id
int64 1
213k
| paper_arxiv_id
stringlengths 10
13
| path
float64 | caption
stringlengths 9
127k
⌀ | label
stringlengths 7
125
⌀ | table_text
stringlengths 33
131k
⌀ |
|---|---|---|---|---|---|
212,187
|
2310.17230v1
| null |
\caption{Example Code Activations for the \textbf{TokFSM} dataset. The \bb{bolded} digits indicate the token positions that activated the given code. Hyphens (-) are added between each state for readability but are not presented to the model. MLP codes are written in the form \texttt{layer.code-id}, while attention codes are written in the form \texttt{layer.head.code-id}. More activations are available at \ifdefined\DOUBLEBLIND [redacted for anonymity] \else \url{https://huggingface.co/spaces/taufeeque/codebook-features} \fi.}
|
\label{tab:activations-fsm}
|
\begin{tabular}{p{2cm}p{3.5cm}p{7cm}}
\toprule
\textbf{Code} & \textbf{Interpretation} & \textbf{Example Activations} \\
\midrule
MLP 0.2523 & \bb{1} digit & 3\bb{1}-83-40-87-80-78-38-76-03-86-\bb{1}7-97-76-09-\bb{1}5 \\
& & \bb{1}0-57-62-43-92-3\bb{1}-83-82-23-65-94-33-23-49-4\bb{1} \\
& & \bb{1}9-83-3\bb{1}-73-29-47-04-\bb{1}5-77-05-79-23-47-89-95 \\
\addlinespace
MLP 1.2527 & 48\bb{9} trigram (either pos.) & 86-04-8\bb{9}-80-17-03-40-74-24-09-93-35-59-61-49 \\
& & 40-46-50-38-47-04-8\bb{9}-80-91-82-94-33-41-77-59 \\
& & 18-94-55-55-48-24-68-48-\bb{9}0-43-97-50-74-77-59 \\
\addlinespace
MLP 2.2543 & 4\bb{0} bigram (2nd pos.) & 80-04-70-50-4\bb{0}-59-07-73-28-02-71-54-31-62-40 \\
& & 74-05-13-72-95-66-52-31-98-20-88-4\bb{0}-59-22-19 \\
& & 4\bb{0}-46-44-01-88-66-51-14-41-57-18-84-89-60-51 \\
\addlinespace
Attn 1.2.3207 & Tokens after 44 bigram& 44-\bb{2}7-74-05-59-64-67-72-42-93-35-09-67-39-96 \\
& & 44-\bb{2}7-74-05-22-65-98-75-83-20-00-60-80-57-94 \\
& & \bb{7}7-69-28-02-34-46-52-72-94-18-84-12-16-64-4\bb{6} \\
\addlinespace
Attn 2.0.3044 & Tokens on or after 59 & 74-05-5\bb{9}-64-67-72-42-93-35-09-67-39-96-07-96 \\
& & 88-40-5\bb{9}-\bb{2}2-19-33-31-93-42-53-75-94-33-31-76 \\
& & 87-14-40-59-\bb{2}4-72-86-04-30-04-81-56-01-17-30 \\
\bottomrule
\end{tabular}
|
212,188
|
2310.17230v1
| null |
\caption{\textbf{Number of active codes in $k=1$ attention + MLP codebook model trained on TokFSM}. Each codebook has 10,000 codes; most of the codes in each codebook are not active by the end of training.}
|
\label{tab:fsm-active-codes}
|
\begin{tabular}{lrrrrr}
\toprule
Layer & Head 0 & Head 1 & Head 2 & Head 3 & MLP \\
\midrule
0 & 40 & 45 & 41 & 49 & 11 \\
1 & 293 & 367 & 657 & 460 & 1027 \\
2 & 1482 & 3071 & 1103 & 1499 & 943 \\
3 & 690 & 282 & 315 & 1233 & 247 \\
\bottomrule
\end{tabular}
|
212,189
|
2310.17230v1
| null |
\caption{\textbf{Ablation studies.} Next-token accuracy (for TinyStories and WikiText-103) and next-state transition accuracies (for TokFSM) across various ablation studies. \textbf{\textit{Legend}}: \textbf{Attn CB}: Codebook applied to the attention layers. \textbf{Random Init}: Codebooks applied to a randomly-initialized model instead of a pretrained model (then finetuned end-to-end as usual). \textbf{Train Only CB}: Train only the codebook layers with the original loss while keeping the base model frozen. \textbf{Only AE Loss}: Only apply the autoencoding loss to the codebooks; do not update the model parameters. \textbf{Attn + MLP CB} Codebooks applied to the attention and MLP codebooks simultaneously.}
|
\label{tab:lm-ablations}
|
\begin{tabular}{@{}lccccc@{}}
\toprule
\textbf{Model} & \textbf{Attn CB} & \textbf{Random Init} & \textbf{Train Only CB} & \textbf{Only AE Loss} \\
\midrule
\textbf{TinyStories-1L} & \textbf{57.91} & 55.67 & 47.08 & 51.73 \\
\textbf{FSM-4L} & \textbf{96.39} & 52.35 & 58.48 & 43.44 \\
\textbf{WikiText-103-24L} & \textbf{46.16} & 38.53 & 31.22 & 28.35 \\
\bottomrule
\end{tabular}
|
212,190
|
2310.17230v1
| null |
\caption{410m Parameters}
|
\label{tab:runtime-cb-attn}
|
\begin{tabular}{@{}lccc@{}}
\toprule
Model & Tok/s & $\Delta$ FAISS & $\Delta$ Base \\
\midrule
Base & 14.8 & & \\
CB w/ FAISS & 7.2 & 56.2\% & -51.5\% \\
CB no FAISS & 4.6 & & -68.9\% \\
\bottomrule
\end{tabular}
|
212,191
|
2310.17230v1
| null |
\caption{70m Parameters}
| null |
\begin{tabular}{@{}lccc@{}}
\toprule
Model & Tok/s & $\Delta$ FAISS & $\Delta$ Base \\
\midrule
Base & 57.5 & & \\
CB w/ FAISS & 37.4 & 34.2\% & -34.9\% \\
CB no FAISS & 27.9 & & -51.5\% \\
\bottomrule
\end{tabular}
|
212,192
|
2310.17230v1
| null |
\caption{410m Parameters}
| null |
\begin{tabular}{@{}lccc@{}}
\toprule
Model & Tok/s & $\Delta$ FAISS & $\Delta$ Base \\
\midrule
Base & 14.8 & & \\
CB w/ FAISS & 7.2 & 56.2\% & -51.5\% \\
CB no FAISS & 4.6 & & -68.9\% \\
\bottomrule
\end{tabular}
|
212,193
|
2310.17230v1
| null |
\caption{Example generations from language models. The prompts are highlighted in bold. While the factuality of the completions is unreliable for all models, all models generate largely grammatical text.}
|
\label{tab:lm_generations}
|
\begin{tabular}{p{0.10\textwidth}p{0.4\textwidth}p{0.4\textwidth}}
\toprule
\textbf{Language Model} & \textbf{TinyStories 1-Layer Model} & \textbf{WikiText-103 Model} \\
\midrule
Base & \textbf{Once upon a time} there was a little boy named Timmy. Timmy loved to play outside in the rain. He would jump in puddles and splash around. One day, Timmy saw a big puddle in the park. He jumped in it and got all wet.[...]
& \textbf{The war was fought} against the Ottoman Empire and the Kingdom of Hungary. The Ottoman Turks, their king, and several of their princes were killed and many more captured, and the kingdom was divided among the Hungarian monarchs ; [...]
\\
\addlinespace
Codebooks (Attn) & \textbf{Once upon a time}, there was a little girl named Lily. She loved to play with her toys and her friends. One day, Lily's mom told her that they were going to buy a new toy. Lily was very excited and asked, ``Can I play with your toys, please?"[...]
& \textbf{The war was fought} by France and the British Empire, and by the Axis powers. With the exception of the Italians and Americans, whose armies won the war against the Axis Powers, the victorious Allies suffered the most of the war, a terrible defeat on both fronts. [...]
\\
\addlinespace
Codebooks (MLP) & \textbf{Once upon a time}, there was a little boy named Timmy. Timmy loved to play with his toy cars and trucks. One day, Timmy's mom took him to the store to buy a new toy. Timmy saw a big red truck and asked his mommy if they could get it, but she said they had to wait until they got to the store.
& \textbf{The war was fought} between the United States and France. The French responded by launching an invasion of the Allied continent in June 1917 with the aim of defeating the Allied armies in northern France. [...]
\\
\bottomrule
\end{tabular}
|
212,194
|
2310.17230v1
| null |
\caption{TinyStories}
|
\label{tab:lm-activations}
|
\begin{tabular}{p{1.2cm}p{2.5cm}p{9cm}}
\toprule
\textbf{Code} & \textbf{Interpretation} & \textbf{Example Activations} \\
\midrule
0.2 & Fighting & The two cats started to \bb{quarrel loudly} over the bone\\
& & They ran around the house, fighting over \bb{the} thread \\
& & But then, they got into a fight \bb{over} who got to play with the toy \\
\addlinespace
0.3 & Negative emotions & He feels angry and \bb{scared}. He tries to catch the boat, but it \\
& & She started to feel \bb{nervous} because she thought she wouldn't be able to \\
& & Lily and Tom felt \bb{fearful}. They did not like storms. \\
\addlinespace
0.6 & ``You'' dialogue & The dragon smiled and said, "\bb{You are} too small. It's not possible." \\
& & The happy fish thanked her and said "\bb{You must} be very persistent to complete this task. \\
& & John smiled and said, "\bb{You won}! You were really fast."
\\
\addlinespace
1.2 & Fire & The fire \bb{spread to} the \bb{cans and bottles and} made more explosions. The garage was full of smoke \\
& & Lily knew that fire could be dangerous and she \bb{always} remembered to be careful \bb{when} playing with matches or \bb{li}ghters. \\
& & Mom hugged them and said, "I know, but \bb{fire} is \bb{not a} toy. \bb{It can hurt you and} the plants \bb{and} animals. \\
\addlinespace
5.3 & Discovered/found & Lily found a delicate flower in the garden and \bb{showed} it to her sister.\\
& & had discovered an amazing reef and helped a turtle in \bb{need}. \\
& & One day, Tom and Mia found a ball in \bb{the} hut. \\
\bottomrule
\end{tabular}
|
212,206
|
2103.07601v3
| null |
\caption{Table \ref{tab:maintable} with 2 decimal results. Section \ref{sec:additionaltables}}
|
\label{tab:maintable2dec}
|
\begin{tabular}{ccccccccc}
\hline
Task & $\xi (\alpha, \beta^{\un})$ & $\xi(\beta^{\un}, \beta^{\prx})$ & $\xi(\Delta, \beta^{\un})$ & $\xi(\alpha, \beta)$ & $\xi^{*}$ \\
\hline
IMDB & 86.81 & 88.88\phantom{*} & 86.81 & 86.81 & 89.55 \\
Yelp & 90.39 & 95.22\phantom{*} & 95.31 & 93.59 & 96.20 \\
AG News & 93.54 & 96.05$^{*}$ & 94.32 & 94.50 & 96.05 \\
20 NG & 91.16 & 60.75\phantom{*} & 84.57 & 84.57 & 94.22 \\
SST & 78.16 & 83.53$^{*}$ & 78.16 & 82.38 & 83.53 \\
Amzn & 82.48 & 90.38$^{*}$ & 82.48 & 88.07 & 90.38 \\
\hline
Muti30k & 43.45 & 43.45\phantom{*} & 43.45 & 48.58 & 66.29 \\
IWSLT14 & 35.82 & 35.82\phantom{*} & 32.98 & 44.09 & 67.36 \\
News It-Pt & 21.82 & 25.06\phantom{*} & 21.82 & 25.06 & 55.41 \\
News En-Nl & 18.59 & 23.21\phantom{*} & 18.59 & 26.79 & 62.90 \\
News En-Pt & 19.12 & 19.12\phantom{*} & 19.12 & 27.85 & 67.75 \\
\hline
\end{tabular}
|
212,195
|
2310.17230v1
| null |
\caption{WikiText-103}
| null |
\begin{tabular}{p{1.2cm}p{2.5cm}p{9cm}}
\toprule
\textbf{Code} & \textbf{Interpretation} & \textbf{Example Activations} \\
\midrule
7.12.7884 & Months (after preposition) & at Toulon in \bb{August} The ship began trials [\ldots] and spent three weeks in \bb{September} attached to \\
& & 14 : 30 on 7 \bb{December}. The division had the [\ldots] a major attack until 8 \bb{December} \\
& & on \bb{August} 31, a Utah [\ldots] On \bb{September} 1, 1987 \\
\addlinespace
4.15.6101 & Evaluative words & Initially , the New Zealand attack progressed \bb{well} \\
& & Superman from the main timeline is \bb{successfully} teleported into \\
& & only HWMs evaluated as "\bb{excellent}" are used by NHC \\
\addlinespace
1.9.295 & Names starting with `B` & In one account from the Bah\bb{amas} , a mating pair ascended \\
& & while John and Roy B\bb{oulting} noted that [...] \\
& & B\bb{ocks}car, sometimes called B\bb{ock}'s Car, is the name of the United States Army Air Forces B\bb{-29} bomber \\
\addlinespace
4.14.4742 & Years in 2000s & As of \bb{2011} , the International Shark Attack File lists \\
& & In \bb{2014} , a study at the University of Amsterdam with \\
& & Fabian Cancellara kicked off his \bb{2010} campaign with an overall victory at the Tour of\\
\addlinespace
9.3.3727 & Square Units & Atlanta encompasses 134.0 \bb{square miles} (347.\bb{1}km\bb{2}) \\
& & it covered more than 55 square metres (590 \bb{sq} ft) \\
& & 6 percent or 101,593 \bb{square} kilometres (39,\bb{225 sq} mi) of [...] \\
\bottomrule
\end{tabular}
|
212,196
|
2310.17230v1
| null |
\caption{TinyStories}
| null |
\begin{tabular}{p{1.2cm}p{2.5cm}p{9cm}}
\toprule
\textbf{Code} & \textbf{Interpretation} & \textbf{Example Activations} \\
\midrule
0.2 & Fighting & The two cats started to \bb{quarrel loudly} over the bone\\
& & They ran around the house, fighting over \bb{the} thread \\
& & But then, they got into a fight \bb{over} who got to play with the toy \\
\addlinespace
0.3 & Negative emotions & He feels angry and \bb{scared}. He tries to catch the boat, but it \\
& & She started to feel \bb{nervous} because she thought she wouldn't be able to \\
& & Lily and Tom felt \bb{fearful}. They did not like storms. \\
\addlinespace
0.6 & ``You'' dialogue & The dragon smiled and said, "\bb{You are} too small. It's not possible." \\
& & The happy fish thanked her and said "\bb{You must} be very persistent to complete this task. \\
& & John smiled and said, "\bb{You won}! You were really fast."
\\
\addlinespace
1.2 & Fire & The fire \bb{spread to} the \bb{cans and bottles and} made more explosions. The garage was full of smoke \\
& & Lily knew that fire could be dangerous and she \bb{always} remembered to be careful \bb{when} playing with matches or \bb{li}ghters. \\
& & Mom hugged them and said, "I know, but \bb{fire} is \bb{not a} toy. \bb{It can hurt you and} the plants \bb{and} animals. \\
\addlinespace
5.3 & Discovered/found & Lily found a delicate flower in the garden and \bb{showed} it to her sister.\\
& & had discovered an amazing reef and helped a turtle in \bb{need}. \\
& & One day, Tom and Mia found a ball in \bb{the} hut. \\
\bottomrule
\end{tabular}
|
212,197
|
2310.17230v1
| null |
\caption{TinyStories}
|
\label{tab:topic-steering-regex}
|
\begin{tabular}{@{}ll@{}}
\toprule
\textbf{Topic} & \textbf{Regex} \\
\midrule
Dragon & dragon \\
Slide & slide \\
Friend & friend \\
Tom \& Sam & tom$\mid$ sam \\
Flower & flower \\
Fire & fire \\
Baby & baby \\
Princess & prince$\mid$ crown$\mid$ king$\mid$ castle \\
\bottomrule
\end{tabular}
|
212,198
|
2310.17230v1
| null |
\caption{Wikitext}
| null |
\begin{tabular}{@{}ll@{}}
\toprule
\textbf{Topic} & \textbf{Regex} \\
\midrule
Football & football$\mid$ soccer$\mid$ goal$\mid$ stadium$\mid$ fifa$\mid$ player$\mid$ trophy$\mid$ league \\
Movie & movie$\mid$ tv$\mid$ television$\mid$ film$\mid$ media \\
Video Game & game \\
Song & song$\mid$ music$\mid$ mtv \\
\bottomrule
\end{tabular}
|
212,199
|
2310.17230v1
| null |
\caption{TinyStories}
| null |
\begin{tabular}{@{}ll@{}}
\toprule
\textbf{Topic} & \textbf{Regex} \\
\midrule
Dragon & dragon \\
Slide & slide \\
Friend & friend \\
Tom \& Sam & tom$\mid$ sam \\
Flower & flower \\
Fire & fire \\
Baby & baby \\
Princess & prince$\mid$ crown$\mid$ king$\mid$ castle \\
\bottomrule
\end{tabular}
|
212,200
|
2310.17230v1
| null |
\caption{Steering the TinyStories language model with topic codes. While the language model generations are often incoherent, the topics are largely introduced in grammatical ways}
|
\label{tab:lm-interventions-ts-full}
|
\begin{tabular}{p{1.8cm}C{0.7cm}p{10cm}}
\toprule
\textbf{Topic} & \textbf{Codes} & \textbf{Steered generations} \\
\midrule
Null (default generation) & - & \textbf{Once upon a time,} there was a little girl named Lily. She loved to play with her toys and eat yummy food. One day, she went to the store with her mommy to buy some apples. She saw a big basket and wanted to buy it. Her mommy said, "Let's count the apples and buy the yummiest apples!" Lily was so happy and said, "Yes, please!" \\
Dragon & 8 & \textbf{Once upon a time,} there was a little girl named Lily. She was very excited to go outside and explore. She flew over the trees and saw a big, scary dragon. The dragon was very scary. But Lily knew that things were not real and she would hurt her. \\
Slide & 12 & \textbf{Once upon a time,} there was a big cake. He wanted to escape and explore the world, but he was scared. He thought and thought about how to use the slide. He asked his mom for some help. \\
Friend & 9 & \textbf{Once upon a time,} there was a little girl named Lily. She had a friend named Sue. Sue was a nosy cat. She liked to play with her brother Tom. One day, they went to the park and saw a big dog. The dog was not nice. \\
Flower & 8 & \textbf{Once upon a time,} there was a little girl named Lily. She liked to pick flowers in the meadow. One day, she saw a big, green caterpillar on a leaf. She wanted to take it home and sell it to someone else. As she picked the flower, it started to bloom and made it look pretty. \\
Fire & 16 & \textbf{Once upon a time,} there was a little boy named Timmy. Timmy loved his new toy. He always felt like a real fireman. One day, Timmy's mom made him some hot soup and gave him some medicine to help his mommy feel better. Timmy was scared that the fire would be gone, but he didn't feel happy. \\
Baby & 15 & \textbf{Once upon a time,} there was a little girl named Lily. She loved going to the gym with her mommy. One day, Lily's mom asked her to help put the baby in the crib. \\
Princess & 14 & \textbf{Once upon a time,} there was a little bird named Tweety. One day, the princess had a dream that she was invited to a big castle. She was very excited and said, ``I want to be a princess and ride the big, pretty castle!" \\
\bottomrule
\end{tabular}
|
212,207
|
2103.07601v3
| null |
\caption{translation results from Table \ref{tab:maintable2dec} except with performance measured by bleu rather than token accuracy. Section \ref{sec:additionaltables}}
|
\label{tab:maintable2decbleu}
|
\begin{tabular}{ccccccccc}
\hline
Task & $\xi (\alpha, \beta^{\un})$ & $\xi (\beta^{\un}, \beta^{\prx})$ & $\xi(\Delta, \beta^{\un})$ & $\xi(\alpha, \beta)$ & $\xi^{*}$ \\
\hline
Muti30k & 11.43 & 11.43 & 11.43 & 16.41 & 37.89 \\
IWSLT14 & \phantom{0}6.71 & \phantom{0}6.71 & \phantom{0}5.31 & 9.89 & 32.95 \\
News It-Pt & \phantom{0}1.29 & \phantom{0}2.16 & \phantom{0}1.29 & 2.16 & 24.71 \\
News En-Nl & \phantom{0}0.94 & \phantom{0}2.39 & \phantom{0}0.94 & 4.12 & 29.42 \\
News En-Pt & \phantom{0}0.74 & \phantom{0}0.74 & \phantom{0}0.74 & 4.28 & 37.04 \\
\hline
\end{tabular}
|
212,293
|
2303.14969v1
| null |
\caption{10-shot learning performance of Ours with different number of attention heads in Matching module.}
|
\label{tab:attention_heads}
|
\begin{tabular}{c|cc}
\toprule
\multirow{4}{*}{Number of Attention Heads} &
\multicolumn{2}{c}{Tasks} \\
\cmidrule{2-3}
&
\multicolumn{2}{c}{Fold 1} \\
\cmidrule{2-3}
&
SS & SN \\
&
mIoU ↑ & mErr ↓ \\
\midrule
1 &
0.3702 & 12.5936 \\
4 &
\textbf{0.4097} & \textbf{11.4391} \\
\bottomrule
\end{tabular}
|
212,201
|
2310.17230v1
| null |
\caption{\textbf{Steering language model generations with topic codes}.
