Mercurial > ift6266
comparison writeup/nips2010_submission.tex @ 488:6c9ff48e15cd
Moved the tables into a separate supplementary material file
| author | dumitru@dumitru.mtv.corp.google.com |
|---|---|
| date | Mon, 31 May 2010 19:07:35 -0700 |
| parents | 6beaf3328521 |
| children | ee9836baade3 |
comparison
equal
deleted
inserted
replaced
| 485:6beaf3328521 | 488:6c9ff48e15cd |
|---|---|
| 600 %\bibliographystyle{plainnat} | 600 %\bibliographystyle{plainnat} |
| 601 \bibliographystyle{unsrtnat} | 601 \bibliographystyle{unsrtnat} |
| 602 %\bibliographystyle{apalike} | 602 %\bibliographystyle{apalike} |
| 603 } | 603 } |
| 604 | 604 |
| 605 \newpage | |
| 606 | |
| 607 \centerline{APPENDIX FOR {\bf Deep Self-Taught Learning for Handwritten Character Recognition}} | |
| 608 | |
| 609 \vspace*{1cm} | |
| 610 | |
| 611 \begin{table}[h] | |
| 612 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits + | |
| 613 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training | |
| 614 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture | |
| 615 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07) | |
| 616 and using a validation set to select hyper-parameters and other training choices. | |
| 617 \{SDA,MLP\}0 are trained on NIST, | |
| 618 \{SDA,MLP\}1 are trained on NISTP, and \{SDA,MLP\}2 are trained on P07. | |
| 619 The human error rate on digits is a lower bound because it does not count digits that were | |
| 620 recognized as letters. For comparison, the results found in the literature | |
| 621 on NIST digits classification using the same test set are included.} | |
| 622 \label{tab:sda-vs-mlp-vs-humans} | |
| 623 \begin{center} | |
| 624 \begin{tabular}{|l|r|r|r|r|} \hline | |
| 625 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline | |
| 626 Humans& 18.2\% $\pm$.1\% & 39.4\%$\pm$.1\% & 46.9\%$\pm$.1\% & $1.4\%$ \\ \hline | |
| 627 SDA0 & 23.7\% $\pm$.14\% & 65.2\%$\pm$.34\% & 97.45\%$\pm$.06\% & 2.7\% $\pm$.14\%\\ \hline | |
| 628 SDA1 & 17.1\% $\pm$.13\% & 29.7\%$\pm$.3\% & 29.7\%$\pm$.3\% & 1.4\% $\pm$.1\%\\ \hline | |
| 629 SDA2 & 18.7\% $\pm$.13\% & 33.6\%$\pm$.3\% & 39.9\%$\pm$.17\% & 1.7\% $\pm$.1\%\\ \hline | |
| 630 MLP0 & 24.2\% $\pm$.15\% & 68.8\%$\pm$.33\% & 78.70\%$\pm$.14\% & 3.45\% $\pm$.15\% \\ \hline | |
| 631 MLP1 & 23.0\% $\pm$.15\% & 41.8\%$\pm$.35\% & 90.4\%$\pm$.1\% & 3.85\% $\pm$.16\% \\ \hline | |
| 632 MLP2 & 24.3\% $\pm$.15\% & 46.0\%$\pm$.35\% & 54.7\%$\pm$.17\% & 4.85\% $\pm$.18\% \\ \hline | |
| 633 \citep{Granger+al-2007} & & & & 4.95\% $\pm$.18\% \\ \hline | |
| 634 \citep{Cortes+al-2000} & & & & 3.71\% $\pm$.16\% \\ \hline | |
| 635 \citep{Oliveira+al-2002} & & & & 2.4\% $\pm$.13\% \\ \hline | |
| 636 \citep{Milgram+al-2005} & & & & 2.1\% $\pm$.12\% \\ \hline | |
| 637 \end{tabular} | |
| 638 \end{center} | |
| 639 \end{table} | |
| 640 | |
| 641 \begin{table}[h] | |
| 642 \caption{Relative change in error rates due to the use of perturbed training data, | |
| 643 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models. | |
| 644 A positive value indicates that training on the perturbed data helped for the | |
| 645 given test set (the first 3 columns on the 62-class tasks and the last one is | |
| 646 on the clean 10-class digits). Clearly, the deep learning models did benefit more | |
| 647 from perturbed training data, even when testing on clean data, whereas the MLP | |
| 648 trained on perturbed data performed worse on the clean digits and about the same | |
| 649 on the clean characters. } | |
| 650 \label{tab:perturbation-effect} | |
| 651 \begin{center} | |
| 652 \begin{tabular}{|l|r|r|r|r|} \hline | |
| 653 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline | |
| 654 SDA0/SDA1-1 & 38\% & 84\% & 228\% & 93\% \\ \hline | |
| 655 SDA0/SDA2-1 & 27\% & 94\% & 144\% & 59\% \\ \hline | |
| 656 MLP0/MLP1-1 & 5.2\% & 65\% & -13\% & -10\% \\ \hline | |
| 657 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline | |
| 658 \end{tabular} | |
| 659 \end{center} | |
| 660 \end{table} | |
| 661 | |
| 662 \begin{table}[h] | |
| 663 \caption{Test error rates and relative change in error rates due to the use of | |
| 664 a multi-task setting, i.e., training on each task in isolation vs training | |
| 665 for all three tasks together, for MLPs vs SDAs. The SDA benefits much | |
| 666 more from the multi-task setting. All experiments on only on the | |
| 667 unperturbed NIST data, using validation error for model selection. | |
| 668 Relative improvement is 1 - single-task error / multi-task error.} | |
| 669 \label{tab:multi-task} | |
| 670 \begin{center} | |
| 671 \begin{tabular}{|l|r|r|r|} \hline | |
| 672 & single-task & multi-task & relative \\ | |
| 673 & setting & setting & improvement \\ \hline | |
| 674 MLP-digits & 3.77\% & 3.99\% & 5.6\% \\ \hline | |
| 675 MLP-lower & 17.4\% & 16.8\% & -4.1\% \\ \hline | |
| 676 MLP-upper & 7.84\% & 7.54\% & -3.6\% \\ \hline | |
| 677 SDA-digits & 2.6\% & 3.56\% & 27\% \\ \hline | |
| 678 SDA-lower & 12.3\% & 14.4\% & 15\% \\ \hline | |
| 679 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline | |
| 680 \end{tabular} | |
| 681 \end{center} | |
| 682 \end{table} | |
| 683 | |
| 684 | 605 |
| 685 \end{document} | 606 \end{document} |