comparison writeup/nips2010_submission.tex @ 483:b9cdb464de5f

pointeur a la demo
author Yoshua Bengio <bengioy@iro.umontreal.ca>
date Mon, 31 May 2010 17:57:45 -0400
parents ce69aa9204d8
children 9a757d565e46
comparison
equal deleted inserted replaced
482:ce69aa9204d8 483:b9cdb464de5f
114 There are two main parts in the pipeline. The first one, 114 There are two main parts in the pipeline. The first one,
115 from slant to pinch below, performs transformations. The second 115 from slant to pinch below, performs transformations. The second
116 part, from blur to contrast, adds different kinds of noise. 116 part, from blur to contrast, adds different kinds of noise.
117 117
118 {\large\bf Transformations}\\ 118 {\large\bf Transformations}\\
119
119 {\bf Slant.} 120 {\bf Slant.}
120 We mimic slant by shifting each row of the image 121 We mimic slant by shifting each row of the image
121 proportionnaly to its height: $shift = round(slant \times height)$. 122 proportionnaly to its height: $shift = round(slant \times height)$.
122 The $slant$ coefficient can be negative or positive with equal probability 123 The $slant$ coefficient can be negative or positive with equal probability
123 and its value is randomly sampled according to the complexity level: 124 and its value is randomly sampled according to the complexity level:
178 The actual value is given by bilinear interpolation considering the pixels 179 The actual value is given by bilinear interpolation considering the pixels
179 around the (non-integer) source position thus found. 180 around the (non-integer) source position thus found.
180 Here $pinch \sim U[-complexity, 0.7 \times complexity]$.\\ 181 Here $pinch \sim U[-complexity, 0.7 \times complexity]$.\\
181 182
182 {\large\bf Injecting Noise}\\ 183 {\large\bf Injecting Noise}\\
184
183 {\bf Motion Blur.} 185 {\bf Motion Blur.}
184 This GIMP filter is a ``linear motion blur'' in GIMP 186 This GIMP filter is a ``linear motion blur'' in GIMP
185 terminology, with two parameters, $length$ and $angle$. The value of 187 terminology, with two parameters, $length$ and $angle$. The value of
186 a pixel in the final image is the approximately mean value of the $length$ first pixels 188 a pixel in the final image is the approximately mean value of the $length$ first pixels
187 found by moving in the $angle$ direction. 189 found by moving in the $angle$ direction.
550 \resizebox{.99\textwidth}{!}{\includegraphics{images/improvements_charts.pdf}}\\ 552 \resizebox{.99\textwidth}{!}{\includegraphics{images/improvements_charts.pdf}}\\
551 \caption{Charts corresponding to tables \ref{tab:perturbation-effect} (left) and \ref{tab:multi-task} (right).} 553 \caption{Charts corresponding to tables \ref{tab:perturbation-effect} (left) and \ref{tab:multi-task} (right).}
552 \label{fig:improvements-charts} 554 \label{fig:improvements-charts}
553 \end{figure} 555 \end{figure}
554 556
555 557 A Flash demo of the recognizer (where both the MLP and the SDA can be compared)
558 can be executed on-line at {\tt http://deep.host22.com}.
556 559
557 \section{Conclusions} 560 \section{Conclusions}
558 561
559 The conclusions are positive for all the questions asked in the introduction. 562 The conclusions are positive for all the questions asked in the introduction.
560 \begin{itemize} 563 \begin{itemize}