Full table available at \ifdefined\DOUBLEBLIND [redacted for anonymity] \else \href{https://docs.google.com/spreadsheets/d/10gbVwSFe_G1josj51n4J9iCu1vnILJSHx05SA7Oln8w/edit?usp=sharing}{link} \fi.
At symbol (@) delimiters present in WikiText-103 data have been omitted for readability.}
|
\label{tab:lm-interventions-wikitext}
|
\begin{tabular}{p{1cm}p{0.6cm}p{5.4cm}p{5.4cm}}
\toprule
\textbf{Topic} & \textbf{Codes} & \textbf{Original generations} & \textbf{Steered generations} \\
\midrule
Video game & 18
& \textbf{The war was fought} on two fronts. The war was initiated in 1914 between Austria-Hungary and Serbia, when the Entente Powers signed a treaty of friendship between the two countries. In October 1914, Tschichky was sent to defend the German Empire'
& \textbf{The war was fought} on both sides, and was only the second game to deal with one-on-one battles, following SimCity 2D Blade II. The game was released to critical acclaim, with praise particularly directed to the new console \\
% \hline
Football & 18
& \textbf{The war was fought} on two fronts. The war was initiated in 1914 between Austria-Hungary and Serbia, when the Entente Powers signed a treaty of friendship between the two countries. In October 1914, Tschichky was sent to defend the German Empire'
& \textbf{The war was fought} in its first forty years. In the summer of 1946, the Cardinals of the All-America Football Conference (AAFC) were rapidly becoming the favorites for NFL Hall-of-Fame coach Jim Mora, who had \\
% \hline
Movie & 12
& \textbf{The novel was published in} November 2009 by MacChinnacle, a London publishing house. The book's publishers, Syco, published the book in the United Kingdom and the United States on 1 November 2009. The book received generally positive reviews from critics, who praised the
& \textbf{The novel was published in} the United States and Canada. The film was directed by Joe Hahn and stars Steven Spielberg as Lucas, Neil Patrick Harris, and Jude Lawder as Lucas's best friend, Jonathan Miller. The plot follows a character (Lucas \\
% \hline
Song & 17
& \textbf{The team won} their first ever Grand Prix and the first since the 1990 season. The team finished in third place behind Williams and Ralf Schumacher, with the Ferraris of David Coulthard and Jarno Trulli finishing in the top three.
& \textbf{The team won} the Grammy Awards for Best Gospel Album.
= = Background = =
In 2004, The Dream released their third studio album, The Beacon Street Collection, which produced the singles "HOV Lane" and "Wishing Machine \\
\bottomrule
\end{tabular}
|
212,208
|
2103.07601v3
| null |
\caption{Table \ref{tab:maintable} except with correlations and performance metrics taken over the training set instead of the validation set. Section \ref{sec:additionaltables}}
|
\label{tab:maintabletrain}
|
\begin{tabular}{ccccccccc}
\hline
Task & $\xi (\alpha, \beta^{\un})$ & $\xi (\beta^{\un}, \beta^{\prx})$ & $\xi(\Delta, \beta^{\un})$ & $\xi(\alpha, \beta)$ & $\xi^{*}$ \\
\hline
IMDB & \phantom{0}90.40 & \phantom{0}99.95$^{*}$ & \phantom{0}90.40 & \phantom{0}95.01 & \phantom{0}99.95 \\
Yelp & \phantom{0}75.61 & \phantom{0}96.54\phantom{*} & \phantom{0}96.19 & \phantom{0}94.44 & \phantom{0}98.22 \\
AG News & \phantom{0}93.57 & \phantom{0}98.42$^{*}$ & \phantom{0}94.63 & \phantom{0}95.54 & \phantom{0}98.42 \\
20 NG & 100.00 & 65.40 & 100.00 & \phantom{0}100.0 & 100.00 \\
SST & \phantom{0}97.72 & 100.00$^{*}$ & \phantom{0}84.11 & 100.0$^{*}$ & 100.00 \\
Amzn & \phantom{0}87.96 & \phantom{0}99.58$^{*}$ & \phantom{0}80.98 & \phantom{0}91.09 & \phantom{0}99.58 \\
\hline
Muti30k & \phantom{0}43.27 & \phantom{0}43.27\phantom{*} & \phantom{0}43.27 & 51.97 & \phantom{0}80.76 \\
IWSLT14 & \phantom{0}35.94 & \phantom{0}35.94\phantom{*} & \phantom{0}35.94 & 44.18 & \phantom{0}71.18 \\
News It-Pt & \phantom{0}22.69 & \phantom{0}25.96\phantom{*} & \phantom{0}22.69 & 39.98 & \phantom{0}77.10 \\
News En-Nl & \phantom{0}18.85 & \phantom{0}23.56\phantom{*} & \phantom{0}18.85 & 40.09 & \phantom{0}74.49 \\
News En-Pt & \phantom{0}19.33 & \phantom{0}19.33\phantom{*} & \phantom{0}19.33 & 42.41 & \phantom{0}77.97 \\
\hline
\end{tabular}
|
212,209
|
2103.07601v3
| null |
\caption{translation results from Table Table \ref{tab:maintabletrain} except with performance measured by bleu rather than token accuracy. Section \ref{sec:additionaltables}}
|
\label{tab:maintabletrainbleu}
|
\begin{tabular}{ccccccccc}
\hline
Task & $\xi (\alpha, \beta^{\un})$ & $\xi (\beta^{\un}, \beta^{\prx})$ & $\xi(\Delta, \beta^{\un})$ & $\xi^{*}$ \\
\hline
Muti30k & 11.87 & 11.87 & 11.87 & 52.28 \\
IWSLT14 & \phantom{0}6.82 & \phantom{0}6.82 & \phantom{0}6.82 & 36.23 \\
News It-Pt & \phantom{0}1.30 & \phantom{0}2.30 & \phantom{0}1.30 & 42.40 \\
News En-Nl & \phantom{0}1.11 & \phantom{0}2.29 & \phantom{0}1.11 & 39.40 \\
News En-Pt & \phantom{0}0.83 & \phantom{0}0.83 & \phantom{0}0.83 & 46.57 \\
\hline
\end{tabular}
|
212,210
|
2103.07601v3
| null |
\caption{statistics for each dataset. Median sequence length in the training set and train set size. Note: src refers to the input "source" sequence, and trg refers to the output "target" sequence. Section \ref{sec:datasetsummarize}}
|
\label{tab:datasetsize}
|
\begin{tabular}{ccc}
Data & vocab size \\
\hline
IMDB & 60338 \\
AG News & 31065 \\
NewsG & 31065 \\
SST & 11022 \\
Amzn & 37110 \\
Yelp & 41368 \\
\hline
IWSLT14 & \phantom{0}8000 \\
Multi-30k & \phantom{0}8000 \\
News-en-nl & \phantom{0}8000 \\
News-en-pt & \phantom{0}8000 \\
News-it-pt & \phantom{0}8000 \\
\hline
\end{tabular}
|
212,211
|
2103.07601v3
| null |
\caption{We report the correlation between $\alpha^{\fr}$ and $\alpha$ on classification datasets, and compare it against $\ag(\alpha, \beta^{\un})$, the same column defined in Table \ref{tab:maintable}. Section \ref{sec:frozenbeta}}
|
\label{tab:frozen-beta-correlation}
|
\begin{tabular}{ccc}
Dataset & $\ag(\alpha, \alpha^{\fr})$ & $\ag(\alpha, \beta^{\un})$ \\
\hline
IMDB & \phantom{0}9 & 53\\
AG News & 17 & 39\\
20 NG & 19 & 65\\
SST & 14 & 20\\
Amzn & 15 & 52\\
Yelp & \phantom{0}8 & 18\\
\hline
\end{tabular}
|
212,212
|
2103.07601v3
| null |
\caption{we report $\ag(\beta^{\un}, \beta^{\logr})$ to demonstrate its effective equivalence to $\ag(\beta^{\un}, \beta^{\prx})$. These values are not exactly the same due to differences in regularization strategies.}
|
\label{tab:logregtab}
|
\begin{tabular}{ccc}
Task & $\ag(\beta^{\un}, \beta^{\prx})$ & $\ag(\beta^{\un}, \beta^{\logr})$ \\
\hline
IMDB & 0.81 & 0.84 \\
Yelp & 0.74 & 0.76 \\
AG News & 0.57 & 0.58 \\
20 NG & 0.40 & 0.45 \\
SST & 0.39 & 0.46 \\
Amzn & 0.53 & 0.60 \\
\hline
\end{tabular}
|
212,213
|
2208.01545v1
| null |
\caption{
\textbf{
The diversity coefficient of MiniImagenet (MI) and Cifar-fs is low.}
The diversity coefficient was computed using the cosine distance between different standard 5-way, 20-shot classification tasks from the few-shot learning benchmark using the Task2Vec method described in section \ref{def_div}.
We used 20 shots (number of examples per class) since we can use the whole task data to compute the diversity coefficient (no splitting of support and query set required for the diversity coefficient).
We used Resnet18 and Resnet34 networks as probe networks -- both pre-trained on ImageNet (indicated as ``pt" on table) and randomly initialized (indicated as ``rand" on table).
We observe that both type of networks and weights give similar diversity results.
All confidence intervals were at 95\%.
To compute results, we used 500 few-shot learning tasks and only compare pairs of different tasks.
This results in $(500^2 - 500)/2 = 124,750$ pair-wise distances used to compute the diversity coefficient. % (500**2-500)/2 = 124,750
}
|
\label{div_mi_table}
|
\begin{tabular}{lll}
\toprule
Probe Network & Diversity on MI & Diversity on Cifar-fs \\
\midrule
Resnet18 (pt) & 0.117 $\pm$ 2.098e-5 & 0.100 $\pm$ 2.18e-5 \\
Resnet18 (rand) & 0.0955 $\pm$ 1.29e-5 & 0.103 $\pm$ 1.05e-5 \\
Resnet34 (pt) & 0.0999 $\pm$ 1.95e-5 & 0.0847 $\pm$ 3.06e-5 \\
Resnet34 (rand) & 0.0620 $\pm$ 8.12e-6 & 0.0643 $\pm$ 9.64e-6 \\
\bottomrule
\end{tabular}
|
212,214
|
2208.01545v1
| null |
\caption{
\textbf{
MAML trained representations and supervised trained representation have statistically equivalent meta-test accuracy on MiniImagenet -- which has low diversity.}
% More precisely, a model trained with 5 inner steps (MAML5) when adapted with 5 inner steps using MAL has meta-test accuracy to a transfer model trained with union supervise learning (USL) that adapted only the final layer during inference as in \citep{rfs}.
The transfer model's adaptation is labeled as ``Adapted Head only (with LR)" -- which stands for ``Logistic Regression (LR)" used in \citep{rfs}.
More precisely, we used Logistic Regression (LR) with LBFGS with the default value for the l2 regularization parameter given by Python's Sklearn.
Note that an increase in inner steps from 5 to 10 with the MAML5 trained model does not provide an additional meta-test accuracy boost, consistent with previous work \citep{Miranda2020}.
Note that the fact that the MAML5 representation matches the USL representation when both use the same adaptation method is not surprising -- given that:
1) previous work has shown that the distance between the body of an adapted MAML model is minimal compared to the unadapted MAML (which we reproduce in \ref{feature_extractor_sl_vs_maml_vs_adapted_maml} in the green line)
and 2) the fact that a MAML5 adaptation is only 5 steps of MAML while LR fully converges the prediction layer.
We want to highlight that only the MAML5 model achieved the maximum meta-test performance of 0.6 with the MAML5 adaptation -- suggesting that the USL and MAML5 meta-learning algorithms might learn different representations.
For USL to have a fair comparison during meta-test time when using the MAML adaptation, we provide the MAML final layer learned initialization parameters to the USL model (but any is fine due to convexity when using a fixed feature extractor).
This is needed since during meta-training USL is trained with a union of all the labels (64) -- so it does not even have the right output size of 5 for few-shot prediction.
Meta-testing was done in the standard 5-way, 5-shot regime.
}
|
\label{sl_vs_maml_accuracy_comparison_table}
|
\begin{tabular}{lll}
\toprule
Meta-train Initialization & Adaptation at Inference & Meta-test Accuracy \\
\midrule
Random & no adaptation & 19.3 $\pm$ 0.80 \\
MAML0 & no adaptation & 20.0 $\pm$ 0.00 \\
USL & no adaptation & 15.0 $\pm$ 0.26 \\
\midrule
Random & MAML5 adaptation & 34.2 $\pm$ 1.16 \\
\textbf{MAML5} & \textbf{MAML5 adaptation} & \textbf{62.4 $\pm$ 1.64} \\
USL & MAML5 adaptation & 25.1 $\pm$ 0.98 \\
\midrule
Random & MAML10 adaptation & 34.1 $\pm$ 1.23 \\
\textbf{MAML5} & \textbf{MAML10 adaptation} & \textbf{62.3 $\pm$ 1.50} \\
USL & MAML10 adaptation & 25.1 $\pm$ 0.97 \\
\midrule
Random & Adapt Head only (with LR) & 40.2 $\pm$ 1.30 \\
MAML5 & Adapt Head only (with LR) & 59.7 $\pm$ 1.37 \\
\textbf{USL} & \textbf{Adapt Head only (with LR)} & \textbf{60.1 $\pm$ 1.37} \\
\bottomrule
\end{tabular}
|
212,215
|
2208.01545v1
| null |
\caption{
\textbf{
Benchmarks of increasing diversity are created by increasing $\sigma_m$, or the standard deviation of the class mean }
A larger $\sigma_m$ increases the variance of the class means, making their respective class distributions farther apart on average and causing both the Hellinger-based distribution diversity and Task2Vec-based task diversity coefficients to increase. We varied $\sigma_m$ from 0.01 to 1000 and fixed all remaining benchmark parameters to obtain 7 different Gaussian benchmarks. The corresponding Hellinger-based distribution diversity coefficients were obtained by numerically approximating the expected Hellinger distance between two classes sampled from the benchmark and computing the 95\% confidence interval of the approximation. We also computed Task2Vec-based task diversity coefficients as an alternative measure to diversity using a random 3-layer fully connected probe network described in Section \ref{experimental_details}. Figure \ref{1d_gaussian_heatmaps} visualizes the Task2Vec task diversities among the synthetic benchmarks via a heatmap showing the relative pairwise distance between sampled tasks.
