Mercurial > ift6266
comparison writeup/techreport.tex @ 452:b0622f78cfec
Add a small paragraph mentionning the distribution differences and a figure illustrating the difference.
| author | Arnaud Bergeron <abergeron@gmail.com> |
|---|---|
| date | Mon, 10 May 2010 13:52:57 -0400 |
| parents | 18841eeb433f |
| children | c0f738f0cef0 |
comparison
equal
deleted
inserted
replaced
| 451:227ebc0be7ae | 452:b0622f78cfec |
|---|---|
| 1 \documentclass[12pt,letterpaper]{article} | 1 \documentclass[12pt,letterpaper]{article} |
| 2 \usepackage[utf8]{inputenc} | 2 \usepackage[utf8]{inputenc} |
| 3 \usepackage{graphicx} | 3 \usepackage{graphicx} |
| 4 \usepackage{times} | 4 \usepackage{times} |
| 5 \usepackage{mlapa} | 5 \usepackage{mlapa} |
| 6 \usepackage{subfigure} | |
| 6 | 7 |
| 7 \begin{document} | 8 \begin{document} |
| 8 \title{Generating and Exploiting Perturbed Training Data for Deep Architectures} | 9 \title{Generating and Exploiting Perturbed Training Data for Deep Architectures} |
| 9 \author{The IFT6266 Gang} | 10 \author{The IFT6266 Gang} |
| 10 \date{April 2010, Technical Report, Dept. IRO, U. Montreal} | 11 \date{April 2010, Technical Report, Dept. IRO, U. Montreal} |
| 299 \item {\bf NISTP} {\em ne pas utiliser PNIST mais NISTP, pour rester politically correct...} | 300 \item {\bf NISTP} {\em ne pas utiliser PNIST mais NISTP, pour rester politically correct...} |
| 300 NISTP is equivalent to P07 except that we only apply transformations from slant to pinch. Therefore, the character is transformed | 301 NISTP is equivalent to P07 except that we only apply transformations from slant to pinch. Therefore, the character is transformed |
| 301 but no additionnal noise is added to the image, this gives images closer to the NIST dataset. | 302 but no additionnal noise is added to the image, this gives images closer to the NIST dataset. |
| 302 \end{itemize} | 303 \end{itemize} |
| 303 | 304 |
| 305 We noticed that the distribution of the training sets and the test sets differ. | |
| 306 Since our validation sets are sampled from the training set, they have approximately the same distribution, but the test set has a completely different distribution as illustrated in figure \ref {setsdata}. | |
| 307 | |
| 308 \begin{figure} | |
| 309 \subfigure[NIST training]{\includegraphics[width=0.5\textwidth]{images/nisttrainstats}} | |
| 310 \subfigure[NIST validation]{\includegraphics[width=0.5\textwidth]{images/nistvalidstats}} | |
| 311 \subfigure[NIST test]{\includegraphics[width=0.5\textwidth]{images/nistteststats}} | |
| 312 \subfigure[NISTP validation]{\includegraphics[width=0.5\textwidth]{images/nistpvalidstats}} | |
| 313 \caption{Proportion of each class in some of the data sets} | |
| 314 \label{setsdata} | |
| 315 \end{figure} | |
| 316 | |
| 304 \subsection{Models and their Hyperparameters} | 317 \subsection{Models and their Hyperparameters} |
| 305 | 318 |
| 306 \subsubsection{Multi-Layer Perceptrons (MLP)} | 319 \subsubsection{Multi-Layer Perceptrons (MLP)} |
| 307 | 320 |
| 308 An MLP is a family of functions that are described by stacking layers of of a function similar to | 321 An MLP is a family of functions that are described by stacking layers of of a function similar to |