.org/abs/2206.03400v1" }

{ "text_length": 1105 }

{ "abstract": "Knowledge is central to human and scientific developments. Natural Language\nProcessing (NLP) allows automated analysis and creation of knowledge. Data is a\ncrucial NLP and machine learning ingredient. The scarcity of open datasets is a\nwell-known problem in machine and deep learning research. This is very much the\ncase for textual NLP datasets in English and other major world languages. For\nthe Bangla language, the situation is even more challenging and the number of\nlarge datasets for NLP research is practically nil. We hereby present Potrika,\na large single-label Bangla news article textual dataset curated for NLP\nresearch from six popular online news portals in Bangladesh (Jugantor,\nJaijaidin, Ittefaq, Kaler Kontho, Inqilab, and Somoyer Alo) for the period\n2014-2020. The articles are classified into eight distinct categories\n(National, Sports, International, Entertainment, Economy, Education, Politics,\nand Science \\& Technology) providing five attributes (News Article, Category,\nHeadline, Publication Date, and Newspaper Source). The raw dataset contains\n185.51 million words and 12.57 million sentences contained in 664,880 news\narticles. Moreover, using NLP augmentation techniques, we create from the raw\n(unbalanced) dataset another (balanced) dataset comprising 320,000 news\narticles with 40,000 articles in each of the eight news categories. Potrika\ncontains both the datasets (raw and balanced) to suit a wide range of NLP\nresearch. By far, to the best of our knowledge, Potrika is the largest and the\nmost extensive dataset for news classification.", "title": "Potrika: Raw and Balanced Newspaper Datasets in the Bangla Language with Eight Topics and Five Attributes", "url": "http://arxiv.org/abs/2210.09389v1" }

{ "text_length": 1714 }

{ "abstract": "Research and education in machine learning needs diverse, representative, and\nopen datasets that contain sufficient samples to handle the necessary training,\nvalidation, and testing tasks. Currently, the Recommender Systems area includes\na large number of subfields in which accuracy and beyond accuracy quality\nmeasures are continuously improved. To feed this research variety, it is\nnecessary and convenient to reinforce the existing datasets with synthetic\nones. This paper proposes a Generative Adversarial Network (GAN)-based method\nto generate collaborative filtering datasets in a parameterized way, by\nselecting their preferred number of users, items, samples, and stochastic\nvariability. This parameterization cannot be made using regular GANs. Our GAN\nmodel is fed with dense, short, and continuous embedding representations of\nitems and users, instead of sparse, large, and discrete vectors, to make an\naccurate and quick learning, compared to the traditional approach based on\nlarge and sparse input vectors. The proposed architecture includes a DeepMF\nmodel to extract the dense user and item embeddings, as well as a clustering\nprocess to convert from the dense GAN generated samples to the discrete and\nsparse ones, necessary to create each required synthetic dataset. The results\nof three different source datasets show adequate distributions and expected\nquality values and evolutions on the generated datasets compared to the source\nones. Synthetic datasets and source codes are available to researchers.", "title": "Creating Synthetic Datasets for Collaborative Filtering Recommender Systems using Generative Adversarial Networks", "url": "http://arxiv.org/abs/2303.01297v1" }

{ "text_length": 1670 }

{ "abstract": "Dataset licensing is currently an issue in the development of machine\nlearning systems. And in the development of machine learning systems, the most\nwidely used are publicly available datasets. However, since the images in the\npublicly available dataset are mainly obtained from the Internet, some images\nare not commercially available. Furthermore, developers of machine learning\nsystems do not often care about the license of the dataset when training\nmachine learning models with it. In summary, the licensing of datasets for\nmachine learning systems is in a state of incompleteness in all aspects at this\nstage.\n Our investigation of two collection datasets revealed that most of the\ncurrent datasets lacked licenses, and the lack of licenses made it impossible\nto determine the commercial availability of the datasets. Therefore, we decided\nto take a more scientific and systematic approach to investigate the licensing\nof datasets and the licensing of machine learning systems that use the dataset\nto make it easier and more compliant for future developers of machine learning\nsystems.", "title": "An investigation of licensing of datasets for machine learning based on the GQM model", "url": "http://arxiv.org/abs/2303.13735v1" }