}
|
\label{div_gaussian_table}
|
\begin{tabular}{lll}
\toprule
Benchmark Parameters \\ $(\mu_m, \sigma_m, \mu_s, \sigma_s)$ & Hellinger-based Distribution Diversity & Task2Vec-based Task Diversity \\
\midrule
(0, 0.01, 1, 0.01) & 7.475e-05 $\pm$ 4.891e-07 & 0.247 $\pm$ 1.04e-3 \\
(0, 1, 1, 0.01) & 0.183 $\pm$ 1.24e-3 & 0.271 $\pm$ 1.15e-3 \\
(0, 3, 1, 0.01) & 0.574 $\pm$ 2.28e-3 & 0.393 $\pm$ 1.79e-3 \\
(0, 10, 1, 0.01) & 0.860 $\pm$ 1.75e-3 & 0.470 $\pm$ 2.35e-3 \\
(0, 20, 1, 0.01) & 0.929 $\pm$ 1.31e-3 & 0.533 $\pm$ 2.47e-3 \\
(0, 30, 1, 0.01) & 0.952 $\pm$ 1.10e-3 & 0.537 $\pm$ 2.57e-3 \\
(0, 1000, 1, 0.01) & 0.998 $\pm$ 2.07e-4 & 0.546 $\pm$ 2.74e-3 \\
\bottomrule
\end{tabular}
|
212,216
|
2302.13834v2
| null |
\caption{Results on NICE target, we performed 30 runs with different seeds for each approach. For MCD we used the results from \citep{doucet2022annealed}. \label{tab:was}}
| null |
\begin{tabular}{@{}ccccc@{}}
\toprule
Method & DDS & PIS & SMC & MCD \\ \midrule
$\ln Z$ Estimate & $-3.204 \pm 0.645$ & $-3.933 \pm 0.754$ & $-4.255 \pm 2.043$ & $-6.25 $ \\
$\gW^2_{\gamma=0.01}(\pi_{\mathrm{true}}, \hat{\pi})$ & $658.079$ & $658.778$ & $750.245$ & NA \\
\bottomrule
\end{tabular}
|
212,217
|
2302.13834v2
| null |
\caption{Training time per ELBO gradient update on 300 samples. \label{fig:training_time}}
| null |
\begin{tabular}{@{}ccccc@{}}
\toprule
& \multicolumn{4}{c}{Training Times per Iteration - seconds (LGCP)} \\ \cmidrule(l){1-5}
& $k=64$ & $k=128$ & $k=256$ & $k=512$ \\ \cmidrule(l){2-5}
DDS & $0.104 \pm 0.0004$ & $0.205 \pm 0.0004$ & $0.406 \pm 0.0004$ & $0.809 \pm 0.0004$ \\
PIS & $0.104 \pm 0.0004$ & $0.205 \pm 0.0003$ & $0.406 \pm 0.0005$ & $0.808 \pm 0.0006$ \\ \bottomrule
\end{tabular}
|
212,218
|
2302.13834v2
| null |
\caption{Comparing EM vs Exponential integrators on Funnel and Ionosphere datasets. We can see how EM significantly overestimates $\ln Z$.\label{fig:euvsexp}}
| null |
\begin{tabular}{@{}llccccc@{}}
\toprule
\multirow{2}{*}{} & \multicolumn{3}{c}{Funnel} & \multicolumn{3}{c}{Ionosphere} \\ \cmidrule(l){2-7}
& Exponential Integrator & Euler Mayurama & Ground Truth & \multicolumn{1}{l}{Exponential Integrator} & Euler Mayurama & Ground Truth \\ \midrule
$k=64$ & $-0.206 \pm 0.059$ & $2.425 \pm 0.266$ & \multirow{4}{*}{$0$} & $-111.693 \pm 0.169$ & $-106.364 \pm 0.371$ & \multirow{4}{*}{$-111.560$} \\
$k=128$ & $-0.176\pm0.099$ & $2.011\pm 0.532$ & & $-111.587 \pm 0.136$ & $-111.158\pm 0.399$ & \\
$k=256$ & $-0.221 \pm 0.076$ & \multicolumn{1}{l}{$-0.086 \pm 0.124$} & & $-111.575\pm 0.163$ & \multicolumn{1}{l}{$-112.271 \pm 0.366$} & \\
$k=512$ & $-0.176 \pm 0.068$ & $-0.154 \pm 0.066$ & & $-111.582\pm 0.136$ & $-112.912 \pm 0.353$ & \\ \bottomrule
\end{tabular}
|
212,219
|
2302.13834v2
| null |
\caption{Ablation for $\cos^2$ scheduling.}
|
\label{tab:cos_sq}
|
\begin{tabular}{@{}llllllll@{}}
\toprule
Method & Funnel & LGCP & VAE & Sonar & Ionosphere & Brownian & NICE \\ \midrule
PIS Cos decay & $-0.256 \pm 0.094$ & $501.060 \pm 0.881$ & $-110.025 \pm 0.0720$ & $-108.843 \pm 0.242$ & $-111.618 \pm 0.127$ & $1.035 \pm 0.097$ & $-4.212 \pm 0.620$ \\
PIS$_{\mathrm{uniform}}$ & $-0.268 \pm 0.07$ & $502.002 \pm 0.957$ & $-110.025 \pm0.070$ & $-109.349 \pm 0.356$ & $-111.602 \pm 0.145$ & $0.960 \pm 0.145$ & $-3.985 \pm 0.349$ \\
DDS & $-0.176 \pm 0.098$ & $497.944 \pm 0.994$ & $-110.012 \pm 0.071$ & $-108.903 \pm 0.226$ & $-111.587 \pm 0.136$ & $1.047 \pm 0.156$ & $-3.490 \pm 0.568$ \\
DDS$_{\mathrm{uniform}}$ & $-0.275 \pm 0.127$ & $477.680 \pm 1.338$ & $-110.245 \pm 0.188$ & $-109.556 \pm 0.468$ & $-111.736 \pm 0.161$ & $0.475 \pm 0.166$ & $-7.047 \pm 0.691$ \\ \bottomrule
\end{tabular}
|
212,220
|
2302.13834v2
| null |
\caption{Results for PIS grad with and without detaching the gradient. \label{tab:detach}}
| null |
\begin{tabular}{@{}llllllll@{}}
\toprule
Method & Funnel & LGCP & VAE & Sonar & Ionosphere & Brownian & NICE \\ \midrule
PIS Grad Detach & $-0.268 \pm 0.07$ & $502.002 \pm 0.957$ & $-110.025 \pm0.0704$ & $-109.349 \pm 0.356$ & $-111.602 \pm 0.145$ & $0.960 \pm 0.145$ & $-3.985 \pm 0.349$ \\
PIS Grad & $-0.268 \pm 0.09$ & $501.273 \pm 0.798$ & $-110.033 \pm 0.0647$ & $-109.377 \pm 0.361$ & $-111.705 \pm 0.190$ & $-5.793 \pm 0.499$ & $-3.927 \pm 0.666$ \\ \bottomrule
\end{tabular}
|
212,233
|
2111.13420v3
| null |
\caption{Classification accuracy ($\%$) of different DG methods on VLCS with AlexNet as the backbone.}
|
\label{tab: vlcs}
|
\begin{tabular}{c|ccccc}
\hline
\multirow{2}{*}{Methods} & \multicolumn{5}{c}{VLCS} \\ \cline{2-6}
& Caltech & Labelme & Pascal & Sun & Avg. \\ \hline
MLDG
& 97.9 & 59.5 &66.4 & 64.8 & 72.2 \\
Epi-FCR & 94.1 & 64.3 & 67.1 & 65.9 & 72.9 \\
JiGen & 96.93 & 60.9 & 70.6 & 64.3 & 73.2 \\
MMLD & 96.6 & 58.7 & \textbf{72.1} & 66.8 & 73.5 \\
MASF & 94.8 & \textbf{64.9} & 69.1 & 67.6 & 74.1 \\\hline
ERM & 96.3 & 59.7 & 70.6 & 64.5 & 72.8 \\
ERM+MAML & 97.8 & 58.0 & 67.1 & 64.1 & 71.8 \\
\rowcolor[gray]{0.9}
ERM+\ours & 97.8 & 60.1 & 69.7 & 67.3 & 73.7 \\\hline
ERM$^*$ & 96.4 & 60.7 & 68.6 & 66.2 & 73.0 \\
ERM$^*$+MAML & 98.1 & 58.2 & 69.6 & 64.5 & 72.6 \\
\rowcolor[gray]{0.9}
ERM$^*$+\ours & \textbf{98.1} & 62.4 & 69.3 & \textbf{69.1} & \textbf{74.7} \\ \hline
\end{tabular}
|
212,221
|
2302.13834v2
| null |
\caption{Fitted hyperparameters for funnel, lgcp and ion experiments. \label{hype:base}}
| null |
\begin{tabular}{@{}llllllllll@{}}
\toprule
& \multicolumn{3}{l}{funnel} & \multicolumn{3}{l}{lgcp} & \multicolumn{3}{l}{ion} \\ \midrule
& DDS & PIS & UDMP & DDS & PIS & UDMP & DDS & PIS & UDMP \\
$K=64$ & $\sigma=1.075,\alpha=1.075$ & $\sigma=1.068$ & $\sigma= 1.85, \alpha= 1.67, m= 0.9$ & $\sigma= 2.1, \alpha=1.50$ & $\sigma= 1.068$ & $\sigma=1.1, \alpha=2.5, m= 0.4$ & $\sigma = 0.688, \alpha= 1.463$ & $\sigma =0.253$ & $\sigma= 0.6, \alpha= 3.85, m= 0.600$ \\
$K=128$ & $\sigma=1.075, \alpha= 0.6875$ & $\sigma= 0.416$ & $\sigma=1.85, \alpha= 3.7, m=0.9$ & $\sigma= 2.1, \alpha=0.75$ & $\sigma = 0.742$ & $\sigma=1.4, \alpha=2.5, m=0.4$ & $\sigma =0.3, \alpha=1.075$ & $\sigma= 0.09$ & $\sigma=0.6, \alpha =3.85, m = 0.600$ \\
$K=256$ & $\sigma=1.85, \alpha=0.3$ & $\sigma= 0.742$ & $\sigma=1.075, \alpha = 2.5, m= 0.9$ & $\sigma= 2.1, \alpha=0.900$ & $\sigma = 0.579$ & $\sigma=1.4, \alpha=4.5, m = 0.4$ & $\sigma= 0.3, \alpha =0.688$ & $\sigma = 0.416$ & $\sigma= 0.6, \alpha=3.85, m= 1.0$ \\
$K=512$ & $\sigma=1.463, \alpha=0.3$ & $\sigma = 0.253$ & $\sigma =0.688, \alpha=3.7, m= 0.9$ & $\sigma =2.1, \alpha=1.500$ & $\sigma= 0.416$ & $\sigma=1.7, \alpha=4.5, m= 0.4$ & $\sigma =0.688, \alpha =0.688$ & $\sigma= 0.253$ & $\sigma=0.6, \alpha=3.85, m=1.0$ \\ \bottomrule
\end{tabular}
|
212,222
|
2302.13834v2
| null |
\caption{Fitted hyperparameters for sonar, brownian, vae and nice experiments. \label{hype:logreg}}
| null |
\begin{tabular}{@{}lllllllllllll@{}}
\toprule
& \multicolumn{3}{l}{lr\_sonar} & \multicolumn{3}{l}{vae} & \multicolumn{3}{l}{brownian} & \multicolumn{3}{l}{nice} \\ \midrule
& DDS & PIS & UDMP & DDS & PIS & UDMP & DDS & PIS & UDMP & DDS & PIS & UDMP \\
$K=64$ & $\sigma= 0.3, \alpha=1.650$ & $\sigma= 0.253$ & $\sigma= 1.15, \alpha=1.7, m= 2.2$ & $\sigma= 0.61, \alpha=2.2$ & $\sigma=0.253$ & NA & $\sigma= 0.1, \alpha=2.35$ & $\sigma =0.084$ & $\sigma=0.115, \alpha=4.8, m=2.2$ & $\sigma= 1.5, \alpha =2.125$ & $\sigma =0.75$ & $\sigma=1.2, \alpha= 1.0, m= 0.9$ \\
$K=128$ & $\sigma=0.3, \alpha=1.2$ & $\sigma= 0.253$ & $\sigma= 0.55, \alpha=1.7, m= 3.1$ & $\sigma 0.61, \alpha=1.670$ & $\sigma= 0.2523$ & NA & $\sigma =0.1, \alpha=1.8$ & $\sigma= 0.093$ & $\sigma=0.115, \alpha=4.8, m=2.2$ & $\sigma= 1.5, \alpha =1.75$ & $\sigma =0.588$ & $\sigma=1.2, \alpha= 1.0 m=1.65$ \\
$K=256$ & $\sigma=0.3, \alpha=0.75$ & $\sigma= 0.253$ & $\sigma=0.55, \alpha= 2.9, m= 2.2$ & $\sigma= 0.61, \alpha=1.140$ & $\sigma= 0.506$ & NA & $\sigma =0.1, \alpha=2.35$ & $\sigma= 0.043$ & $\sigma=0.115, \alpha=3.75, m=2.2$ & $\sigma= 1.5, \alpha =1.75$ & $\sigma =0.425$ & $\sigma=1.2, \alpha= 2.5, m=0.9$ \\
$K=512$ & $\sigma=0.3, \alpha=0.75$ & $\sigma= 0.253$ & $\sigma= 0.55, \alpha=2.9, m=3.1$ & $\sigma =0.61, \alpha =1.670$ & $\sigma=0.416$ & NA & $\sigma =0.1, \alpha=1.8$ & $\sigma= 0.0408$ & $\sigma=0.115, \alpha=4.8, m=2.2$ & $\sigma= 1.5, \alpha =2.5$ & $\sigma =0.263$ & $\sigma=1.2, \alpha= 2.5, m=1.65$ \\ \bottomrule
\end{tabular}
|
212,223
|
2301.12245v1
| null |
\caption{Results on CIFAR-100.}
|
\label{tbl:results-cifar100}
|
\begin{tabular}{@{}lcccccc@{}}
\toprule
{\bf Setting} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & $\tau = 1$ & $\tau = 4$ & $\tau = 1$ & $\tau = 4$ & \\
\midrule
ResNet-56 $\rightarrow$ LeNet-5x8 & 47.3 $\pm$ 0.6 & 50.1 $\pm$ 0.4 & 59.9 $\pm$ 0.2 & 61.9 $\pm$ 0.2 & \best{66.1 $\pm$ 0.4} & 72.0\\
ResNet-56 $\rightarrow$ ResNet-20 & 67.7 $\pm$ 0.5 & 68.2 $\pm$ 0.3 & \best{71.6 $\pm$ 0.2} & 69.6 $\pm$ 0.3 & 71.4 $\pm$ 0.3 & 72.0\\
ResNet-110 $\rightarrow$ LeNet-5x8 & 47.2 $\pm$ 0.5 & 48.6 $\pm$ 0.8 & 59.0 $\pm$ 0.3 & 60.8 $\pm$ 0.2 & \best{65.8 $\pm$ 0.2} & 73.4\\
ResNet-110 $\rightarrow$ ResNet-20 & 67.8 $\pm$ 0.3 & 67.8 $\pm$ 0.2 & 71.2 $\pm$ 0.0 & 69.0 $\pm$ 0.3 & \best{71.4 $\pm$ 0.0} & 73.4\\
\bottomrule
\end{tabular}
|
212,224
|
2301.12245v1
| null |
\caption{Results on Tiny ImageNet.}
|
\label{tbl:results-imagenet}
|
\begin{tabular}{@{}lcccccc@{}}
\toprule
{\bf Setting} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & $\tau = 1$ & $\tau = 2$ & $\tau = 1$ & $\tau = 2$ & \\
\midrule
MobileNet-V3-125 $\rightarrow$ MobileNet-V3-35 & 58.5 $\pm$ 0.2 & 59.2 $\pm$ 0.1 & 60.2 $\pm$ 0.2 & 60.7 $\pm$ 0.2 & \best{62.3 $\pm$ 0.3} & 62.7 \\
ResNet-101 $\rightarrow$ MobileNet-V3-35 & 58.5 $\pm$ 0.2& 59.4 $\pm$ 0.5 & 61.6 $\pm$ 0.2 & 61.1 $\pm$ 0.3 & \best{62.0 $\pm$ 0.3} & 66.0 \\
MobileNet-V3-125 $\rightarrow$ VGG-16 & 48.9 $\pm$ 0.3 & 54.1 $\pm$ 0.4 & 59.4 $\pm$ 0.4 & 58.9 $\pm$ 0.7 & \best{62.3 $\pm$ 0.3} & 62.7 \\
ResNet-101 $\rightarrow$ VGG-16 & 48.6 $\pm$ 0.4 & 53.1 $\pm$ 0.4 & 60.6 $\pm$ 0.2 & 60.4 $\pm$ 0.2 & \best{64.0 $\pm$ 0.1} & 66.0 \\
\bottomrule
\end{tabular}
|
212,225
|
2301.12245v1
| null |
\caption{Results on binary CIFAR-100. Every second line is an MSE student.}
|
\label{tbl:results-cifar100-binary}
|
\begin{tabular}{@{}lcccccc@{}}
\toprule
{\bf Setting} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & $\tau = 1$ & $\tau = 4$ & $\tau = 1$ & $\tau = 4$ & \\
\midrule
ResNet-56 $\rightarrow$ LeNet-5x8 & 71.5 $\pm$ 0.2 & 72.4 $\pm$ 0.1 & 73.6 $\pm$ 0.2 & 74.7 $\pm$ 0.2 & \best{76.1 $\pm$ 0.2} & 77.9 \\
ResNet-56 $\rightarrow$ LeNet-5x8 & 71.5 $\pm$ 0.4 & 71.9 $\pm$ 0.3 & 73.0 $\pm$ 0.3 & \best{75.1 $\pm$ 0.3} & \best{75.1 $\pm$ 0.1} & 77.9 \\
\midrule
ResNet-56 $\rightarrow$ ResNet-20 & 75.8 $\pm$ 0.5 & 76.1 $\pm$ 0.2 & 77.1 $\pm$ 0.6 & 77.8 $\pm$ 0.3 & \best{78.1 $\pm$ 0.1} & 77.9 \\
ResNet-56 $\rightarrow$ ResNet-20 & 76.1 $\pm$ 0.5 & 76.0 $\pm$ 0.2 & 77.4 $\pm$ 0.3 & 78.0 $\pm$ 0.2 & \best{78.4 $\pm$ 0.3} & 77.9 \\
\midrule
ResNet-110 $\rightarrow$ LeNet-5x8 & 71.4 $\pm$ 0.4 & 71.9 $\pm$ 0.1 & 72.9 $\pm$ 0.3 & 74.3 $\pm$ 0.3 & \best{75.4 $\pm$ 0.3} & 78.4 \\
ResNet-110 $\rightarrow$ LeNet-5x8 & 71.6 $\pm$ 0.2 & 71.5 $\pm$ 0.4 & 72.6 $\pm$ 0.4 & \best{74.8 $\pm$ 0.4} & 74.6 $\pm$ 0.2 & 78.4 \\
\midrule
ResNet-110 $\rightarrow$ ResNet-20 & 76.0 $\pm$ 0.3 & 76.0 $\pm$ 0.2 & 77.0 $\pm$ 0.1 & 77.3 $\pm$ 0.2 & \best{78.0 $\pm$ 0.4} & 78.4 \\
ResNet-110 $\rightarrow$ ResNet-20 & 76.1 $\pm$ 0.2 & 76.4 $\pm$ 0.3 & 77.6 $\pm$ 0.3 & 77.9 $\pm$ 0.2 & \best{78.1 $\pm$ 0.1} & 78.4 \\
\bottomrule
\end{tabular}
|
212,226
|
2301.12245v1
| null |
\caption{Initial learning rates for different dataset and model pairs.}
|
\label{tab:learning-rate-table}
|
\begin{tabular}{llc}
\toprule
Dataset & Model & Learning rate\\
\midrule
\multirow{4}*{CIFAR-10, CIFAR-100, binary CIFAR-100} & ResNet-56 (teacher) & 0.1 \\
& ResNet-110 (teacher) & 0.1 \\
& ResNet-20 (CE or MSE students) & 0.1 \\
& LeNet-5x8 (CE or MSE students) & 0.04 \\
\midrule
\multirow{3}*{Tiny ImageNet} & MobileNet-V3-125 (teacher) & 0.04\\
& ResNet-101 (teacher) & 0.1 \\
& MobileNet-V3-35 (student) & 0.04 \\
& VGG-16 (student) & 0.01 \\
\bottomrule
\end{tabular}
|
212,227
|
2301.12245v1
| null |
\caption{Results on CIFAR-10. Every second line is an MSE student.}
|
\label{tbl:results-cifar10}
|
\begin{tabular}{@{}lcccccc@{}}
\toprule
{\bf Setting} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & $\tau = 1$ & $\tau = 4$ & $\tau = 1$ & $\tau = 4$ & \\
\midrule
ResNet-56 $\rightarrow$ LeNet-5x8 & 81.8 $\pm$ 0.5 & 82.4 $\pm$ 0.5 & 86.0 $\pm$ 0.2 & 86.8 $\pm$ 0.2 & \best{88.6 $\pm$ 0.1} & 93.2 \\
ResNet-56 $\rightarrow$ LeNet-5x8 & 83.4 $\pm$ 0.3 & 83.1 $\pm$ 0.2 & 84.9 $\pm$ 0.1 & 85.6 $\pm$ 0.1 & \best{87.1 $\pm$ 0.1} & 93.2 \\
\midrule
ResNet-110 $\rightarrow$ LeNet-5x8 & 81.7 $\pm$ 0.3 & 81.9 $\pm$ 0.5 & 85.8 $\pm$ 0.1 & 86.5 $\pm$ 0.1 & \best{88.8 $\pm$ 0.1} & 93.9 \\
ResNet-110 $\rightarrow$ LeNet-5x8 & 83.2 $\pm$ 0.4 & 83.2 $\pm$ 0.1 & 85.0 $\pm$ 0.3 & 85.6 $\pm$ 0.1 & \best{87.1 $\pm$ 0.2} & 93.9 \\
\midrule
ResNet-110 $\rightarrow$ ResNet-20 & 91.4 $\pm$ 0.2 & 91.4 $\pm$ 0.1 & 92.8 $\pm$ 0.0 & 92.2 $\pm$ 0.3 & \best{93.1 $\pm$ 0.1} & 93.9 \\
ResNet-110 $\rightarrow$ ResNet-20 & 90.9 $\pm$ 0.1 & 90.9 $\pm$ 0.2 & 91.6 $\pm$ 0.2 & 91.2 $\pm$ 0.1 & \best{92.1 $\pm$ 0.2} & 93.9 \\
\bottomrule
\end{tabular}
|
212,228
|
2301.12245v1
| null |
\caption{\rebuttal{Knowledge distillation results on CIFAR-100 with varying loss mixture coefficient $\alpha$.}}
|
\label{tbl:cifar100-kd-alpha}
|
\begin{tabular}{@{}llcccccc@{}}
\toprule
{\bf Setting} & {\bf $\alpha$} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & & $\tau = 1$ & $\tau = 4$ & $\tau = 1$ & $\tau = 4$ & \\
\midrule
\multirow{5}{*}{{\centering \begin{tabular}{c}ResNet-56 $\rightarrow$\\LeNet-5x8\end{tabular}}} & 0.2 & \multirow{5}{*}{47.3 $\pm$ 0.6} & 47.6 $\pm$ 0.7 & 57.6 $\pm$ 0.2 & 54.3 $\pm$ 0.7 & 59.0 $\pm$ 0.6 & \multirow{5}{*}{72.0}\\
& 0.4 & & 48.9 $\pm$ 0.3 & 58.9 $\pm$ 0.4 & 56.7 $\pm$ 0.5 & 62.5 $\pm$ 0.2 & \\
& 0.6 & & 49.4 $\pm$ 0.5 & 59.7 $\pm$ 0.0 & 61.1 $\pm$ 0.0 & 65.3 $\pm$ 0.2 & \\
& 0.8 & & 49.8 $\pm$ 0.1 & 60.1 $\pm$ 0.1 & 62.0 $\pm$ 0.1 & 65.9 $\pm$ 0.2 & \\
& 1.0 & & 50.1 $\pm$ 0.4 & 59.9 $\pm$ 0.2 & 61.9 $\pm$ 0.2 & 66.1 $\pm$ 0.4 & \\
\midrule
