Mercurial > ift6266

comparison writeup/nips2010_submission_supplementary.tex @ 556:a7193b092b0a

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
cleaner le supplementary material
author Yoshua Bengio <bengioy@iro.umontreal.ca>
date Thu, 03 Jun 2010 08:14:08 -0400
parents 8b7e054d22bd
children
comparison
equal deleted inserted replaced
555:b6dfba0a110c 556:a7193b092b0a
14 14
15 \maketitle 15 \maketitle
16 16
17 \section*{Appendix I: Full results} 17 \section*{Appendix I: Full results}
18 18
19 These tables correspond to Figures 3 and 4 and contain the raw error rates for each model and dataset considered. 19 These tables correspond to Figures 2 and 3 and contain the raw error rates for each model and dataset considered.
20 They also contain additional data such as test errors on P07 and standard errors.
20 21
21 \begin{table}[h] 22 \begin{table}[ht]
22 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits + 23 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits +
23 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training 24 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training
24 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture 25 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture
25 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07) 26 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07)
26 and using a validation set to select hyper-parameters and other training choices. 27 and using a validation set to select hyper-parameters and other training choices.
46 \citep{Milgram+al-2005} & & & & 2.1\% $\pm$.12\% \\ \hline 47 \citep{Milgram+al-2005} & & & & 2.1\% $\pm$.12\% \\ \hline
47 \end{tabular} 48 \end{tabular}
48 \end{center} 49 \end{center}
49 \end{table} 50 \end{table}
50 51
51 \begin{table}[h] 52 \begin{table}[ht]
52 \caption{Relative change in error rates due to the use of perturbed training data, 53 \caption{Relative change in error rates due to the use of perturbed training data,
53 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models. 54 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models.
54 A positive value indicates that training on the perturbed data helped for the 55 A positive value indicates that training on the perturbed data helped for the
55 given test set (the first 3 columns on the 62-class tasks and the last one is 56 given test set (the first 3 columns on the 62-class tasks and the last one is
56 on the clean 10-class digits). Clearly, the deep learning models did benefit more 57 on the clean 10-class digits). Clearly, the deep learning models did benefit more
67 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline 68 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline
68 \end{tabular} 69 \end{tabular}
69 \end{center} 70 \end{center}
70 \end{table} 71 \end{table}
71 72
72 \begin{table}[h] 73 \begin{table}[ht]
73 \caption{Test error rates and relative change in error rates due to the use of 74 \caption{Test error rates and relative change in error rates due to the use of
74 a multi-task setting, i.e., training on each task in isolation vs training 75 a multi-task setting, i.e., training on each task in isolation vs training
75 for all three tasks together, for MLPs vs SDAs. The SDA benefits much 76 for all three tasks together, for MLPs vs SDAs. The SDA benefits much
76 more from the multi-task setting. All experiments on only on the 77 more from the multi-task setting. All experiments on only on the
77 unperturbed NIST data, using validation error for model selection. 78 unperturbed NIST data, using validation error for model selection.
89 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline 90 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline
90 \end{tabular} 91 \end{tabular}
91 \end{center} 92 \end{center}
92 \end{table} 93 \end{table}
93 94
94 {\small 95
96 \newpage
97
98 \vspace*{10mm}
99 %{\small
95 \bibliography{strings,strings-short,strings-shorter,ift6266_ml,aigaion-shorter,specials} 100 \bibliography{strings,strings-short,strings-shorter,ift6266_ml,aigaion-shorter,specials}
96 %\bibliographystyle{plainnat} 101 %\bibliographystyle{plainnat}
97 \bibliographystyle{unsrtnat} 102 \bibliographystyle{unsrtnat}
98 %\bibliographystyle{apalike} 103 %\bibliographystyle{apalike}
99 } 104 %}
100 105
101 \end{document} 106 \end{document}