{ "text_length": 1212 }

{ "abstract": "This work advances the field of sign language machine translation by focusing\non dataset quality and simplification of the translation system. We introduce\nSignBank+, a clean version of the SignBank dataset, optimized for machine\ntranslation. Contrary to previous works that employ complex factorization\ntechniques for translation, we advocate for a simplified text-to-text\ntranslation approach. Our evaluation shows that models trained on SignBank+\nsurpass those on the original dataset, establishing a new benchmark and\nproviding an open resource for future research.", "title": "SignBank+: Multilingual Sign Language Translation Dataset", "url": "http://arxiv.org/abs/2309.11566v1" }

{ "text_length": 661 }

{ "abstract": "Fairness (also known as equity interchangeably) in machine learning is\nimportant for societal well-being, but limited public datasets hinder its\nprogress. Currently, no dedicated public medical datasets with imaging data for\nfairness learning are available, though minority groups suffer from more health\nissues. To address this gap, we introduce Harvard Glaucoma Fairness\n(Harvard-GF), a retinal nerve disease dataset with both 2D and 3D imaging data\nand balanced racial groups for glaucoma detection. Glaucoma is the leading\ncause of irreversible blindness globally with Blacks having doubled glaucoma\nprevalence than other races. We also propose a fair identity normalization\n(FIN) approach to equalize the feature importance between different identity\ngroups. Our FIN approach is compared with various the-state-of-the-art fairness\nlearning methods with superior performance in the racial, gender, and ethnicity\nfairness tasks with 2D and 3D imaging data, which demonstrate the utilities of\nour dataset Harvard-GF for fairness learning. To facilitate fairness\ncomparisons between different models, we propose an equity-scaled performance\nmeasure, which can be flexibly used to compare all kinds of performance metrics\nin the context of fairness. The dataset and code are publicly accessible via\n\\url{https://ophai.hms.harvard.edu/datasets/harvard-glaucoma-fairness-3300-samples/}.", "title": "Harvard Glaucoma Fairness: A Retinal Nerve Disease Dataset for Fairness Learning and Fair Identity Normalization", "url": "http://arxiv.org/abs/2306.09264v2" }

{ "text_length": 1531 }

{ "abstract": "Previous work on learning physical systems from data has focused on\nhigh-resolution grid-structured measurements. However, real-world knowledge of\nsuch systems (e.g. weather data) relies on sparsely scattered measuring\nstations. In this paper, we introduce a novel simulated benchmark dataset,\nDynaBench, for learning dynamical systems directly from sparsely scattered data\nwithout prior knowledge of the equations. The dataset focuses on predicting the\nevolution of a dynamical system from low-resolution, unstructured measurements.\nWe simulate six different partial differential equations covering a variety of\nphysical systems commonly used in the literature and evaluate several machine\nlearning models, including traditional graph neural networks and point cloud\nprocessing models, with the task of predicting the evolution of the system. The\nproposed benchmark dataset is expected to advance the state of art as an\nout-of-the-box easy-to-use tool for evaluating models in a setting where only\nunstructured low-resolution observations are available. The benchmark is\navailable at https://anonymous.4open.science/r/code-2022-dynabench/.", "title": "DynaBench: A benchmark dataset for learning dynamical systems from low-resolution data", "url": "http://arxiv.org/abs/2306.05805v2" }