\multirow{5}{*}{{\centering \begin{tabular}{c}ResNet-56 $\rightarrow$\\ResNet-20\end{tabular}}} & 0.2 & \multirow{5}{*}{67.7 $\pm$ 0.5} & 67.9 $\pm$ 0.3 & 70.3 $\pm$ 0.3 & 68.2 $\pm$ 0.3 & 70.3 $\pm$ 0.1 & \multirow{5}{*}{72.0}\\
& 0.4 & & 67.9 $\pm$ 0.1 & 71.0 $\pm$ 0.2 & 68.7 $\pm$ 0.2 & 71.4 $\pm$ 0.2 & \\
& 0.6 & & 68.1 $\pm$ 0.3 & 71.3 $\pm$ 0.1 & 69.6 $\pm$ 0.4 & 71.5 $\pm$ 0.2 & \\
& 0.8 & & 68.3 $\pm$ 0.2 & 71.4 $\pm$ 0.4 & 69.8 $\pm$ 0.3 & 71.1 $\pm$ 0.3 & \\
& 1.0 & & 68.2 $\pm$ 0.3 & 71.6 $\pm$ 0.2 & 69.6 $\pm$ 0.3 & 71.4 $\pm$ 0.3 & \\
\bottomrule
\end{tabular}
|
212,229
|
2301.12245v1
| null |
\caption{\rebuttal{Results on CIFAR-100.}}
|
\label{tbl:results-cifar1000-tau=2}
|
\begin{tabular}{@{}lcccccc@{}}
\toprule
{\bf Setting} & {\bf No KD} & \multicolumn{2}{c}{\bf Offline KD} & \multicolumn{2}{c}{\bf Online KD} & {\bf Teacher} \\
& & $\tau = 1$ & $\tau = 2$ & $\tau = 1$ & $\tau = 2$ & \\
\midrule
ResNet-56 $\rightarrow$ LeNet-5x8 & 47.3 $\pm$ 0.6 & 50.1 $\pm$ 0.4 & 55.2 $\pm$ 0.1 & 61.9 $\pm$ 0.2 & \best{64.7 $\pm$ 0.2} & 72.0\\
ResNet-56 $\rightarrow$ ResNet-20 & 67.7 $\pm$ 0.5 & 68.2 $\pm$ 0.3 & 70.4 $\pm$ 0.3 & 69.6 $\pm$ 0.3 & \best{70.8 $\pm$ 0.3} & 72.0\\
ResNet-110 $\rightarrow$ LeNet-5x8 & 47.2 $\pm$ 0.5 & 48.6 $\pm$ 0.8 & 54.0 $\pm$ 0.5 & 60.8 $\pm$ 0.2 & \best{63.9 $\pm$ 0.2} & 73.4\\
ResNet-110 $\rightarrow$ ResNet-20 & 67.8 $\pm$ 0.3 & 67.8 $\pm$ 0.2 & 69.3 $\pm$ 0.2 & 69.0 $\pm$ 0.3 & \best{70.5 $\pm$ 0.3} & 73.4\\
\bottomrule
\end{tabular}
|
212,230
|
2111.13420v3
| null |
\caption{Classification accuracy ($\%$) of different DG methods on PACS with ResNet-18 and Digits-DG with a CNN backbone ~\citep{zhou2021domainMixstyle}. $\ddagger$: our reimplemented results using the official code are different from the ones reported in the original paper.}
|
\label{tab:pacs}
|
\begin{tabular}{c|cccc|c|c|ccccc}
\hline \hline
\multirow{2}{*}{Method} & \multicolumn{5}{c|}{PACS} & & \multicolumn{5}{c}{Digits-DG} \\ \cline{2-6} \cline{8-12}
& A & C & P & S & Avg. & & MINIST & MINIST-M & SVHN & \multicolumn{1}{c|}{SYN} & Avg. \\ \cline{1-6} \cline{8-12}
MMD-AAE & 75.2 & 72.7 & 96.0 & 64.2 & 77.0 & & 96.5 & 58.4 & 65.0 & \multicolumn{1}{c|}{78.4} & 74.6 \\
CCSA & 80.5 & 76.9 & 93.6 & 66.8 & 79.4 & & 95.2 & 58.2 & 65.5 & \multicolumn{1}{c|}{79.1} & 74.5 \\
JiGen & 79.4 & 75.3 & 96.0 & 71.6 & 80.5 & & 96.5 & 61.4 & 63.7 & \multicolumn{1}{c|}{74.0} & 73.9 \\
CrossGrad & 79.8 & 76.8 & 96.0 & 70.2 & 80.7 & & 96.7 & 61.1 & 65.3 & \multicolumn{1}{c|}{80.2} & 75.8 \\
MLDG & 79.5 & 77.3 & 94.3 & 71.5 & 80.7 & & 94.7 & 60.3 & 61.5 & \multicolumn{1}{l|}{75.4} & 72.6 \\
MASF & 80.3 & 77.2 & 95.0 & 71.7 & 81.1 & & - & - & - &\multicolumn{1}{c|}{-}& - \\
MetaReg & 83.7 & 77.2 & 95.5 & 70.3 & 81.7 & & - & - & - & \multicolumn{1}{c|}{-} & - \\
RSC & 83.4 & 80.3 & 96.0 & 80.9 & 85.2 & & - & - & - & \multicolumn{1}{c|}{-} & - \\
MatchDG & 81.3 & \textbf{80.7} & \textbf{96.5} & 79.7 & 84.6 & & - & - & - & \multicolumn{1}{c|}{-} & - \\
MixStyle$^\ddagger$ & 83.0 & 78.6 & 96.3 & 71.2 & 82.3 & & 96.5 & 63.5 & 64.7 & \multicolumn{1}{c|}{81.2} & 76.5 \\
FACT & \textbf{85.4} & 78.4 & 95.2 & 79.2 & 84.5 & & \textbf{97.9} & 65.6 & 72.4 & \multicolumn{1}{c|}{\textbf{90.3}} & 81.5 \\
\hline \hline
ERM & 77.0 & 75.9 & 96.0 & 69.2 & 79.5 & & 95.8 & 58.8 & 61.7 & \multicolumn{1}{c|}{78.6} & 73.7 \\
ERM+MAML & 77.0 & 74.5 & 94.8 & 72.1 & 79.6 & & 96.0 & 63.1 & 65.0 & \multicolumn{1}{c|}{81.1} & 76.5 \\
\rowcolor[gray]{0.9}
ERM+\ours & 80.7 & 76.9 & 95.6 & 74.5 & 81.9 & & 95.8 & 63.7 & 65.8 & \multicolumn{1}{c|}{80.7} & 76.5 \\\cline{1-6} \cline{8-12}
ERM$^*$ & 82.5 & 74.2 & 95.4 & 76.5 & 82.1 & & 96.1 & 65.0 & 73.0 & \multicolumn{1}{c|}{84.6} & 79.7 \\
ERM$^*$+MAML & 81.8 & 73.2 & 94.8 & 75.7 & 81.4 & & 96.2 &67.0 & 74.0 & \multicolumn{1}{c|}{84.1} & 80.3 \\
\rowcolor[gray]{0.9}
%%%ERM$^*$+\ours & \textbf{84.2} & 72.8 & 95.1 & \textbf{80.5} & \textbf{83.2} & & 95.6 & \textbf{68.8} & \textbf{76.5} & \multicolumn{1}{c|}{\textbf{86.0}} & \textbf{81.7} \\ \cline{1-6} \cline{8-12}
ERM$^*$+\ours & 84.2 & 78.8 & 95.1 & \textbf{83.2} & \textbf{85.3} & & 95.6 & \textbf{68.8} & \textbf{76.5} & \multicolumn{1}{c|}{86.0} & \textbf{81.7} \\ \cline{1-6} \cline{8-12}
\hline \hline
\end{tabular}
|
212,231
|
2111.13420v3
| null |
\caption{Classification accuracy (\%) of different DG methods on PACS with ResNet-50.}
|
\label{tab:resnet50}
|
\begin{tabular}{c|cccc|c}
\hline \hline
Method & A & C & P & S & Avg. \\ \hline
MatchDG & 85.6 & 82.1 & 97.9 & 78.8 & 86.1 \\
RSC & 87.9 & \textbf{82.2} & 97.9 & 83.4 & 87.8 \\
FACT & 89.6 & 81.8 & 96.8 & 84.5 & 88.2 \\
Fish & - & - & - & - & 85.5 \\ \hline \hline
\multicolumn{1}{c|}{ERM$^*$} & \multicolumn{1}{c}{88.0} & \multicolumn{1}{c}{78.8} & \multicolumn{1}{c}{\textbf{98.2}} & \multicolumn{1}{c|}{81.7} & \multicolumn{1}{c}{86.7} \\
\rowcolor[gray]{0.9} \multicolumn{1}{c|}{ERM$^*$+\ours} & \textbf{89.7} & \multicolumn{1}{c}{\textbf{82.2}} & 97.9 & \multicolumn{1}{c|}{\textbf{86.2}} & \multicolumn{1}{c}{\textbf{89.0}} \\ \hline \hline
\end{tabular}
|
212,232
|
2111.13420v3
| null |
\caption{Classification accuracy ($\%$) of different DG methods on Office-Home with ResNet-18 as the backbone.}
|
\label{tab: office-home}
|
\begin{tabular}{c|ccccc}
\hline
% \small
\multirow{2}{*}{Method} & \multicolumn{5}{c}{Office-Home} \\ \cline{2-6}
& Art & Clipart & Product & Realworld & Avg. \\ \cline{1-6}
JiGen & 53.0 & 47.5 & 71.5 & 72.8 & 61.2 \\
MMD-AAE & 56.5 & 47.3 & 72.1 & 74.8 & 62.7 \\
MLDG & 57.8 &50.3 &70.6 &73.0 & 63.0 \\
CrossGrad & 58.4 & 49.4 & 73.9 & 75.8 & 64.4 \\
CCSA & \textbf{59.9} & 49.9 & 74.1 & 75.7 & 64.9 \\
MixStyle & 58.7 & 53.4 & 74.2 & \textbf{75.9} & 65.5 \\ \hline
ERM & 58.1 & 48.7 & 74.0 & 75.6 & 64.2 \\
ERM+MAML & 56.8 & 52.5 & 74.0 & 74.7 & 64.5 \\
\rowcolor[gray]{0.9}
ERM+\ours & 57.1 & 52.0 & 74.1 & 75.6 & 64.7 \\ \cline{1-6}
ERM$^*$ & 59.6 & 53.0 & \textbf{74.3} & 75.4 & 65.6 \\
ERM$^*$+MAML & 56.2 & 56.1 & 72.6 & 73.2 & 64.5 \\
\rowcolor[gray]{0.9}
ERM$^*$+\ours & 59.3 & \textbf{56.2} & 74.2 & 75.1 & \textbf{66.2} \\ \hline
\end{tabular}
|
212,281
|
2203.07720v3
| null | null |
\label{tab:region-select}
|
\begin{tabular}{l|lll}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
& R@1 & R@5 & R@10 \\ \hline
Track Object & 33.6 & 59.2 & 69.7 \\ \hline
Sorted Selection & 36.0 & 61.0 & 71.8 \\ \hline
\end{tabular}
|
212,234
|
2206.06720v2
| null |
\caption{Root mean squared error results on \hyperref[para:large_regression]{large scale
regression datasets.}}
|
\label{tab:large_regression}
|
\begin{tabular}{lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}l}
\toprule
& \multicolumn{4}{c}{Single-layer} & \multicolumn{8}{c}{Ours} & \multicolumn{2}{c}{\citetalias{salimbeni2017doubly}} \\
\cmidrule(lr){2-5} \cmidrule(lr){6-13} \cmidrule(lr){14-15}
& \multicolumn{2}{c}{SGP} & \multicolumn{2}{c}{VIP} & \multicolumn{2}{c}{DVIP 2} & \multicolumn{2}{c}{DVIP 3} &
\multicolumn{2}{c}{DVIP 4} & \multicolumn{2}{c}{DVIP 5} & \multicolumn{2}{c}{DGP 3} \\
\midrule
Year & 9.15 & 0.01 & 10.27 & 0.01 & 9.61 & 0.03 & 9.34 & 0.02 & 9.30 & 0.03 & 9.27 & 0.03 & $\bm{8.94}$ & $\bm{0.03}$ \\
Airline & 38.61 & 0.05 & 38.90 & 0.06 & 37.96 & 0.03 & 37.91 & 0.05 & 37.83 & 0.03 & $\bm{37.80}$ & $\bm{0.05}$ & 37.95& 0.04\\
Taxi & 554.22& 0.32 & 554.60 & 0.19 & 549.28 & 0.59 & $\bm{531.42}$ &$\bm{1.59}$ & 547.33 &1.03 & 538.94 & 2.23 & 552.90 & 0.33\\
\bottomrule
\end{tabular}
|
212,235
|
2206.06720v2
| null |
\caption{Results on \hyperref[para:image]{image classification datasets.}}
|
\label{tab:image}
|
\begin{tabular}{lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}l}
\toprule
\multirow{2.5}{*}{\textbf{MNIST}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{4}{c}{Ours} &
\multicolumn{6}{c}{\citetalias{salimbeni2017doubly}} \\
\cmidrule(lr){2-5} \cmidrule(lr){6-9} \cmidrule(lr){10-15}
& \multicolumn{2}{c}{SGP} & \multicolumn{2}{c}{VIP}
& \multicolumn{2}{c}{DVIP 2} & \multicolumn{2}{c}{DVIP 3}
& \multicolumn{6}{c}{
\begin{tabular}{c@{\hskip 0.4in}c}
DGP 2 & DGP 3\\
\end{tabular}
} \\
\midrule
Accuracy (\%) & 96.25 & 0.04 & 97.99 & 0.03 & $\bm{98.39}$ & $\bm{0.05}$ & 98.36 & 0.04 & \multicolumn{6}{c}{
\begin{tabular}{r@{$\pm$}l@{\hskip 0.4in}r@{$\pm$}l}
97.75 & 0.04 &97.86 & 0.05 \\
\end{tabular}
} \\
Likelihood & -0.146 & 0.00 & -0.144 & 0.00 & -0.074 & 0.00 & -0.080 & 0.00 & \multicolumn{6}{c}{
\begin{tabular}{r@{$\pm$}l@{\hskip 0.25in}r@{$\pm$}l}
-0.082 & 0.00 & $\bm{-0.072}$ & $\bm{0.00}$ \\
\end{tabular}
} \\
\midrule \midrule
\multirow{2.5}{*}{\textbf{Rectangles}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{8}{c}{Ours} & \multicolumn{2}{c}{\citetalias{salimbeni2017doubly}} \\
\cmidrule(lr){2-5} \cmidrule(lr){6-13} \cmidrule(lr){14-15}
& \multicolumn{2}{c}{SGP} & \multicolumn{2}{c}{VIP}
& \multicolumn{2}{c}{DVIP 2} & \multicolumn{2}{c}{DVIP 3}
& \multicolumn{2}{c}{DVIP 4} & \multicolumn{2}{c}{DVIP 5} & \multicolumn{2}{c}{DGP 3}\\
Accuracy (\%) & 72.54 & 0.14 & 85.63 & 0.18 & 87.84 & 0.20 & $\bm{88.21}$ & $\bm{0.12}$ & 87.43 & 0.20& 86.49 & 0.17 & 75.16 & 0.16\\
Likelihood & -0.518 & 0.00 & -0.348 & 0.00 & -0.306 & 0.00& $\bm{-0.295}$ &$\bm{0.00}$ &-0.309 & 0.00 & -0.320 & 0.00& -0.470 & 0.00\\
AUC & 0.828 & 0.00 & 0.930 & 0.00 &0.950 & 0.00 & $\bm{0.953}$ & $\bm{0.00}$ & 0.947 & 0.00 & 0.939 & 0.00 & 0.858 & 0.00\\
\bottomrule
\end{tabular}
|
212,236
|
2206.06720v2
| null |
\caption{Results on \hyperref[para:large_class]{large classification datasets.}}
|
\label{tab:susy_higgs}
|
\begin{tabular}{lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}lr@{$\pm$}l}
\toprule
\multirow{ 2.5}{*}{\textbf{SUSY}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{8}{c}{Ours} & \multicolumn{2}{c}{\citetalias{salimbeni2017doubly}} \\
\cmidrule(lr){2-5} \cmidrule(lr){6-13} \cmidrule(lr){14-15}
& \multicolumn{2}{c}{SGP} & \multicolumn{2}{c}{VIP} & \multicolumn{2}{c}{DVIP 2} &
\multicolumn{2}{c}{DVIP 3} & \multicolumn{2}{c}{DVIP 4} & \multicolumn{2}{c}{DVIP 5} & \multicolumn{2}{c}{DGP 4}\\
\midrule
Accuracy (\%) & 79.75 & 0.02 & 78.68 & 0.02 & 80.11 & 0.03 & 80.13 & 0.01 & 80.22 & 0.01& $\bm{80.24}$ & $\bm{0.02}$ & 80.06 & 0.01\\
Likelihood & -0.436 & 0.00 & -0.456 & 0.00 & -0.429 & 0.00 & -0.429 & 0.00 & $\bm{-0.427}$ & $\bm{0.00}$ & $\bm{-0.427}$ & $\bm{0.00}$& -0.432 & 0.00\\
AUC & 0.8727 & 0.00 & 0.8572 & 0.00 & 0.8742 & 0.00 & 0.8749 & 0.00& 0.8755 & 0.00 & $\bm{0.8756}$ & $\bm{0.00}$ & 0.8751 & 0.00 \\
\midrule\midrule
\textbf{HIGGS} & \multicolumn{2}{c}{SGP} & \multicolumn{2}{c}{VIP} & \multicolumn{2}{c}{DVIP 2} & \multicolumn{2}{c}{DVIP 3} & \multicolumn{2}{c}{DVIP 4} & \multicolumn{2}{c}{DVIP 5} & \multicolumn{2}{c}{DGP 5}\\
\midrule
Accuracy (\%) & 69.95 & 0.03 & 57.42 & 0.03 & 66.09 & 0.02 & 69.85 & 0.02& 70.43 & 0.01& 72.01 & 0.02 & $\bm{74.92}$ & $\bm{0.01}$\\
Likelihood & -0.573 & 0.00 & -0.672 & 0.00 & -0.611 & 0.00& -0.575& 0.00 & -0.565 & 0.00 & -0.542 & 0.00 & $\bm{-0.501}$ & $\bm{0.00}$\\
AUC & 0.7693 & 0.00 & 0.6247 & 0.00 & 0.7196 & 0.00 & 0.7704 & 0.00 &0.7782 & 0.00 & 0.7962 & 0.00 & $\bm{0.8324}$ & $\bm{0.00}$\\
\bottomrule
\end{tabular}
|
212,237
|
2206.06720v2
| null |
\caption{Results on Boston and Energy dataset using the constrained and unconstrained BNN prior for VIP with \(\alpha = 0\).}
|
\label{tab:prior_overfitting}
|
\begin{tabular}{lcccccc}
\toprule
\textbf{Unconstrained }& RMSE Train & RMSE Test & NLL Train & NLL Test & CRPS Train & CRPS Test \\
\midrule
Boston & \(0.05 \pm 0.00\) & \(5.85 \pm 0.14\) & \(-1.21 \pm 0.19\) & \(5126.07 \pm 274.79\) &\(0.03 \pm 0.00\) & \(4.31 \pm 0.08\) \\
Energy & \(0.14 \pm 0.00\) & \(0.57 \pm 0.01\) & \(-0.51 \pm 0.01\) & \(6.52 \pm 0.42\) & \( 0.079 \pm 0.00\) & \( 0.36 \pm 0.01\)\\
\midrule\midrule
\textbf{Constrained }& RMSE Train & RMSE Test & NLL Train & NLL Test & CRPS Train & CRPS Test \\
\midrule
Boston & \(3.90 \pm 0.02\) & \(4.73 \pm 0.24\) & \(2.77 \pm 0.00\) & \(23.03 \pm 0.07\) &\(2.06 \pm 0.01\) & \(2.40 \pm 0.08\) \\
Energy & \(2.35 \pm 0.03\) & \(2.57 \pm 0.08\) & \(2.27 \pm 0.01\) & \(2.07 \pm 0.02\) & \(1.28 \pm 0.01\) & \(1.27 \pm 0.04\)\\
\bottomrule
\end{tabular}
|
212,238
|
2206.06720v2
| null |
\caption{Results on Power and Protein datasets (UCI) using DVIP with 3 layers and different values of the number of prior samples \(S\).}
|
\label{tab:prior_samples}
|
\begin{tabular}{l*{9}{c}}
\toprule
\textbf{Power} & \(S = 10\) & \(S = 20\)& \(S = 30\)& \(S = 40\)& \(S = 50\) \\
\midrule
RMSE & \(4.03 \pm 0.04\) & \(4.01 \pm 0.04\) & \(3.94 \pm 0.04\) & \(3.92 \pm 0.04\) & \( 3.94 \pm 0.04\)\\
NLL & \(2.81 \pm 0.01\) & \(2.81 \pm 0.00\) & \(2.79 \pm 0.01\) & \(2.78 \pm 0.01\) & \(2.79 \pm 0.01\)\\
CRPS & \(2.19 \pm 0.01\) & \(2.18 \pm 0.01\) & \(2.14 \pm 0.01\) & \(2.11 \pm 0.01\) & \(2.13 \pm 0.01\)\\
CPU Time (s) & \(2693 \pm 19\) & \(2806 \pm 22\) & \(3152 \pm 20\) & \(3451 \pm 63\) & \(3742 \pm 55\)\\
\midrule
\midrule
\textbf{Protein} & \(S = 10\) & \(S = 20\)& \(S = 30\)& \(S = 40\)& \(S = 50\) \\
\midrule
RMSE & \(4.53 \pm 0.01\) & \(4.40 \pm 0.01\) & \(4.28 \pm 0.01\) & \(4.27 \pm 0.01\) & \( 4.21 \pm 0.01\)\\
NLL & \(2.92 \pm 0.00\) & \(2.90 \pm 0.00\) & \(2.87 \pm 0.00\) & \(2.86 \pm 0.00\) & \(2.85 \pm 0.00\)\\
CRPS & \(2.52 \pm 0.00\) & \(2.43 \pm 0.00\) & \(2.36 \pm 0.00\) & \(2.34 \pm 0.00\) & \(2.31 \pm 0.00\)\\
CPU Time (s) & \(2334 \pm 28\) & \(2734 \pm 19\) & \(3616 \pm 11\) & \(3727 \pm 35\) & \(4330 \pm 52\)\\
\bottomrule
\end{tabular}
|
212,239
|
2206.06720v2
| null |
\caption{Results on Power and Protein datasets (UCI) using DVIP with different prior BNN architectures.}
|
\label{tab:prior_structure}
|
\begin{tabular}{lcccccc}
\toprule
\textbf{} & \multicolumn{3}{c}{BNN 10-10-10} & \multicolumn{3}{c}{BNN 20-20} \\
\midrule
\textbf{Power} & DVIP \(2\) & DVIP \(3\) & DVIP \(4\)& DVIP \(2\) & DVIP \(3\) & DVIP \(4\)\\
\midrule
RMSE & \(4.02 \pm 0.04\) & \(3.94 \pm 0.04\) & \(3.97 \pm 0.04\) & \(4.02 \pm 0.04\) & \( 3.99 \pm 0.04\)& \( 3.93 \pm 0.04\)\\
NLL & \(2.81 \pm 0.01\) & \(2.79 \pm 0.00\) & \(2.80 \pm 0.01\) & \(2.81 \pm 0.01\) & \(2.80 \pm 0.01\)& \(2.79 \pm 0.01\)\\
CRPS & \(2.18 \pm 0.01\) & \(2.13 \pm 0.01\) & \(2.15 \pm 0.01\) & \(2.18 \pm 0.01\) & \(2.16 \pm 0.01\) & \(2.13 \pm 0.01\) \\
\midrule
\midrule
