Mercurial > ift6266

comparison writeup/nips2010_submission.tex @ 534:4d6493d171f6

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
added all sizes
author Yoshua Bengio <bengioy@iro.umontreal.ca>
date Tue, 01 Jun 2010 22:12:13 -0400
parents 22d5cd82d5f0
children caf7769ca19c f0ee2212ea7c
comparison
equal deleted inserted replaced
533:22d5cd82d5f0 534:4d6493d171f6
392 widely used for training and testing character 392 widely used for training and testing character
393 recognition systems~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005}. 393 recognition systems~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005}.
394 The dataset is composed of 814255 digits and characters (upper and lower cases), with hand checked classifications, 394 The dataset is composed of 814255 digits and characters (upper and lower cases), with hand checked classifications,
395 extracted from handwritten sample forms of 3600 writers. The characters are labelled by one of the 62 classes 395 extracted from handwritten sample forms of 3600 writers. The characters are labelled by one of the 62 classes
396 corresponding to ``0''-``9'',``A''-``Z'' and ``a''-``z''. The dataset contains 8 parts (partitions) of varying complexity. 396 corresponding to ``0''-``9'',``A''-``Z'' and ``a''-``z''. The dataset contains 8 parts (partitions) of varying complexity.
397 The fourth partition (called $hsf_4$), experimentally recognized to be the most difficult one, is the one recommended 397 The fourth partition (called $hsf_4$, 82587 examples),
398 experimentally recognized to be the most difficult one, is the one recommended
398 by NIST as a testing set and is used in our work as well as some previous work~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005} 399 by NIST as a testing set and is used in our work as well as some previous work~\citep{Granger+al-2007,Cortes+al-2000,Oliveira+al-2002-short,Milgram+al-2005}
399 for that purpose. We randomly split the remainder into a training set and a validation set for 400 for that purpose. We randomly split the remainder (731668 examples) into a training set and a validation set for
400 model selection. The sizes of these data sets are: 651668 for training, 80000 for validation, 401 model selection.
401 and 82587 for testing.
402 The performances reported by previous work on that dataset mostly use only the digits. 402 The performances reported by previous work on that dataset mostly use only the digits.
403 Here we use all the classes both in the training and testing phase. This is especially 403 Here we use all the classes both in the training and testing phase. This is especially
404 useful to estimate the effect of a multi-task setting. 404 useful to estimate the effect of a multi-task setting.
405 Note that the distribution of the classes in the NIST training and test sets differs 405 Note that the distribution of the classes in the NIST training and test sets differs
406 substantially, with relatively many more digits in the test set, and more uniform distribution 406 substantially, with relatively many more digits in the test set, and more uniform distribution
443 All data sets contain 32$\times$32 grey-level images (values in $[0,1]$) associated with a label 443 All data sets contain 32$\times$32 grey-level images (values in $[0,1]$) associated with a label
444 from one of the 62 character classes. 444 from one of the 62 character classes.
445 %\begin{itemize} 445 %\begin{itemize}
446 446
447 %\item 447 %\item
448 {\bf NIST.} This is the raw NIST special database 19~\citep{Grother-1995}. 448 {\bf NIST.} This is the raw NIST special database 19~\citep{Grother-1995}. It has
449 \{651668 / 80000 / 82587\} \{training / validation / test} examples.
449 450
450 %\item 451 %\item
451 {\bf P07.} This dataset is obtained by taking raw characters from all four of the above sources 452 {\bf P07.} This dataset is obtained by taking raw characters from all four of the above sources
452 and sending them through the transformation pipeline described in section \ref{s:perturbations}. 453 and sending them through the transformation pipeline described in section \ref{s:perturbations}.
453 For each new example to generate, a data source is selected with probability $10\%$ from the fonts, 454 For each new example to generate, a data source is selected with probability $10\%$ from the fonts,
454 $25\%$ from the captchas, $25\%$ from the OCR data and $40\%$ from NIST. We apply all the transformations in the 455 $25\%$ from the captchas, $25\%$ from the OCR data and $40\%$ from NIST. We apply all the transformations in the
455 order given above, and for each of them we sample uniformly a \emph{complexity} in the range $[0,0.7]$. 456 order given above, and for each of them we sample uniformly a \emph{complexity} in the range $[0,0.7]$.
457 It has \{81920000 / 80000 / 20000\} \{training / validation / test} examples.
456 458
457 %\item 459 %\item
458 {\bf NISTP.} This one is equivalent to P07 (complexity parameter of $0.7$ with the same proportions of data sources) 460 {\bf NISTP.} This one is equivalent to P07 (complexity parameter of $0.7$ with the same proportions of data sources)
459 except that we only apply 461 except that we only apply
460 transformations from slant to pinch. Therefore, the character is 462 transformations from slant to pinch. Therefore, the character is
461 transformed but no additional noise is added to the image, giving images 463 transformed but no additional noise is added to the image, giving images
462 closer to the NIST dataset. 464 closer to the NIST dataset.
465 It has \{81920000 / 80000 / 20000\} \{training / validation / test} examples.
463 %\end{itemize} 466 %\end{itemize}
464 467
465 \vspace*{-1mm} 468 \vspace*{-1mm}
466 \subsection{Models and their Hyperparameters} 469 \subsection{Models and their Hyperparameters}
467 \vspace*{-1mm} 470 \vspace*{-1mm}