{ "text_length": 1261 }

{ "abstract": "As technology grows and evolves rapidly, it is increasingly clear that mobile\ndevices are more commonly used for sensitive matters than ever before. A need\nto authenticate users continuously is sought after as a single-factor or multi\nfactor authentication may only initially validate a user, which does not help\nif an impostor can bypass this initial validation. The field of touch dynamics\nemerges as a clear way to non intrusively collect data about a user and their\nbehaviors in order to develop and make imperative security related decisions in\nreal time. In this paper we present a novel dataset consisting of tracking 25\nusers playing two mobile games Snake.io and Minecraft each for 10 minutes,\nalong with their relevant gesture data. From this data, we ran machine learning\nbinary classifiers namely Random Forest and K Nearest Neighbor to attempt to\nauthenticate whether a sample of a particular users actions were genuine. Our\nstrongest model returned an average accuracy of roughly 93% for both games,\nshowing touch dynamics can differentiate users effectively and is a feasible\nconsideration for authentication schemes. Our dataset can be observed at\nhttps://github.com/zderidder/MC-Snake-Results", "title": "Continuous User Authentication Using Machine Learning and Multi-Finger Mobile Touch Dynamics with a Novel Dataset", "url": "http://arxiv.org/abs/2207.13648v1" }

{ "text_length": 1357 }

{ "abstract": "Machine learning models deployed in healthcare systems face data drawn from\ncontinually evolving environments. However, researchers proposing such models\ntypically evaluate them in a time-agnostic manner, with train and test splits\nsampling patients throughout the entire study period. We introduce the\nEvaluation on Medical Datasets Over Time (EMDOT) framework and Python package,\nwhich evaluates the performance of a model class over time. Across five medical\ndatasets and a variety of models, we compare two training strategies: (1) using\nall historical data, and (2) using a window of the most recent data. We note\nchanges in performance over time, and identify possible explanations for these\nshocks.", "title": "Model Evaluation in Medical Datasets Over Time", "url": "http://arxiv.org/abs/2211.07165v1" }

{ "text_length": 786 }

{ "abstract": "Understanding and interpreting human actions is a long-standing challenge and\na critical indicator of perception in artificial intelligence. However, a few\nimperative components of daily human activities are largely missed in prior\nliterature, including the goal-directed actions, concurrent multi-tasks, and\ncollaborations among multi-agents. We introduce the LEMMA dataset to provide a\nsingle home to address these missing dimensions with meticulously designed\nsettings, wherein the number of tasks and agents varies to highlight different\nlearning objectives. We densely annotate the atomic-actions with human-object\ninteractions to provide ground-truths of the compositionality, scheduling, and\nassignment of daily activities. We further devise challenging compositional\naction recognition and action/task anticipation benchmarks with baseline models\nto measure the capability of compositional action understanding and temporal\nreasoning. We hope this effort would drive the machine vision community to\nexamine goal-directed human activities and further study the task scheduling\nand assignment in the real world.", "title": "LEMMA: A Multi-view Dataset for Learning Multi-agent Multi-task Activities", "url": "http://arxiv.org/abs/2007.15781v1" }

{ "text_length": 1226 }

{ "abstract": "We introduce WikiLingua, a large-scale, multilingual dataset for the\nevaluation of crosslingual abstractive summarization systems. We extract\narticle and summary pairs in 18 languages from WikiHow, a high quality,\ncollaborative resource of how-to guides on a diverse set of topics written by\nhuman authors. We create gold-standard article-summary alignments across\nlanguages by aligning the images that are used to describe each how-to step in\nan article. As a set of baselines for further studies, we evaluate the\nperformance of existing cross-lingual abstractive summarization methods on our\ndataset. We further propose a method for direct crosslingual summarization\n(i.e., without requiring translation at inference time) by leveraging synthetic\ndata and Neural Machine Translation as a pre-training step. Our method\nsignificantly outperforms the baseline approaches, while being more cost\nefficient during inference.", "title": "WikiLingua: A New Benchmark Dataset for Cross-Lingual Abstractive Summarization", "url": "http://arxiv.org/abs/2010.03093v1" }

{ "text_length": 1034 }

{ "abstract": "Physical symptoms caused by high stress commonly happen in our daily lives,\nleading to the importance of stress recognition systems. This study aims to\nimprove stress classification by selecting appropriate features from\nThermal-stress data, ANUstressDB. We explored three different feature selection\ntechniques: correlation analysis, magnitude measure, and genetic algorithm.\nSupport Vector Machine (SVM) and Artificial Neural Network (ANN) models were\ninvolved in measuring these three algorithms. Our result indicates that the\ngenetic algorithm combined with ANNs can improve the prediction accuracy by\n19.1% compared to the baseline. Moreover, the magnitude measure performed best\namong the three feature selection algorithms regarding the balance of\ncomputation time and performance. These findings are likely to improve the\naccuracy of current stress recognition systems.", "title": "Feature Selection on Thermal-stress Dataset", "url": "http://arxiv.org/abs/2109.03755v1" }