\textbf{Protein} & DVIP \(2\) & DVIP \(3\) & DVIP \(4\)& DVIP \(2\) & DVIP \(3\) & DVIP \(4\)\\ \midrule
RMSE & \(4.57 \pm 0.01\) & \(4.37 \pm 0.01\) & \(4.29 \pm 0.01\) & \(4.55 \pm 0.01\) & \(4.43 \pm 0.01\)& \(4.31 \pm 0.01\)\\
NLL & \(2.93\pm 0.00\) & \(2.89 \pm 0.00\) & \(2.87 \pm 0.00\) & \(2.93 \pm 0.00\) & \(2.90 \pm 0.00\) & \(2.87 \pm 0.00\)\\
CRPS & \(2.55 \pm 0.00\) & \(2.42 \pm 0.00\) & \(2.36 \pm 0.00\) & \(2.54 \pm 0.00\) & \(2.45 \pm 0.00\) & \(2.37 \pm 0.00\)\\
\bottomrule
\end{tabular}
|
212,240
|
2206.06720v2
| null |
\caption{Results on Year dataset of VIP with the usual BNN prior with \(2\) hidden units of
width \(10\) and tanh activations, VIP using a BNN that approximates a GP
with RBF kernel, VIP with the last prior and \(100\) inducing points and a
sparse GP with \(100\) inducing points. Experiments with VIP are trained using \(\alpha = 0\). }
|
\label{tab:gp_prior}
|
\begin{tabular}{l*{8}{c}}
\toprule
\textbf{Year} & VIP & VIP-GP (linear regression) & VIP-GP (inducing points) & SGP \\
\midrule
RMSE & \(10.27 \pm 0.01\) & \(10.23 \pm 0.01\) & \(9.28 \pm 0.01\) & \(9.15 \pm 0.01\) \\
NLL & \(3.74 \pm 0.00\) & \(3.77 \pm 0.00\) & \(3.64 \pm 0.00\) & \(3.62 \pm 0.00\)\\
CRPS & \(5.45 \pm 0.01\) & \(5.45 \pm 0.02\) & \(4.85 \pm 0.01 \) & \(4.83 \pm 0.01\)\\
CPU Time (s) & \(1217 \pm 257\) & \(1687\pm 271\) & \(30867 \pm 326\) & \(1874 \pm 265\)\\
\bottomrule
\end{tabular}
|
212,241
|
2206.06720v2
| null |
\caption{Results on Protein UCI dataset using DVIP with the usual BNN prior, the approximated GP prior and deep GPs. Experiments with VIP are trained using \(\alpha = 0\). }
|
\label{tab:Protein}
|
\begin{tabular}{l*{9}{c}}
\toprule
\textbf{BNN Prior} & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 \\
\midrule
RMSE & \(4.76 \pm 0.01\) & \(4.24 \pm 0.01\)& \( 4.14 \pm 0.01\)& \(4.14 \pm 0.01\)& \(\bm{4.09 \pm 0.01}\)& \\
NLL & \(2.98 \pm 0.00\)& \(2.86 \pm 0.00\)& \(2.84 \pm 0.00\)& \(2.84 \pm 0.00\)& \(\bm{2.83 \pm 0.00}\)& \\
CRPS & \(2.68 \pm 0.00\) &\(2.34 \pm 0.00\)& \( 2.26 \pm 0.00\)& \( 2.25 \pm 0.00\)& \(\bm{2.21 \pm 0.00}\) & \\
CPU time (s) & \(3086 \pm 173\) & \(3981 \pm 182\)& \(8604 \pm 774\)& \(9931 \pm 616\)& \( 12568 \pm 327\)\\
\midrule
\textbf{GP Prior} & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 \\
\midrule
RMSE & \(4.89\pm 0.01\) & \(4.26\pm 0.01\)& \(4.07\pm 0.01\) & \(4.02\pm 0.01\) & \(\bm{4.01\pm 0.01}\)\\
NLL & \(3.00 \pm 0.00\)& \(2.87 \pm 0.00\)& \(2.82 \pm 0.00\)& \(2.81 \pm 0.00\)& \(\bm{2.81 \pm 0.00}\) \\
CRPS & \(2.77 \pm 0.00\) &\(2.35 \pm 0.00\)& \(2.21 \pm 0.00\)&\(2.17 \pm 0.00\)&\(\bm{2.16 \pm 0.00}\)\\
CPU time (s) & \(5880\pm 249\) & \(12293\pm 564\) & \(20351\pm 1110\) & \(18514\pm 575\) & \( 25835\pm 1259\)\\
\midrule
\citetalias{salimbeni2017doubly} & SGP & DGP 2 & DGP 3 & DGP 4 & DGP 5 \\
\midrule
RMSE & \(4.56 \pm 0.01\) & \( 4.17 \pm 0.01\) & \(\bm{4.00 \pm 0.01}\)& \( 4.01 \pm 0.01\)& \(4.02 \pm 0.01\)\\
NLL & \(2.93\pm 0.00\) & \( 2.84 \pm 0.00\)& \(\bm{2.79 \pm 0.00}\)& \(\bm{2.79 \pm 0.00}\)& \(2.80 \pm 0.00\) \\
CRPS & \(2.56 \pm 0.00\) & \( 2.31 \pm 0.00\)& \(\bm{2.19 \pm 0.00}\)& \(\bm{2.19 \pm 0.00}\)& \(2.20 \pm 0.00\) \\
CPU time (s) & \(2690 \pm 114\) & \(10031 \pm 129\) & \(17528\pm 689\) & \(16128\pm 190\) & \(20653\pm 969\)\\
\bottomrule
\end{tabular}
|
212,242
|
2206.06720v2
| null |
\caption{Results on Kin8nm UCI dataset using DVIP with the usual BNN prior, the approximated GP prior and deep GPs. Experiments with VIP are trained using \(\alpha = 0\). }
|
\label{tab:Kin8nm}
|
\begin{tabular}{l*{9}{c}}
\toprule
\textbf{BNN Prior} & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 \\
\midrule
RMSE & \(0.15 \pm 0.00\) & \(\bm{0.07 \pm 0.00}\)& \( \bm{0.07 \pm 0.00}\)& \(\bm{0.07 \pm 0.00}\)& \(\bm{0.07 \pm 0.00}\)& \\
NLL & \(-0.47 \pm 0.00\)& \(-1.13 \pm 0.00\)& \(\bm{-1.18 \pm 0.00}\)& \(-1.16 \pm 0.00\)& \(-1.17 \pm 0.00\)& \\
CRPS & \(0.08 \pm 0.00\) &\(\bm{0.04 \pm 0.00}\)& \(\bm{0.04 \pm 0.00}\)& \( \bm{0.04 \pm 0.00}\)& \(\bm{0.04 \pm 0.00}\) & \\
CPU time (s) & \(2109 \pm 57\) & \(5086 \pm 232\)& \(6927 \pm 27\)& \(9459 \pm 77\)& \(11763 \pm 141\)\\
\midrule
\textbf{GP Prior} & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 \\
\midrule
RMSE & \(0.14 \pm 0.00\) & \(0.07 \pm 0.00\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\) \\
NLL & \(-0.43\pm 0.00\) & \(-1.20 \pm 0.00\)& \(-1.25 \pm 0.00\)& \(\bm{-1.26 \pm 0.00}\)& \(- 1.25 \pm 0.00\) \\
CRPS & \(0.08 \pm 0.00\) &\(0.04 \pm 0.00\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\) \\
CPU time (s) & \(6672 \pm 442\) & \(14300 \pm 847\) & \(17573 \pm 1033\) & \(22561 \pm 980\) & \( 22669 \pm 657\)\\
\midrule
\citetalias{salimbeni2017doubly} & SGP & DGP 2 & DGP 3 & DGP 4 & DGP 5 \\
\midrule
RMSE & \(0.09 \pm 0.00\) & \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\) \\
NLL & \(-0.91\pm 0.00\) & \(-1.29 \pm 0.00\)& \(-1.32 \pm 0.00\)& \(\bm{-1.33 \pm 0.00}\)& \(- 1.30 \pm 0.00\) \\
CRPS & \(0.05 \pm 0.00\) &\( \bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\) \\
CPU time (s) & \(2053 \pm 81\) & \(6375 \pm 331\) & \(11147 \pm 472\) & \(17502 \pm 1060\) & \(21846 \pm 1246\)\\
\bottomrule
\end{tabular}
|
212,243
|
2206.06720v2
| null |
\caption{Negative Log Likelihood and Continuous Ranked Probability Score results on \hyperref[para:large_regression]{large scale regression datasets.}}
|
\label{tab:full_large_regression}
|
\begin{tabular}{l*{9}{c}}
\toprule
\multirow{ 2.5}{*}{\textbf{NLL}}& \multicolumn{2}{c}{Single-layer} & \multicolumn{4}{c}{Ours} & \citetalias{salimbeni2017doubly} \\ \cmidrule(lr){2-3} \cmidrule(lr){4-7} \cmidrule(lr){8-8}
& SGP & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 & Best DGP \\
\midrule
Year & \(3.62 \pm 0.00 \) & \(3.74 \pm 0.00\) & \(3.68 \pm 0.00\) & \(3.64 \pm 0.00\) & \(3.64 \pm 0.00\) & \(3.63 \pm 0.00\) & \(\bm{3.59 \pm 0.00}\)\\
Airline & \(5.10 \pm 0.00\) & \(5.11 \pm 0.00\) & \(5.08 \pm 0.00\)& \(5.07 \pm 0.00\) & \(5.07 \pm 0.00\)& \(\bm{5.06 \pm 0.00}\)& \(5.07 \pm 0.00\)\\
Taxi & \(7.73 \pm 0.00\) & \(7.73 \pm 0.00\) & \(7.72 \pm 0.00\) & \(\bm{7.69 \pm 0.00}\) & \(7.72 \pm 0.00\) & \(7.70 \pm 0.00\) & \(7.73 \pm 0.00\)\\
\midrule \midrule
\multirow{ 2.5}{*}{\textbf{CRPS}}& \multicolumn{2}{c}{Single-layer} & \multicolumn{4}{c}{Ours} & \citetalias{salimbeni2017doubly} \\ \cmidrule(lr){2-3} \cmidrule(lr){4-7} \cmidrule(lr){8-8}
& SGP & VIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 & Best DGP \\
\midrule
Year & \(4.83 \pm 0.01 \) & \(5.45 \pm 0.01\) & \(5.13 \pm 0.02 \) & \(4.96 \pm 0.01 \) & \(4.93 \pm 0.01\) & \(4.91 \pm 0.02\) & \(\bm{4.680 \pm 0.01}\)\\
Airline & \(17.90 \pm 0.05\) & \(17.93 \pm 0.04\) & \(17.53 \pm 0.05\)& \(17.54 \pm 0.07\) & \(17.51 \pm 0.05\) & \(\bm{17.45 \pm 0.04}\)& \(\bm{17.47 \pm 0.03}\)\\
Taxi & \(283.79 \pm 0.19\) & \(284.22 \pm 0.20\) & \(282.09 \pm 0.32\) & \(\bm{274.65 \pm 0.68}\) & \(281.28 \pm 0.44\) & \(277.60 \pm 0.90\) & \(282.99 \pm 0.21\)\\
\bottomrule
\end{tabular}
|
212,244
|
2206.06720v2
| null |
\caption{Negative Log Likelihood, Root Mean Squared Error and Continuous Ranked Probability Score results on \hyperref[para:uci]{regression UCI benchmark datasets.}}
|
\label{tab:full_uci}
|
\begin{tabular}{l*{13}{c}}
\toprule
\multirow{ 2.5}{*}{\textbf{NLL}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{4}{c}{Ours} & \multicolumn{4}{c}{\citetalias{salimbeni2017doubly}} \\ \cmidrule(lr){2-5} \cmidrule(lr){6-9} \cmidrule(lr){10-13}
& SGP & VIP & VIP 200 & SIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 & DGP 2 & DGP 3 & DGP 4 & DGP 5\\
\midrule
\texttt{Boston} & \(2.62 \pm 0.05\) & \(2.76 \pm 0.05\)& \(2.69 \pm 0.03\)& \(2.72 \pm 0.03\) & \(2.85 \pm 0.09\)& \(\bm{2.59 \pm 0.06}\)& \(2.67 \pm 0.09\)& \(2.66 \pm 0.08\)& \(2.63 \pm 0.05\)& \(2.63 \pm 0.05\)& \(2.64 \pm 0.05\)& \(2.65 \pm 0.05\) \\
\texttt{Energy}& \(1.54 \pm 0.02\) & \(2.07 \pm 0.02\)& \(2.07 \pm 0.02\)& \(1.17 \pm 0.02\) &\(0.76 \pm 0.02\)& \(\bm{0.70 \pm 0.01}\)& \(\bm{0.70 \pm 0.01}\)& \(0.73 \pm 0.01\)& \(0.72 \pm 0.01\)& \(0.74 \pm 0.01\)& \(0.72 \pm 0.01\)& \(0.73 \pm 0.01\) \\
\texttt{Concrete}& \(3.16 \pm 0.01\) & \(3.45 \pm 0.02\)& \(3.48 \pm 0.01\)& \(3.60 \pm 0.05\) & \(3.24 \pm 0.04\)& \(3.20 \pm 0.05\)& \(\bm{3.03 \pm 0.02}\)& \(3.06 \pm 0.02\)& \(3.17 \pm 0.01\)& \(3.20 \pm 0.01\)& \(3.13 \pm 0.01\)& \(3.12 \pm 0.01\) \\
\texttt{Winered}& \(\bm{0.93 \pm 0.01}\) & \(0.94 \pm 0.01\)& \(0.96 \pm 0.01\)& \(0.97 \pm 0.01\) &\(0.94 \pm 0.01\)& \(0.94 \pm 0.01\)& \(0.94 \pm 0.01\)& \(0.95 \pm 0.01\)& \(0.94 \pm 0.01\)& \(0.94 \pm 0.01\)& \( 0.94 \pm 0.01\)& \(\bm{0.93 \pm 0.01}\) \\
\texttt{Power}& \(2.84 \pm 0.00\) & \(2.85 \pm 0.00\)& \( 2.86 \pm 0.00\)& \(2.84 \pm 0.00\) & \(2.82 \pm 0.01\)& \(2.81 \pm 0.00\)& \(\bm{2.79 \pm 0.01}\)& \(\bm{2.79 \pm 0.01}\)& \(2.81 \pm 0.01\)& \(2.80 \pm 0.00\)& \(2.80 \pm 0.00\)& \(2.80 \pm 0.01\) \\
\texttt{Protein}& \(2.93 \pm 0.00\) & \(3.03 \pm 0.00\)& \(3.03 \pm 0.00\)& \(3.00 \pm 0.00\) & \(2.93 \pm 0.00\)& \(2.89 \pm 0.00\)& \(2.88 \pm 0.00\)& \(2.86 \pm 0.00\)& \( 2.84 \pm 0.00\)& \(\bm{2.79 \pm 0.00}\)& \(\bm{2.79 \pm 0.00}\)& \(2.80 \pm 0.00\) \\
\texttt{Naval}& \(-6.11 \pm 0.06\) & \(-4.50 \pm 0.02\)& \(-4.31 \pm 0.00\)& \(-2.79 \pm 0.00\)& \(-5.89 \pm 0.02\)& \(-5.98 \pm 0.01\)& \(-5.90 \pm 0.01\)& \(-5.92 \pm 0.01\)& \(\bm{-6.35 \pm 0.09}\)& \(-6.21 \pm 0.04\)& \(-6.27 \pm 0.06\)& \(-6.21 \pm 0.08\) \\
\texttt{Kin8nm}& \( -0.91 \pm 0.00\) & \(-0.31 \pm 0.00\)& \(-0.25 \pm 0.00\)& \(-0.27 \pm 0.02\) &\(-1.00 \pm 0.00\)& \(-1.13 \pm 0.00\)& \(-1.15 \pm 0.00\)& \(-1.16 \pm 0.00\)& \(-1.29 \pm 0.00\)& \(-1.32 \pm 0.00\)& \(\bm{-1.33 \pm 0.00}\)& \(1.30 \pm 0.00\) \\
\midrule\midrule
\multirow{ 2.5}{*}{\textbf{RMSE}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{4}{c}{Ours} & \multicolumn{4}{c}{\citetalias{salimbeni2017doubly}} \\ \cmidrule(lr){2-5} \cmidrule(lr){6-9} \cmidrule(lr){10-13}
& SGP & VIP & VIP 200 & SIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 & DGP 2 & DGP 3 & DGP 4 & DGP 5\\
\midrule
\texttt{Boston} & \(\bm{3.48 \pm 0.17}\) & \(4.78 \pm 0.28\)& \(4.49 \pm 0.28\)& \(5.10 \pm 0.32\)& \(3.87 \pm 0.19\)& \(3.50 \pm 0.20\)& \(3.60 \pm 0.19\)& \(3.66 \pm 0.21\)& \(3.51 \pm 0.18\)& \(3.53 \pm 0.19\)& \(3.55 \pm 0.20\)& \(3.56 \pm 0.20\) \\
\texttt{Energy}& \(1.07 \pm 0.03\) & \(2.57 \pm 0.08\)& \(2.68 \pm 0.07\)& \(3.27 \pm 0.09\)& \(0.52 \pm 0.01\)& \(0.47 \pm 0.01\)& \(\bm{0.46 \pm 0.01}\)& \(0.47 \pm 0.01\)& \(\bm{0.46 \pm 0.01}\)& \(0.47 \pm 0.01\)& \(\bm{0.46 \pm 0.01}\)& \(\bm{0.46 \pm 0.01}\) \\
\texttt{Concrete}& \(5.84 \pm 0.12\) & \(7.75 \pm 0.15\)& \(8.06 \pm 0.16\)& \(8.70 \pm 0.43\) &\(6.01 \pm 0.16\)& \(5.68 \pm 0.18\)& \(\bm{5.13 \pm 0.12}\)& \(5.27 \pm 0.13\)& \(5.86 \pm 0.12\)& \(6.01 \pm 0.12\)& \(5.54 \pm 0.11\)& \(5.52 \pm 0.12\) \\
\texttt{Winered}& \(\bm{0.61 \pm 0.00}\) & \(0.62 \pm 0.00\)& \(0.63 \pm 0.00\)& \(0.64 \pm 0.00\)& \(0.62 \pm 0.00\)& \(0.62 \pm 0.00\)& \(0.62 \pm 0.00\)& \(0.62 \pm 0.00\)& \(0.62 \pm 0.00\)& \(0.62 \pm 0.00\)& \( 0.62 \pm 0.00\)& \(0.62 \pm 0.00\) \\
\texttt{Power}& \(4.15 \pm 0.03\) & \(4.21 \pm 0.03\)& \(4.22 \pm 0.03\)& \(4.14 \pm 0.03\) &\(4.06 \pm 0.04\)& \(4.01 \pm 0.04\)& \(3.97 \pm 0.04\)& \(\bm{3.95 \pm 0.04}\)& \(4.00 \pm 0.04\)& \(3.98 \pm 0.03\)& \(3.99 \pm 0.03\)& \(3.96 \pm 0.04\) \\
\texttt{Protein}& \(4.56 \pm 0.01\) & \(5.05 \pm 0.01\)& \(5.04 \pm 0.01\)& \(4.92 \pm 0.02\) &\(4.54 \pm 0.01\)& \( 4.40 \pm 0.01\)& \(4.33 \pm 0.01\)& \(4.26 \pm 0.01\)& \( 4.17 \pm 0.01\)& \(\bm{4.00 \pm 0.01}\)& \( 4.01 \pm 0.01\)& \(4.02 \pm 0.01\) \\
\texttt{Naval}& \(\bm{0.00 \pm 0.00}\) & \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(0.01 \pm 0.00\) &\(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\) \\
\texttt{Kin8nm}& \( 0.09 \pm 0.00\) & \(0.17 \pm 0.00\)& \(0.18 \pm 0.00\)& \(0.18 \pm 0.00\) &\(0.08 \pm 0.00\)& \(0.07 \pm 0.00\)& \(0.07 \pm 0.00\)& \(0.07 \pm 0.00\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\)& \(\bm{0.06 \pm 0.00}\) \\
\midrule \midrule
\multirow{ 2.5}{*}{\textbf{CRPS}}& \multicolumn{4}{c}{Single-layer} & \multicolumn{4}{c}{Ours} & \multicolumn{4}{c}{\citetalias{salimbeni2017doubly}} \\ \cmidrule(lr){2-5} \cmidrule(lr){6-9} \cmidrule(lr){10-13}
& SGP & VIP & VIP 200 & SIP & DVIP 2 & DVIP 3 & DVIP 4 & DVIP 5 & DGP 2 & DGP 3 & DGP 4 & DGP 5\\
\midrule
\texttt{Boston} & \(1.79 \pm 0.05\) & \(2.25 \pm 0.08\)& \(2.13 \pm 0.08\)& \(2.35 \pm 0.11\) &\(1.91 \pm .06\)& \(\bm{1.76 \pm 0.07}\)& \(1.81 \pm 0.07\)& \(1.78 \pm 0.06\)& \(1.79 \pm 0.05\)& \(1.80 \pm 0.06\)& \(1.80 \pm 0.06\)& \(1.81 \pm 0.06\) \\
\texttt{Energy}& \(0.62 \pm 0.01\) & \( 1.27 \pm 0.04\)& \(1.30 \pm 0.03\)& \(1.21 \pm 0.04\) &\(0.28 \pm 0.00\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\)& \(\bm{0.26 \pm 0.00}\) \\
\texttt{Concrete}& \(3.20 \pm 0.05\) & \(4.29 \pm 0.08\)& \(4.43 \pm 0.08\)& \(4.69 \pm 0.11\) &\(3.26 \pm 0.07\)& \( 3.03 \pm 0.09\)& \(\bm{2.74 \pm 0.05}\)& \(2.83 \pm 0.05\)& \(3.21 \pm 0.05\)& \(3.31 \pm 0.05\)& \(3.05 \pm 0.05\)& \(3.04 \pm 0.05\) \\
\texttt{Winered}& \(\bm{0.34 \pm 0.00}\) & \(\bm{0.34 \pm 0.00}\)& \(0.35 \pm 0.00\)& \(0.35 \pm 0.00\)&\(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\)& \(\bm{0.34 \pm 0.00}\) \\
\texttt{Power}& \(2.27 \pm 0.01\) & \(2.31 \pm 0.01\)& \(2.31 \pm 0.01\)& \(2.27 \pm 0.01\) &\(2.21 \pm 0.01\)& \(2.18 \pm 0.01\)& \(\bm{2.14 \pm 0.01}\)& \(\bm{2.14 \pm 0.01}\)& \(2.17 \pm 0.01\)& \(2.16 \pm 0.01\)& \(2.17 \pm 0.01\)& \(2.15 \pm 0.01\) \\
\texttt{Protein}& \(2.56 \pm 0.00\) & \(2.87 \pm 0.00\)& \(2.86 \pm 0.01\)& \(2.77 \pm 0.00\) &\(2.54 \pm 0.00\)& \( 2.43 \pm 0.00\)& \( 2.38 \pm 0.00\)& \(2.33 \pm 0.00\)& \( 2.31 \pm 0.00\)& \(\bm{2.19 \pm 0.00}\)& \(\bm{2.19 \pm 0.00}\)& \(2.20 \pm 0.00\) \\
\texttt{Naval}& \(\bm{0.00 \pm 0.00}\) & \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\) &\(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\)& \(\bm{0.00 \pm 0.00}\) \\
\texttt{Kin8nm}& \(0.05 \pm 0.00\) & \(0.09 \pm 0.0\)& \(0.10 \pm 0.00\)& \(0.10 \pm 0.00\) &\(0.04 \pm 0.00\)& \(0.04 \pm 0.00\)& \( 0.04 \pm 0.00\)& \(0.04 \pm 0.00\)& \( \bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\)& \(\bm{0.03 \pm 0.00}\) \\
\bottomrule
\end{tabular}
|
212,245
|
2212.08320v2
| null |
\caption{Classification results on ScanObjectNN. Ours$^1$: results trained with no data augmentation. Ours$^2$: results trained with simple point cloud rotation. \texttt{DA}: data augmentation is used during fine-tuning training.
The overall accuracy, \ie, OA (\%) is reported.
}
|
\label{tab:scanobjectnn}
|
\begin{tabular}{lccccc}
\toprule[0.95pt]
Method & \#Params(M) & \texttt{DA} & OBJ\_BG & OBJ\_ONLY & PB\_T50\_RS\\
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{Supervised Learning Only}}\\
\midrule[0.6pt]
PointNet~\citep{PointNet} & 3.5 & $\checkmark$ & 73.3 & 79.2 & 68.0\\
SpiderCNN~\citep{SpiderCNN} & - & $\checkmark$ & 77.1 & 79.5 & 73.7\\
PointNet++~\citep{PointNet++} & 1.5 & $\checkmark$ & 82.3 & 84.3 & 77.9\\
DGCNN~\citep{DGCNN} & 1.8 & $\checkmark$ & 82.8 & 86.2 & 78.1\\
PointCNN~\citep{PointCNN} & 0.6 & $\checkmark$ & 86.1 & 85.5 & 78.5\\
BGA-DGCNN~\citep{BGA} & 1.8 & $\checkmark$ & - & - & 79.7\\
BGA-PN++~\citep{BGA} & 1.5 & $\checkmark$ & - & - & 80.2\\
DRNet~\citep{DRNet} & - & $\checkmark$ & - & - & 80.3\\
GBNet~\citep{GBNet} & 8.8 & $\checkmark$ & - & - & 80.5\\
SimpleView~\citep{SimpleView} & - & $\checkmark$ & - & - & 80.5$\pm$0.3\\
PRANet~\citep{PRANet} & 2.3 & $\checkmark$ & - & - & 81.0\\
MVTN~\citep{MVTN} & - & $\checkmark$ & - & - &82.8\\
PointMLP~\citep{PointMLP} & 13.2 & $\checkmark$ & - & - & 85.4$\pm$0.3\\
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Full})}\\
\midrule[0.6pt]
Transformer~\citep{AttentionIsAllYouNeed} & 22.1 & $\checkmark$ & 79.86 & 80.55 & 77.24\\
OcCo~\citep{OcCo} & 22.1 & $\checkmark$ & 84.85 & 85.54 & 78.79\\
Point-BERT~\citep{PointBERT} & 22.1 & $\checkmark$ & 87.43 & 88.12 & 83.07\\
MaskPoint~\citep{MaskPoint} & 22.1 & $\checkmark$ & 89.30 & 88.10 & 84.30\\
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 90.02 & 88.29 & 85.18 \\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{91.22} & \textbf{89.16} & \textbf{85.81}\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{93.29} & \textbf{91.91} & \textbf{\textbf{88.21}}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 89.31$\pm$0.41 & 87.88$\pm$0.36 & 84.35$\pm$0.31\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{90.06}$\pm$0.56 & \textbf{89.02}$\pm$0.22 & \textbf{85.33}$\pm$0.27 \\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{92.48}$\pm$0.59 & \textbf{91.57}$\pm$0.37 & \textbf{87.88}$\pm$0.36\\
% Linear Probing
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-Linear})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 82.58$\pm$0.58 & 83.52$\pm$0.41 & 73.08$\pm$0.30\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{82.71}$\pm$0.45 & \textbf{84.34}$\pm$0.29 & \textbf{74.17}$\pm$0.05\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{85.20}$\pm$0.83 & \textbf{85.84}$\pm$0.15 & \textbf{76.31}$\pm$0.26\\
% MLP-3
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-$3$})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 84.29$\pm$0.55 & 85.24$\pm$0.67 & 77.34$\pm$0.12\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{85.67}$\pm$0.29 & \textbf{86.79}$\pm$0.30 & \textbf{78.89}$\pm$0.22\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{87.14}$\pm$0.22 & \textbf{88.90}$\pm$0.40 & \textbf{81.52}$\pm$0.19 \\
\bottomrule[0.95pt]
\end{tabular}
|
212,246
|
2212.08320v2
| null | null | null |
\begin{tabular}{lccccc}
\toprule[0.95pt]
Method & \#Params(M) & \texttt{DA} & OBJ\_BG & OBJ\_ONLY & PB\_T50\_RS\\
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{Supervised Learning Only}}\\
\midrule[0.6pt]
PointNet~\citep{PointNet} & 3.5 & $\checkmark$ & 73.3 & 79.2 & 68.0\\
SpiderCNN~\citep{SpiderCNN} & - & $\checkmark$ & 77.1 & 79.5 & 73.7\\
PointNet++~\citep{PointNet++} & 1.5 & $\checkmark$ & 82.3 & 84.3 & 77.9\\
DGCNN~\citep{DGCNN} & 1.8 & $\checkmark$ & 82.8 & 86.2 & 78.1\\
PointCNN~\citep{PointCNN} & 0.6 & $\checkmark$ & 86.1 & 85.5 & 78.5\\
BGA-DGCNN~\citep{BGA} & 1.8 & $\checkmark$ & - & - & 79.7\\
BGA-PN++~\citep{BGA} & 1.5 & $\checkmark$ & - & - & 80.2\\
DRNet~\citep{DRNet} & - & $\checkmark$ & - & - & 80.3\\
GBNet~\citep{GBNet} & 8.8 & $\checkmark$ & - & - & 80.5\\
SimpleView~\citep{SimpleView} & - & $\checkmark$ & - & - & 80.5$\pm$0.3\\
PRANet~\citep{PRANet} & 2.3 & $\checkmark$ & - & - & 81.0\\
MVTN~\citep{MVTN} & - & $\checkmark$ & - & - &82.8\\
PointMLP~\citep{PointMLP} & 13.2 & $\checkmark$ & - & - & 85.4$\pm$0.3\\
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Full})}\\
\midrule[0.6pt]
Transformer~\citep{AttentionIsAllYouNeed} & 22.1 & $\checkmark$ & 79.86 & 80.55 & 77.24\\
OcCo~\citep{OcCo} & 22.1 & $\checkmark$ & 84.85 & 85.54 & 78.79\\
Point-BERT~\citep{PointBERT} & 22.1 & $\checkmark$ & 87.43 & 88.12 & 83.07\\
MaskPoint~\citep{MaskPoint} & 22.1 & $\checkmark$ & 89.30 & 88.10 & 84.30\\
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 90.02 & 88.29 & 85.18 \\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{91.22} & \textbf{89.16} & \textbf{85.81}\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{93.29} & \textbf{91.91} & \textbf{\textbf{88.21}}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 89.31$\pm$0.41 & 87.88$\pm$0.36 & 84.35$\pm$0.31\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{90.06}$\pm$0.56 & \textbf{89.02}$\pm$0.22 & \textbf{85.33}$\pm$0.27 \\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{92.48}$\pm$0.59 & \textbf{91.57}$\pm$0.37 & \textbf{87.88}$\pm$0.36\\
% Linear Probing
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-Linear})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 82.58$\pm$0.58 & 83.52$\pm$0.41 & 73.08$\pm$0.30\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{82.71}$\pm$0.45 & \textbf{84.34}$\pm$0.29 & \textbf{74.17}$\pm$0.05\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{85.20}$\pm$0.83 & \textbf{85.84}$\pm$0.15 & \textbf{76.31}$\pm$0.26\\
% MLP-3
\midrule[0.6pt]
\multicolumn{6}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-$3$})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & 22.1 & $\checkmark$ & 84.29$\pm$0.55 & 85.24$\pm$0.67 & 77.34$\pm$0.12\\
\rowcolor{linecolor1}ACT (Ours$^1$) & 22.1 & $\times$ & \textbf{85.67}$\pm$0.29 & \textbf{86.79}$\pm$0.30 & \textbf{78.89}$\pm$0.22\\
\rowcolor{linecolor}ACT (Ours$^2$) & 22.1 & $\checkmark$ & \textbf{87.14}$\pm$0.22 & \textbf{88.90}$\pm$0.40 & \textbf{81.52}$\pm$0.19 \\
\bottomrule[0.95pt]
\end{tabular}
|
212,247
|
2212.08320v2
| null |
\caption{Classification results on the ModelNet40 dataset. The overall accuracy, \ie, OA (\%) is reported. \texttt{[ST]}: standard Transformer architecture.}
|
\label{tab:modelnet}
|
\begin{tabular}{lccc}
\toprule[0.95pt]
Method & \texttt{[ST]} & \#Point & OA (\%)\\
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Full})}\\
\midrule[0.6pt]
Transformer~\citep{AttentionIsAllYouNeed} & $\checkmark$ & 1k P & 91.4\\
Transformer~\citep{AttentionIsAllYouNeed} & $\checkmark$ & 4k P & 91.2\\
OcCo~\citep{OcCo} & $\checkmark$ & 1k P & 92.1\\
OcCo~\citep{OcCo} & $\checkmark$ & 4k P & 92.2\\
Point-BERT~\citep{PointBERT} & $\checkmark$ & 1k P & 93.2\\
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & \textbf{93.8}\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{93.7}\\
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 93.12$\pm$0.25\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{93.50}$\pm$\textbf{0.08}\\
% Linear Probing
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-Linear})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 91.22$\pm$0.26\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{91.36}$\pm$\textbf{0.17}\\
% MLP-3
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-$3$})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 92.33$\pm$0.09\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{92.69}$\pm$\textbf{0.18} \\
\bottomrule[0.95pt]
\end{tabular}
|
212,248
|
2212.08320v2
| null | null | null |
\begin{tabular}{lccc}
\toprule[0.95pt]
Method & \texttt{[ST]} & \#Point & OA (\%)\\
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{Supervised Learning Only}}\\
\midrule[0.6pt]
PointNet~\citep{PointNet} & - & 1k P & 89.2\\
PointNet++~\citep{PointNet++} & - & 1k P & 90.7\\
PointNet++~\citep{PointNet++} & - & 5k P+N & 91.9\\
PointCNN~\citep{PointCNN} & - & 1k P & 92.5\\
PointConv~\citep{PointConv}& - & 1k P+N & 92.5\\
KPConv~\citep{KPConv} & - & 1k P & 92.9\\
DGCNN~\citep{DGCNN} & - & 1k P & 92.9\\
RS-CNN~\citep{RSCNN} & - & 1k P & 92.9\\
DensePoint~\citep{DensePoint19} & - & 1k P & 93.2\\
PointASNL~\citep{PointASNL20} & - & 1k P & 92.9\\
PosPool~\citep{PosPool20} & - & 5k P & 93.2\\
DRNet~\citep{DRNet} & - & 1k P & 93.1\\
\midrule[0.6pt]
Point Trans.~\citep{PointTrans20} & $\times$ & 1k P & 92.8\\
PCT~\citep{PCT} & $\times$ & 1k P & 93.2\\
PointTransformer~\citep{PointTransformer} & $\times$ & 1k P & 93.7\\
NPCT~\citep{PCT} & $\checkmark$ & 1k P & 91.0\\
\bottomrule[0.95pt]
\end{tabular}
|
212,249
|
2212.08320v2
| null | null | null |
\begin{tabular}{lccc}
\toprule[0.95pt]
Method & \texttt{[ST]} & \#Point & OA (\%)\\
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Full})}\\
\midrule[0.6pt]
Transformer~\citep{AttentionIsAllYouNeed} & $\checkmark$ & 1k P & 91.4\\
Transformer~\citep{AttentionIsAllYouNeed} & $\checkmark$ & 4k P & 91.2\\
OcCo~\citep{OcCo} & $\checkmark$ & 1k P & 92.1\\
OcCo~\citep{OcCo} & $\checkmark$ & 4k P & 92.2\\
Point-BERT~\citep{PointBERT} & $\checkmark$ & 1k P & 93.2\\
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & \textbf{93.8}\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{93.7}\\
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 93.12$\pm$0.25\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{93.50}$\pm$\textbf{0.08}\\
% Linear Probing
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-Linear})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 91.22$\pm$0.26\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{91.36}$\pm$\textbf{0.17}\\
% MLP-3
\midrule[0.6pt]
\multicolumn{4}{c}{\textit{with Self-Supervised Representation Learning} ({\scshape Mlp-$3$})}\\
\midrule[0.6pt]
Point-MAE~\citep{PointMAE} & $\checkmark$ & 1k P & 92.33$\pm$0.09\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & 1k P & \textbf{92.69}$\pm$\textbf{0.18} \\
\bottomrule[0.95pt]
\end{tabular}
|
212,250
|
2212.08320v2
| null |
\caption{Ablation study on the depth of the pretraining decoder.}
|
\label{tab:decoder_depth}
|
\begin{tabular}{cc}
\\
\toprule[0.95pt]
Dec. Depth & OA (\%)$\uparrow$\\
\midrule[0.6pt]
0 & 83.69\\
1 & 85.11\\
2 & \cellcolor{linecolor}{\textbf{85.33}} \\
4 & 84.98\\
\bottomrule[0.95pt]
\end{tabular}
|
212,251
|
2212.08320v2
| null |
\caption{Ablation study on different training strategies of the dVAE tokenizer. The F-Score, Chamfer distance using L1-norm and L2-norm, \ie, CD-$\ell_1$ and CD-$\ell_2$ are reported.}
|
\label{tab:dvae_ablation}
|
\begin{tabular}{lcccccc}
\toprule[0.95pt]
Methods & Num. of Prompt & Prompt Type & Freeze & F-Score$\uparrow$ & CD-$\ell_1\downarrow$ & CD-$\ell_2\downarrow$\\
\midrule[0.6pt]
Point-BERT dVAE & N/A & N/A & N/A & 0.166 & 25.933 & 2.697\\
\midrule[0.6pt]
DeiT-B dVAE & 0 & N/A & $\times$ & 0.175 & 24.589 & 2.380 \\
DeiT-B dVAE & 0 & N/A & $\checkmark$ & 0.180 & 24.090 & 2.274 \\
DeiT-B dVAE & 32 & shallow & $\checkmark$ & 0.188 & 23.769 & 2.196 \\
DeiT-B dVAE & 32 & deep & $\checkmark$ & 0.189 & 23.873 & 2.173\\
DeiT-B dVAE & 64 & deep & $\checkmark$ & 0.189 & 23.229 & 2.127\\
ViT-B dVAE & 64 & deep & $\checkmark$ & 0.193 & 23.524 & 2.110\\
\bottomrule[0.95pt]
\end{tabular}
|
212,252
|
2212.08320v2
| null |
\caption{Study of applying our method as auxiliary knowledge distillation during pretraining.}
|
\label{Tab:ablation_kd}
|
\begin{tabular}{lcc}
\toprule[0.95pt]
Method & KD & OA (\%)$\uparrow$\\
\midrule[0.6pt]
Point-MAE & $\times$ & 85.18\\
Our Impl. & $\checkmark$ & \cellcolor{linecolor}\textbf{86.05}\\
Our Impl. & $\times$ & 84.35$\pm$0.31\\
Our Impl. & $\checkmark$ & \cellcolor{linecolor}\textbf{84.96}$\pm$0.58\\
\bottomrule[0.95pt]
\end{tabular}
|
212,253
|
2212.08320v2
| null |
\caption{Study on the effect of pretrained image Transformer-based 3D Autoencoder.}
|
\label{Tab:ablation_vitindvae}
|
\begin{tabular}{lcc}
\toprule[0.95pt]
Teacher & Target & OA (\%)$\uparrow$\\
\midrule[0.6pt]
Point-BERT & Point-BERT & 83.07\\
Point-BERT & Ours & 83.69\\
Ours & Point-BERT & 82.51\\
Ours & Ours & \cellcolor{linecolor}\textbf{85.81}\\
\bottomrule[0.95pt]
\end{tabular}
|
212,254
|
2212.08320v2
| null |
\caption{Study of applying our method as auxiliary knowledge distillation during pretraining.}
|
\label{Tab:ablation_kd}
|
\begin{tabular}{lcc}
\toprule[0.95pt]
Method & KD & OA (\%)$\uparrow$\\
\midrule[0.6pt]
Point-MAE & $\times$ & 85.18\\
Our Impl. & $\checkmark$ & \cellcolor{linecolor}\textbf{86.05}\\
Our Impl. & $\times$ & 84.35$\pm$0.31\\
Our Impl. & $\checkmark$ & \cellcolor{linecolor}\textbf{84.96}$\pm$0.58\\
\bottomrule[0.95pt]
\end{tabular}
|
212,282
|
2203.07720v3
| null | null |
\label{tab:temporal}
|
\begin{tabular}{l|lll}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
& R@1 & R@5 & R@10 \\ \hline
Time attention & 35.9 & 60.8 & 71.8 \\ \hline
Temporal position & 36.0 & 61.0 & 71.8 \\ \hline
\end{tabular}
|
212,255
|
2212.08320v2
| null |
\caption{Study of different positional embeddings for 2D image transformer in dVAE model. (a) N/A: no positional embedding is used. (b) 2D/$z$: positional embedding with only 2D $xy$ plane coordinates. (c) 3D: positional embedding with all 3D $xyz$ coordinates. The F-Score, Chamfer distance using L1-norm and L2-norm, \ie, CD-$\ell_1$ and CD-$\ell_2$, and OA on ScanObjectNN are reported.}
|
\label{Tab:postion_embed}
|
\begin{tabular}{lccccc}
\toprule[0.95pt]
Methods & pos embed & F-Score$\uparrow$ & CD-$\ell_1\downarrow$ & CD-$\ell_2\downarrow$ & OA (\%)$\uparrow$\\
\midrule[0.6pt]
ViT-B dVAE & N/A & 0.166 & 25.918 & 2.698 & 84.21$\pm$0.45\\
ViT-B dVAE & 2D/$z$ & 0.184 & 24.135 & 2.259 & 85.10$\pm$0.45\\
ViT-B dVAE & 3D & 0.193 & 23.524 & 2.110 & 85.33$\pm$0.27\\
\bottomrule[0.95pt]
\end{tabular}
|
212,256
|
2212.08320v2
| null |
\caption{3D object detection on the ScanNetV2 dataset. The detection performance using mean Average Precision (mAP) at two different IoU thresholds of $0.50$ and $0.25$, \ie, $\text{AP}_{50}$ and $\text{AP}_{25}$ are reported. $xyz$: point cloud coordinates are used.