{ "text_length": 955 }

{ "abstract": "One of the most promising areas of research to obtain practical advantage is\nQuantum Machine Learning which was born as a result of cross-fertilisation of\nideas between Quantum Computing and Classical Machine Learning. In this paper,\nwe apply Quantum Machine Learning (QML) frameworks to improve binary\nclassification models for noisy datasets which are prevalent in financial\ndatasets. The metric we use for assessing the performance of our quantum\nclassifiers is the area under the receiver operating characteristic curve\n(ROC/AUC). By combining such approaches as hybrid-neural networks, parametric\ncircuits, and data re-uploading we create QML inspired architectures and\nutilise them for the classification of non-convex 2 and 3-dimensional figures.\nAn extensive benchmarking of our new FULL HYBRID classifiers against existing\nquantum and classical classifier models, reveals that our novel models exhibit\nbetter learning characteristics to asymmetrical Gaussian noise in the dataset\ncompared to known quantum classifiers and performs equally well for existing\nclassical classifiers, with a slight improvement over classical results in the\nregion of the high noise.", "title": "Binary classifiers for noisy datasets: a comparative study of existing quantum machine learning frameworks and some new approaches", "url": "http://arxiv.org/abs/2111.03372v1" }

{ "text_length": 1335 }

{ "abstract": "Imperfections in data annotation, known as label noise, are detrimental to\nthe training of machine learning models and have an often-overlooked\nconfounding effect on the assessment of model performance. Nevertheless,\nemploying experts to remove label noise by fully re-annotating large datasets\nis infeasible in resource-constrained settings, such as healthcare. This work\nadvocates for a data-driven approach to prioritising samples for re-annotation\n- which we term \"active label cleaning\". We propose to rank instances according\nto estimated label correctness and labelling difficulty of each sample, and\nintroduce a simulation framework to evaluate relabelling efficacy. Our\nexperiments on natural images and on a new medical imaging benchmark show that\ncleaning noisy labels mitigates their negative impact on model training,\nevaluation, and selection. Crucially, the proposed active label cleaning\nenables correcting labels up to 4 times more effectively than typical random\nselection in realistic conditions, making better use of experts' valuable time\nfor improving dataset quality.", "title": "Active label cleaning for improved dataset quality under resource constraints", "url": "http://arxiv.org/abs/2109.00574v2" }

{ "text_length": 1202 }

{ "abstract": "With increasingly more data and computation involved in their training,\nmachine learning models constitute valuable intellectual property. This has\nspurred interest in model stealing, which is made more practical by advances in\nlearning with partial, little, or no supervision. Existing defenses focus on\ninserting unique watermarks in a model's decision surface, but this is\ninsufficient: the watermarks are not sampled from the training distribution and\nthus are not always preserved during model stealing. In this paper, we make the\nkey observation that knowledge contained in the stolen model's training set is\nwhat is common to all stolen copies. The adversary's goal, irrespective of the\nattack employed, is always to extract this knowledge or its by-products. This\ngives the original model's owner a strong advantage over the adversary: model\nowners have access to the original training data. We thus introduce $dataset$\n$inference$, the process of identifying whether a suspected model copy has\nprivate knowledge from the original model's dataset, as a defense against model\nstealing. We develop an approach for dataset inference that combines\nstatistical testing with the ability to estimate the distance of multiple data\npoints to the decision boundary. Our experiments on CIFAR10, SVHN, CIFAR100 and\nImageNet show that model owners can claim with confidence greater than 99% that\ntheir model (or dataset as a matter of fact) was stolen, despite only exposing\n50 of the stolen model's training points. Dataset inference defends against\nstate-of-the-art attacks even when the adversary is adaptive. Unlike prior\nwork, it does not require retraining or overfitting the defended model.", "title": "Dataset Inference: Ownership Resolution in Machine Learning", "url": "http://arxiv.org/abs/2104.10706v1" }