}
|
\label{tab:3d_det}
|
\begin{tabular}{lcccc}
\toprule[0.95pt]
Method & SSL & Input & $\text{AP}_{50}$ & $\text{AP}_{25}$\\
\midrule[0.6pt]
VoteNet~\citep{VoteNet} & $\times$ & $xyz$ & 33.5 & 58.6\\
PointContrast~\citep{PointContrast20} & $\checkmark$ & $xyz$ & 38.0 & 59.2\\
STRL~\citep{STRL} & $\checkmark$ & $xyz$ & 38.4 & 59.5\\
RandomRooms~\citep{RandomRooms21} & $\checkmark$ & $xyz$ & 36.2 & 61.3\\
DepthContrast~\citep{DepthContrast21} & $\checkmark$ & $xyz$ & - & 61.3\\
\midrule[0.6pt]
3DETR~\citep{ThreeDETR} & $\times$ & $xyz$ & 37.9 & 62.1\\
Point-BERT~\citep{PointBERT} & $\checkmark$ & $xyz$ & 38.3 & 61.0\\
MaskPoint~\citep{MaskPoint} & $\checkmark$ & $xyz$ & 40.6 & 63.4\\
\rowcolor{linecolor}ACT (Ours) & $\checkmark$ & $xyz$ & \textbf{42.1} & \textbf{63.8}\\
\bottomrule[0.95pt]
\end{tabular}
|
212,257
|
2212.08320v2
| null |
\caption{Comparison to supervised cross-modal 3D representation learning method on ScanObjectNN. Overall accuracy, \ie, OA (\%) is reported.}
|
\label{tab:comparison_supversied}
|
\begin{tabular}{lccc}
\toprule[0.95pt]
Method & Backbone &\#Params (M) & OA (\%)$\uparrow$\\
\midrule[0.6pt]
P2P~\citep{P2P} & ResNet101 & 42.7 & 87.4\\
P2P~\citep{P2P} & HorNet & 195.8 & 89.3\\
\rowcolor{linecolor} ACT (Ours) & Transformer & 22.1 & 88.2\\
\bottomrule[0.95pt]
\end{tabular}
|
212,258
|
2212.08320v2
| null |
\caption{Part segmentation results on the ShapeNetPart dataset. The mean IoU across all categories, \ie, Cls. mIoU, the mean IoU across all instances, \ie, Ins. mIoU (\%), and IoU (\%) for each category are reported. The best results are \textbf{bolded} and the second best results are \underline{underlined}. }
|
\label{Tab:partseg_category}
|
\begin{tabular}{lcccccccccccccccccc}
\toprule[0.95pt]
Method & \makecell{ Cls.\\mIoU} & \makecell{ Ins.\\mIoU} & aero & bag & cap & car & chair &\makecell{aerp-\\hone} & guitar & knife & lamp & laptop &\makecell{ motor-\\bike} &mug &pistol &rocket &\makecell{ skate-\\board} & table \\
\midrule[0.6pt]
PointNet & 80.39 & 83.7 & 83.4 & 78.7 & 82.5 & 74.9 & 89.6 & 73.0 & 91.5 & 85.9 & 80.8 & 95.3 & 65.2 & 93.0 & 81.2 & 57.9 & 72.8 & 80.6 \\
PointNet++ & 81.85 & 85.1 & 82.4 & 79.0 & 87.7 & 77.3 & 90.8 & 71.8 & 91.0 & 85.9 & 83.7 & 95.3 & 71.6 & 94.1 & 81.3 & 58.7 & 76.4 & 82.6 \\
DGCNN & 82.33 & 85.2 & 84.0 & 83.4 & 86.7 & 77.8 & 90.6 & 74.7 & 91.2 & 87.5 & 82.8 & 95.7 & 66.3 & 94.9 & 81.1 & 63.5 & 74.5 & 82.6 \\
\midrule[0.6pt]
Transformer & 83.42 & 85.1 & 82.9 & 85.4 & 87.7 & 78.8 & 90.5 & \underline{80.8} & 91.1 & 87.7 & 85.3 & 95.6 & 73.9 & 94.9 & 83.5 & 61.2 & 74.9 & 80.6 \\
OcCo & 83.42 & 85.1 & 83.3 & 85.2 & \underline{88.3} & 79.9 & 90.7 & 74.1 & 91.9 & 87.6 & 84.7 & 95.4 & 75.5 & 94.4 & 84.1 & 63.1 & 75.7 & 80.8 \\
Point-BERT & 84.11 & 85.6 & \underline{84.3} & 84.8 & 88.0 & 79.8 & 91.0 & \textbf{81.7} & 91.6 & \textbf{87.9} & 85.2 & 95.6 & \underline{75.6} & 94.7 & 84.3 & 63.4 & 76.3 & 81.5 \\
Point-MAE & 84.19 & \textbf{86.1} & \underline{84.3} & 85.0 & \underline{88.3} & \underline{80.5} & \textbf{91.3} & 78.5 & \underline{92.1} & 87.4 & \underline{86.1} & \textbf{96.1} & 75.2 & 94.6 & 84.7 & 63.5 & \underline{77.1} & \textbf{82.4}\\
\rowcolor{linecolor}ACT (Ours) & \textbf{84.66} & \textbf{86.14} & \textbf{85.2} & \underline{85.2} & \textbf{88.8} & \textbf{81.2} & \textbf{91.3} & 79.4 & \textbf{92.2} & \textbf{87.9} & 85.8 & \underline{96.0} & 75.5 & \textbf{95.5} & \underline{85.2} & \textbf{66.6} & \textbf{77.7} & 81.5\\
\bottomrule[0.95pt]
\end{tabular}
|
212,259
|
2203.07720v3
| null | null |
\label{tab:pretrain-time}
|
\begin{tabular}{l|c|c|c}
\hline
\textbf{Method} & \textbf{Data} & \textbf{GPU Hrs} & \textbf{R@1} \\ \hline
Frozen~\cite{Bain_2021_ICCV} & 5.8M & 4800 & 31.0 \\
UniVL~\cite{luo2020univl} & 132M & 2496 & 21.2 \\
HERO~\cite{li2020hero} & 7.6M & 8064 & 20.5 \\
VIOLET~\cite{fu2021violet} & 185.8M & 2240 & 34.5 \\
ClipBERT~\cite{lei2021less} & 5.6M & 768 & 22.0 \\ \hline
Ours (4F/1.0) & 5.8M & 1600 & 36.3 \\
Ours (1F/1.0) & 5.8M & 800 & 36.0 \\
Ours (1F/0.5) & 2.9M & 416 & 36.4 \\
Ours (1F/0.2) & 1.2M & 104 & 34.6 \\ \hline
\end{tabular}
|
212,260
|
2203.07720v3
| null | null |
\label{tab:infer-ti}
|
\begin{tabular}{l|cc|cc}
\hline
\multirow{2}{*}{Parts} & \multicolumn{2}{c|}{12000 Samples} & \multicolumn{2}{c}{1000 Samples} \\ \cline{2-5}
& \multicolumn{1}{c|}{T$\rightarrow$V} & V$\rightarrow$T & \multicolumn{1}{c|}{T$\rightarrow$V} & V$\rightarrow$T \\ \hline
Language & \multicolumn{1}{c|}{5.6ms} & - & \multicolumn{1}{c|}{4.8ms} & - \\
Video & \multicolumn{1}{c|}{-} & 17.6ms & \multicolumn{1}{c|}{-} & 16.4ms \\
Detector & \multicolumn{1}{c|}{-} & 103.2ms & \multicolumn{1}{c|}{-} & 103.2ms \\ \hline
Rank & \multicolumn{1}{c|}{372.2ms} & 345.8ms & \multicolumn{1}{c|}{31.8ms} & 30.0ms \\ \hline
\end{tabular}
|
212,261
|
2203.07720v3
| null | null |
\label{tab:infer-flops}
|
\begin{tabular}{l|cc|cc}
\hline
\multicolumn{1}{c|}{\multirow{2}{*}{Parts}} & \multicolumn{2}{c|}{Frozen} & \multicolumn{2}{c}{Ours} \\ \cline{2-5}
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{FLOPs} & Param & \multicolumn{1}{c|}{FLOPs} & Param \\ \hline
Language & \multicolumn{1}{c|}{9.7B} & 66M & \multicolumn{1}{c|}{9.7B} & 66M \\
Video & \multicolumn{1}{c|}{590B} & 121.4M & \multicolumn{1}{c|}{41.8B} & 86.8M \\
Detector & \multicolumn{1}{c|}{-} & - & \multicolumn{1}{c|}{455.2B} & 60.6M \\ \hline
All & \multicolumn{1}{c|}{599.7B} & 187.4M & \multicolumn{1}{c|}{506.7B} & 213.4M \\ \hline
\end{tabular}
|
212,262
|
2203.07720v3
| null |
\caption{Effectiveness of our region-word alignment. ``Base'' denotes the baseline that adopts the same region features as input while only using video-sentence global alignment as the objective.}
|
\label{tab:rwa-iclr}
|
\begin{tabular}{llll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{Method}} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
\multicolumn{1}{l|}{} & R@1 & R@5 & R@10 \\ \hline
\multicolumn{1}{l|}{Frozen} & 31.0 & 59.5 & 70.5 \\
\multicolumn{1}{l|}{Frozen+RWA} & 35.1 & 60.7 & 70.9 \\ \hline
\multicolumn{4}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Frozen} & 18.7 & 39.5 & 51.6 \\
\multicolumn{1}{l|}{Frozen+RWA} & 21.7 & 42.2 & 52.2 \\ \hline
\end{tabular}
|
212,263
|
2203.07720v3
| null |
\caption{Results on MSRVTT retrieval with Base.}
|
\label{tab:rwa-base}
|
\begin{tabular}{llll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{Method}} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
\multicolumn{1}{l|}{} & R@1 & R@5 & R@10 \\ \hline
\multicolumn{1}{l|}{Base} & 22.5 & 48.1 & 59.5 \\
\multicolumn{1}{l|}{Ours(Base+RWA)} & 36.0 & 61.0 & 71.8 \\ \hline
\multicolumn{4}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Base} & 16.1 & 34.9 & 44.2 \\
\multicolumn{1}{l|}{Ours(Base+RWA)} & 22.5 & 48.1 & 59.5 \\ \hline
\end{tabular}
|
212,264
|
2203.07720v3
| null |
\caption{Results on MSRVTT retrieval with Frozen.}
|
\label{tab:rwa-frozen}
|
\begin{tabular}{llll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{Method}} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
\multicolumn{1}{l|}{} & R@1 & R@5 & R@10 \\ \hline
\multicolumn{1}{l|}{Frozen} & 31.0 & 59.5 & 70.5 \\
\multicolumn{1}{l|}{Frozen+RWA} & 35.1 & 60.7 & 70.9 \\ \hline
\multicolumn{4}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Frozen} & 18.7 & 39.5 & 51.6 \\
\multicolumn{1}{l|}{Frozen+RWA} & 21.7 & 42.2 & 52.2 \\ \hline
\end{tabular}
|
212,265
|
2203.07720v3
| null | null |
\label{tab:rwa-com}
|
\begin{tabular}{l|lll}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
& R@1 & R@5 & R@10 \\ \hline
Base+TACo & 29.8 & 60.1 & 70.7 \\
Base+FILIP & 27.8 & 55.0 & 65.1 \\ \hline
Ours(Base+RWA) & \textbf{36.0} & \textbf{61.0} & 71.8 \\ \hline
\end{tabular}
|
212,266
|
2203.07720v3
| null | null |
\label{tab:rwa-abl}
|
\begin{tabular}{l|lll}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
& R@1 & R@5 & R@10 \\ \hline
Ours w/o Refinement & 34.5 & 58.3 & 70.2 \\
Select Token & 32.8 & 55.5 & 69.2 \\ \hline
Ours & 36.0 & 61.0 & 71.8 \\ \hline
\end{tabular}
|
212,267
|
2203.07720v3
| null |
\caption{Comparisons with state-of-the-art results on text-to-video retrieval.}
|
\label{tab:video_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Text$\rightarrow$Video}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{VSE} & 12.3 & 30.1 & 42.3 & 14.0 \\
\multicolumn{1}{l|}{VSE++} & 15.4 & 39.6 & 53.0 & 9.0 \\
\multicolumn{1}{l|}{MCues} & 20.3 & 47.8 & 61.1 & 6.0 \\
\multicolumn{1}{l|}{CE} & 19.8 & 49.0 & 63.8 & 6.0 \\
\multicolumn{1}{l|}{SupportSet} & 28.4 & 60.0 & 72.9 & 4.0 \\
\multicolumn{1}{l|}{Frozen} & 33.7 & 64.7 & 76.3 & 3.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{50.9} & \textbf{78.9} & \textbf{87.0} & \textbf{1.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 41.6 & 69.8 & 80.5 & 2.0 \\ \hline
\end{tabular}
|
212,268
|
2203.07720v3
| null |
\caption{LSMDC retrieval}
|
\label{tab:lsmdc_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Text$\rightarrow$Video}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{JSFusion} & 9.1 & 21.2 & 34.1 & 36.0 \\
\multicolumn{1}{l|}{MEE} & 9.3 & 25.1 & 33.4 & 27.0 \\
\multicolumn{1}{l|}{CE} & 11.2 & 26.9 & 34.8 & 25.3 \\
\multicolumn{1}{l|}{MMT} & 12.9 & 29.2 & 38.8 & 19.3 \\
\multicolumn{1}{l|}{AVLNet} & 17.0 & 38.0 & 48.6 & 11.0 \\
\multicolumn{1}{l|}{Dig} & 15.8 & 34.1 & 43.6 & 14.3 \\
\multicolumn{1}{l|}{Frozen} & 15.0 & 34.1 & 39.8 & 20.0 \\
\multicolumn{1}{l|}{VTMCE} & 14.9 & 33.2 & - & 19.0 \\
\multicolumn{1}{l|}{MDMMT} & 18.8 & 38.5 & 47.9 & 12.3 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{25.2} & \textbf{45.5} & \textbf{54.5} & \textbf{7.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 14.3 & 25.8 & 32 & 49.5 \\ \hline
\end{tabular}
|
212,269
|
2203.07720v3
| null |
\caption{MSRVTT retrieval}
|
\label{tab:msrvtt_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Text$\rightarrow$Video}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{MMT} & 26.6 & 57.1 & 69.6 & 4.0 \\
\multicolumn{1}{l|}{ActBERT} & 16.3 & 42.8 & 56.9 & 10.0 \\
\multicolumn{1}{l|}{SupportSet} & 30.1 & 58.5 & 69.3 & 3.0 \\
\multicolumn{1}{l|}{AVLNet} & 27.1 & 55.6 & 66.6 & 4.0 \\
\multicolumn{1}{l|}{TACo} & 29.6 & 59.7 & \textbf{72.7} & 4.0 \\
\multicolumn{1}{l|}{ClipBERT} & 22.0 & 46.8 & 59.9 & 6.0 \\
\multicolumn{1}{l|}{Frozen} & 31.0 & 59.5 & 70.5 & 3.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{36.0} & \textbf{61.0} & 71.8 & \textbf{3.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{SupportSet} & 12.7 & 27.5 & 36.2 & 24.0 \\
\multicolumn{1}{l|}{Frozen} & 18.7 & 39.5 & 51.6 & 10.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{24.0} & \textbf{44.0} & \textbf{52.6} & \textbf{8.0} \\ \hline
\end{tabular}
|
212,270
|
2203.07720v3
| null |
\caption{DiDeMo retrieval}
|
\label{tab:didemo_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Text$\rightarrow$Video}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{S2VT} & 11.9 & 33.6 & - & 13.0 \\
\multicolumn{1}{l|}{FSE} & 13.9 & 36.0 & - & 11.0 \\
\multicolumn{1}{l|}{CE} & 22.6 & 51.1 & - & 5.0 \\
\multicolumn{1}{l|}{ClipBERT} & 20.4 & 44.5 & 56.7 & 7.0 \\
\multicolumn{1}{l|}{Frozen} & 31.0 & 59.8 & 72.4 & 3.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{41.4} & \textbf{67.6} & \textbf{77.6} & \textbf{2.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Frozen} & 21.1 & 46.0 & 56.2 & 7.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{29.6} & \textbf{53.0} & \textbf{65.1} & \textbf{4.0} \\ \hline
\end{tabular}
|
212,283
|
2203.07720v3
| null |
\caption{Results on MSRVTT retrieval with more data.}
|
\label{tab:more-data}
|
\begin{tabular}{c|c|lll}
\hline
\multirow{2}{*}{Data} & \multirow{2}{*}{Data Scale} & \multicolumn{3}{c}{Text$\rightarrow$Video} \\
& & R@1 & R@5 & R@10 \\ \hline
WebVid & 2.5M & {27.6} & {50.4} & {61.7} \\
WebVid+CC3M & 5.8M & 36.0 & 61.0 & 71.8 \\
CC3M+CC7M & 10.3M & 38.7 & 63.8 & \textbf{73.9} \\
WebVid+CC3M+CC7M & 12.8M & \textbf{39.6} & \textbf{64.8} & 73.8 \\ \hline
\end{tabular}
|
212,271
|
2203.07720v3
| null |
\caption{MSVD retrieval}
|
\label{tab:msvd_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Text$\rightarrow$Video}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{VSE} & 12.3 & 30.1 & 42.3 & 14.0 \\
\multicolumn{1}{l|}{VSE++} & 15.4 & 39.6 & 53.0 & 9.0 \\
\multicolumn{1}{l|}{MCues} & 20.3 & 47.8 & 61.1 & 6.0 \\
\multicolumn{1}{l|}{CE} & 19.8 & 49.0 & 63.8 & 6.0 \\
\multicolumn{1}{l|}{SupportSet} & 28.4 & 60.0 & 72.9 & 4.0 \\
\multicolumn{1}{l|}{Frozen} & 33.7 & 64.7 & 76.3 & 3.0 \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{50.9} & \textbf{78.9} & \textbf{87.0} & \textbf{1.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 41.6 & 69.8 & 80.5 & 2.0 \\ \hline
\end{tabular}
|
212,272
|
2203.07720v3
| null |
\caption{Comparisons with state-of-the-art results on video-to-text retrieval.}
|
\label{tab:v2t_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Video$\rightarrow$Text}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{VSE} & 15.8 & 30.2 & 41.4 & 14.0 \\
\multicolumn{1}{l|}{VSE++} & 21.2 & 43.4 & 52.2 & 9.0 \\
\multicolumn{1}{l|}{MCues} & 31.5 & 51.0 & 61.5 & 5.0 \\
\multicolumn{1}{l|}{CE} & - & - & - & - \\
\multicolumn{1}{l|}{SupportSet} & 28.7 & 60.8 & - & 2.0 \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{43.9} & \textbf{74.6} & \textbf{83.8} & \textbf{2.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 41.3 & 67.8 & 78.3 & 2.0 \\ \hline
\end{tabular}
|
212,273
|
2203.07720v3
| null |
\caption{LSMDC retrieval}
|
\label{tab:lsmdc_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Video$\rightarrow$Text}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{JSFusion} & - & - & - & - \\
\multicolumn{1}{l|}{MEE} & - & - & - & - \\
\multicolumn{1}{l|}{CE} & - & - & - & - \\
\multicolumn{1}{l|}{MMT} & - & 28.6 & - & 20.0 \\
\multicolumn{1}{l|}{AVLNet} & - & - & - & - \\
\multicolumn{1}{l|}{Dig} & 14.3 & 33.7 & 43.6 & 15.5 \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{VTMCE} & 15.3 & 34.1 & - & 18.0 \\
\multicolumn{1}{l|}{MDMMT} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{23.8} & \textbf{43.6} & \textbf{54.1} & \textbf{8.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 16.0 & 29.3 & 35.5 & 40.5 \\ \hline
\end{tabular}
|
212,274
|
2203.07720v3
| null |
\caption{MSRVTT retrieval}
|
\label{tab:msrvtt_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Video$\rightarrow$Text}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{MMT} & 27.0 & 57.5 & 69.7 & 3.7 \\
\multicolumn{1}{l|}{ActBERT} & - & - & - & - \\
\multicolumn{1}{l|}{SupportSet} & 28.5 & \textbf{58.6} & 71.6 & \textbf{3.0} \\
\multicolumn{1}{l|}{AVLNet} & - & - & - & - \\
\multicolumn{1}{l|}{TACo} & - & - & - & - \\
\multicolumn{1}{l|}{ClipBERT} & - & - & - & - \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours} } & \textbf{32.7} & 58.4 & 68.4 & 4.0 \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{SupportSet} & 8.7 & 23.0 & 31.1 & 31.0 \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{27.1} & \textbf{45.8} & \textbf{54.5} & \textbf{7.5} \\ \hline
\end{tabular}
|
212,291
|
2303.14969v1
| null |
\caption{Number of task-specific and shared parameters for a single-channel task (in million).}
|
\label{tab:number_of_parameters}
|
\begin{tabular}{cccc}
\toprule
Model & Task-Specific & Shared \\
\midrule
DPT (supervised learning) & 110.55 & 0 \\
InvPT (multi-task learning) & 24.57 & 106.75 \\
Ours (few-shot learning) & 0.0703 & 202.95 \\
\bottomrule
\end{tabular}
|
212,292
|
2303.14969v1
| null |
\caption{MACs of Ours and DPT on a single-query inference for a single-channel task.}
|
\label{tab:computation_cost}
|
\begin{tabular}{ccc}
\toprule
Model & MACs (G) \\
\midrule
DPT & 30.15 \\
Ours after encoding support (10-shot) & 38.79 \\
\bottomrule
\end{tabular}
|
212,275
|
2203.07720v3
| null |
\caption{DiDeMo retrieval}
|
\label{tab:didemo_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Video$\rightarrow$Text}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{S2VT} & 13.2 & 33.6 & - & 15.0 \\
\multicolumn{1}{l|}{FSE} & 13.1 & 33.9 & - & 12.0 \\
\multicolumn{1}{l|}{CE} & 22.5 & 52.3 & - & 5.0 \\
\multicolumn{1}{l|}{ClipBERT} & - &- & - &- \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{36.0} & \textbf{65.7} & \textbf{75.6} & \textbf{3.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{Ours} & 29.8 & 53.1 & 63.6 & 4.0 \\ \hline
\end{tabular}
|
212,276
|
2203.07720v3
| null |
\caption{MSVD retrieval}
|
\label{tab:msvd_retrieval}
|
\begin{tabular}{lllll}
\hline
\multicolumn{1}{l|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c}{\textbf{Video$\rightarrow$Text}} \\
\multicolumn{1}{l|}{} & \textbf{R@1} & \textbf{R@5} & \textbf{R@10} & \textbf{MedR} \\ \hline
\multicolumn{1}{l|}{VSE} & 15.8 & 30.2 & 41.4 & 14.0 \\
\multicolumn{1}{l|}{VSE++} & 21.2 & 43.4 & 52.2 & 9.0 \\
\multicolumn{1}{l|}{MCues} & 31.5 & 51.0 & 61.5 & 5.0 \\
\multicolumn{1}{l|}{CE} & - & - & - & - \\
\multicolumn{1}{l|}{SupportSet} & 28.7 & 60.8 & - & 2.0 \\
\multicolumn{1}{l|}{Frozen} & - & - & - & - \\
\multicolumn{1}{l|}{\textbf{Ours}} & \textbf{43.9} & \textbf{74.6} & \textbf{83.8} & \textbf{2.0} \\ \hline
\multicolumn{5}{l}{Zero-shot} \\ \hline
\multicolumn{1}{l|}{Ours} & 41.3 & 67.8 & 78.3 & 2.0 \\ \hline
\end{tabular}
|
212,277
|
2203.07720v3
| null |
\caption{Comparisons with state-of-the-art results on video question answering.}
|
\label{tab:video_qa}
|
\begin{tabular}{l|c|c}
\hline
\textbf{Method} & \multicolumn{1}{l|}{\textbf{MSRVTT}} & \multicolumn{1}{l}{\textbf{MSVD}} \\ \hline
Co-Mem~\cite{gao2018motion} & 32.0 & 31.7 \\
HMEMA~\cite{fan2019heterogeneous} & 33.0 & 33.7 \\
SSML~\cite{Amrani_Ben-Ari_Rotman_Bronstein_2021} & 35.0 & 35.1 \\
HCRN~\cite{le2020hierarchical} & 35.6 & 36.1 \\
DualVGR~\cite{wang2021dualvgr} & 35.5 & 39.0 \\
ClipBERT~\cite{lei2021less} & 37.4 & - \\
% JustAsk~\cite{Yang_2021_ICCV} & 43.1 & 46.3 \\
% VIOLET~\cite{fu2021violet} & 43.9 & 47.9 \\ \hline
\hline
\textbf{Ours} & \textbf{38.3} & \textbf{39.5} \\ \hline
\end{tabular}
|
212,278
|
2203.07720v3
| null |
\caption{MSRVTT Multiple Choice}
|
\label{tab:msrvtt_mc}
|
\begin{tabular}{l|c}
\hline
\textbf{Method} & \multicolumn{1}{l}{\textbf{MSRVTT}} \\ \hline
JSFusion~\cite{yu2018joint} & 83.4 \\
ActBERT~\cite{zhu2020actbert} & 85.7 \\
ClipBERT~\cite{lei2021less} & 88.2 \\
VideoCLIP~\cite{xu2021videoclip} & 92.1 \\
MERLOT~\cite{zellersluhessel2021merlot} & 90.9 \\
VIOLET~\cite{fu2021violet} & 91.9 \\ \hline
\textbf{Ours} & \textbf{92.4} \\ \hline
\end{tabular}
|
212,279
|
2203.07720v3
| null |
\caption{MSRVTT QA and MSVD QA}
|
\label{tab:msrvtt_msvd_qa}
|
\begin{tabular}{l|c|c}
\hline
\textbf{Method} & \multicolumn{1}{l|}{\textbf{MSRVTT}} & \multicolumn{1}{l}{\textbf{MSVD}} \\ \hline
Co-Mem~\cite{gao2018motion} & 32.0 & 31.7 \\
HMEMA~\cite{fan2019heterogeneous} & 33.0 & 33.7 \\
SSML~\cite{Amrani_Ben-Ari_Rotman_Bronstein_2021} & 35.0 & 35.1 \\
HCRN~\cite{le2020hierarchical} & 35.6 & 36.1 \\
DualVGR~\cite{wang2021dualvgr} & 35.5 & 39.0 \\
ClipBERT~\cite{lei2021less} & 37.4 & - \\
% JustAsk~\cite{Yang_2021_ICCV} & 43.1 & 46.3 \\
% VIOLET~\cite{fu2021violet} & 43.9 & 47.9 \\ \hline
\hline
\textbf{Ours} & \textbf{38.3} & \textbf{39.5} \\ \hline
\end{tabular}
|
212,280
|
2203.07720v3
| null |
\caption{Comparisons with State-of-the-art results on TGIF-FrameQA.}
|
\label{tab:tgif-frameqa}
|
\begin{tabular}{l|c}
\hline
\textbf{Method} & \multicolumn{1}{l}{\textbf{TGIF-FrameQA}} \\ \hline
ST-VQA~\cite{jang2017tgif} & 49.3 \\
Co-Mem~\cite{gao2018motion} & 51.5 \\
PSAC~\cite{li2019beyond} & 55.7 \\
HMEMA~\cite{fan2019heterogeneous} & 53.8 \\
HCRN~\cite{le2020hierarchical} & 55.9 \\
QueST~\cite{jiang2020divide} & 59.7 \\
ClipBERT~\cite{lei2021less} & 60.3 \\
\hline
\textbf{Ours} & \textbf{60.6} \\ \hline
\end{tabular}
|
212,284
|
2203.07720v3
| null |
\caption{Results of MSRVTT caption.}
|
\label{tab:msrvtt-cap}
|
\begin{tabular}{l|cccc}
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c}{MSRVTT} \\
& BLEU4 & METEOR & ROUGE-L & CIDEr \\ \hline
OA-BTG \cite{zhang2019object} & 41.4 & 28.2 & - & 46.9 \\
GRU-EVE \cite{aafaq2019spatio} & 38.3 & 28.4 & 60.7 & 48.1 \\
SAAT \cite{zheng2020syntax} & 39.9 & 27.7 & 61.2 & \textbf{51.0} \\
STG-KD \cite{pan2020spatio} & 40.5 & 28.3 & 60.9 & 47.1 \\ \hline
SupportSet \cite{patrick2021supportset} & 38.9 & 28.2 & 59.8 & 48.6 \\
VideoAsMT \cite{korbar2020video} & 41.7 & 28.5 & - & - \\
UniVL \cite{luo2020univl} & \textbf{41.8} & 28.9 & 60.8 & 50.0 \\ \hline
Ours & 40.8 & \textbf{29.0} & \textbf{61.4} & 50.2 \\ \hline
\end{tabular}
|
212,285
|
2303.14969v1
| null |
\caption{Quantitative comparison on Taskonomy dataset. Few-shot baselines are $10$-shot evaluated on each fold after being trained on the tasks from the other folds, where fully-supervised baselines are trained and evaluated on tasks from each fold (DPT) or all folds (InvPT).}
|
\label{tab:main_table}
|
\begin{tabular}{c|c|cc|cc|cc|cc|cc}
\toprule
\multirow{4}{*}{Supervision} &
\multirow{4}{*}{Model} &
\multicolumn{10}{c}{Tasks} \\
\cmidrule{3-12}
& &
\multicolumn{2}{c|}{Fold 1} & \multicolumn{2}{c|}{Fold 2} & \multicolumn{2}{c|}{Fold 3} &
\multicolumn{2}{c|}{Fold 4} & \multicolumn{2}{c}{Fold 5} \\
\cmidrule{3-12}
& &
SS & SN & ED & ZD & TE & OE & K2 & K3 & RS & PC \\
& &
mIoU ↑ & mErr ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ \\
\midrule
\multirow{2}{*}{Full} &
DPT &
\textbf{0.4449} & \textbf{6.4414} & \textbf{0.0534} & \textbf{0.0268} & \textbf{0.0188} &
\textbf{0.0689} & \textbf{0.0358} & \textbf{0.0357} & \textbf{0.0860} & \textbf{0.0347} \\
&
InvPT &
0.3900 & 12.9249 & 0.0589 & 0.0298 & 0.0517 &
0.0788 & 0.0456 & 0.0384 & 0.0949 & 0.0370 \\
\midrule
\multirow{4}{*}{\shortstack{10-Shot\\($< 0.004\%$)}} &
HSNet &
0.1069 & 24.9120 & 0.2375 & 0.0748 & 0.1746 &
0.1643 & 0.1056 & 0.0651 & 0.2627 & 0.0610 \\
&
VAT &
0.0353 & 25.8134 & 0.2718 & 0.0779 & 0.1719 &
0.1655 & 0.1450 & 0.0678 & 0.2709 & 0.0796 \\
&
DGPNet &
0.0261 & 29.1668 & 0.4579 & 0.2846 & 0.1881 &
0.2130 & 0.1104 & 0.1308 & 0.3680 & 0.3574 \\
&
\textbf{Ours} &
\textbf{0.4097} & \textbf{11.4391} & \textbf{0.0741} & \textbf{0.0316} & \textbf{0.0791} &
\textbf{0.0912} & \textbf{0.0639} & \textbf{0.0519} & \textbf{0.1089} & \textbf{0.0420} \\
\bottomrule
\end{tabular}
|
212,286
|
2303.14969v1
| null |
\caption{Ablation study on matching and task-specific adaptation.}
|
\label{tab:abaltion_table}
|
\begin{tabular}{c|cc|cc|cc|cc|cc}
\toprule
\multirow{4}{*}{Model} &
\multicolumn{10}{c}{Tasks} \\
\cmidrule{2-11}
&
\multicolumn{2}{c|}{Fold 1} & \multicolumn{2}{c|}{Fold 2} & \multicolumn{2}{c|}{Fold 3} &
\multicolumn{2}{c|}{Fold 4} & \multicolumn{2}{c}{Fold 5} \\
\cmidrule{2-11}
&
SS & SN & ED & ZD & TE & OE & K2 & K3 & RS & PC \\
&
mIoU ↑ & mErr ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ \\
\midrule
Ours w/o Matching &
0.2681 & 13.0704 & 0.1111 & 0.0404 & \textbf{0.0778} &
0.1061 & \textbf{0.0613} & 0.0537 & 0.1559 & 0.0445 \\
Ours w/o Adaptation &
0.0002 & 23.4212 & 0.1515 & 0.0641 & 0.1513 &
0.1152 & 0.1110 & 0.0625 & 0.2033 & 0.0632 \\
Ours &
\textbf{0.4097} & \textbf{11.4391} & \textbf{0.0741} & \textbf{0.0316} & 0.0791 &
\textbf{0.0912} & 0.0639 & \textbf{0.0519} & \textbf{0.1089} & \textbf{0.0420} \\
\bottomrule
\end{tabular}
|
212,287
|
2303.14969v1
| null |
\caption{Ablation study on the choice of support set. We disjointly sample four different support sets and report the $10$-shot performance on each set, with the mean and standard deviation.}
|
\label{tab:support_set_choice}
|
\begin{tabular}{c|cc|cc|cc|cc|cc}
\toprule
\multirow{4}{*}{Support set} &
\multicolumn{10}{c}{Tasks} \\
\cmidrule{2-11}
&
\multicolumn{2}{c|}{Fold 1} & \multicolumn{2}{c|}{Fold 2} & \multicolumn{2}{c|}{Fold 3} &
\multicolumn{2}{c|}{Fold 4} & \multicolumn{2}{c}{Fold 5} \\
\cmidrule{2-11}
&
SS & SN & ED & ZD & TE & OE & K2 & K3 & RS & PC \\
&
mIoU ↑ & mErr ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ & RMSE ↓ \\
\midrule
\# 1 &
0.4097 & 11.4391 & 0.0741 & 0.0316 & 0.0791 &
0.0912 & 0.0639 & 0.0519 & 0.1089 & 0.0420 \\
\# 2 &
0.4190 & 11.8860 & 0.0845 & 0.0338 & 0.0839 &
0.0926 & 0.0629 & 0.0497 & 0.1131 & 0.0437 \\
\# 3 &
0.3781 & 11.7418 & 0.0776 & 0.0343 & 0.0807 &
0.0944 & 0.0656 & 0.0494 & 0.1101 & 0.0425 \\
\# 4 &
0.4017 & 11.6203 & 0.0794 & 0.0362 & 0.0799 &
0.0908 & 0.0672 & 0.0502 & 0.1158 & 0.0424 \\
\midrule
Mean &
0.4021 & 11.6718 & 0.0789 & 0.0340 & 0.0809 &
0.0922 & 0.0649 & 0.0503 & 0.1120 & 0.0427 \\
Std. &
0.0152 & 0.1640 & 0.0038 & 0.0016 & 0.0018 &
0.0014 & 0.0016 & 0.0010 & 0.0027 & 0.0006 \\
\bottomrule
\end{tabular}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.