{ "text_length": 1786 }

{ "abstract": "Continuity of care is crucial to ensuring positive health outcomes for\npatients discharged from an inpatient hospital setting, and improved\ninformation sharing can help. To share information, caregivers write discharge\nnotes containing action items to share with patients and their future\ncaregivers, but these action items are easily lost due to the lengthiness of\nthe documents. In this work, we describe our creation of a dataset of clinical\naction items annotated over MIMIC-III, the largest publicly available dataset\nof real clinical notes. This dataset, which we call CLIP, is annotated by\nphysicians and covers 718 documents representing 100K sentences. We describe\nthe task of extracting the action items from these documents as multi-aspect\nextractive summarization, with each aspect representing a type of action to be\ntaken. We evaluate several machine learning models on this task, and show that\nthe best models exploit in-domain language model pre-training on 59K\nunannotated documents, and incorporate context from neighboring sentences. We\nalso propose an approach to pre-training data selection that allows us to\nexplore the trade-off between size and domain-specificity of pre-training\ndatasets for this task.", "title": "CLIP: A Dataset for Extracting Action Items for Physicians from Hospital Discharge Notes", "url": "http://arxiv.org/abs/2106.02524v1" }

{ "text_length": 1350 }

{ "abstract": "It is very important to access a rich music dataset that is useful in a wide\nvariety of applications. Currently, available datasets are mostly focused on\nstoring vocal or instrumental recording data and ignoring the requirement of\nits visual representation and retrieval. This paper attempts to build an\nXML-based public dataset, called SANGEET, that stores comprehensive information\nof Hindustani Sangeet (North Indian Classical Music) compositions written by\nfamous musicologist Pt. Vishnu Narayan Bhatkhande. SANGEET preserves all the\nrequired information of any given composition including metadata, structural,\nnotational, rhythmic, and melodic information in a standardized way for easy\nand efficient storage and extraction of musical information. The dataset is\nintended to provide the ground truth information for music information research\ntasks, thereby supporting several data-driven analysis from a machine learning\nperspective. We present the usefulness of the dataset by demonstrating its\napplication on music information retrieval using XQuery, visualization through\nOmenad rendering system. Finally, we propose approaches to transform the\ndataset for performing statistical and machine learning tasks for a better\nunderstanding of Hindustani Sangeet. The dataset can be found at\nhttps://github.com/cmisra/Sangeet.", "title": "SANGEET: A XML based Open Dataset for Research in Hindustani Sangeet", "url": "http://arxiv.org/abs/2306.04148v1" }

{ "text_length": 1432 }

{ "abstract": "The data article describes the Road Damage Dataset, RDD2022, which comprises\n47,420 road images from six countries, Japan, India, the Czech Republic,\nNorway, the United States, and China. The images have been annotated with more\nthan 55,000 instances of road damage. Four types of road damage, namely\nlongitudinal cracks, transverse cracks, alligator cracks, and potholes, are\ncaptured in the dataset. The annotated dataset is envisioned for developing\ndeep learning-based methods to detect and classify road damage automatically.\nThe dataset has been released as a part of the Crowd sensing-based Road Damage\nDetection Challenge (CRDDC2022). The challenge CRDDC2022 invites researchers\nfrom across the globe to propose solutions for automatic road damage detection\nin multiple countries. The municipalities and road agencies may utilize the\nRDD2022 dataset, and the models trained using RDD2022 for low-cost automatic\nmonitoring of road conditions. Further, computer vision and machine learning\nresearchers may use the dataset to benchmark the performance of different\nalgorithms for other image-based applications of the same type (classification,\nobject detection, etc.).", "title": "RDD2022: A multi-national image dataset for automatic Road Damage Detection", "url": "http://arxiv.org/abs/2209.08538v1" }

{ "text_length": 1284 }