comparison writeup/nips2010_submission.tex @ 464:24f4a8b53fcc

nips2010_submission.tex
author Yoshua Bengio <bengioy@iro.umontreal.ca>
date Fri, 28 May 2010 17:21:21 -0600
parents
children 6205481bf33f
comparison
equal deleted inserted replaced
463:5fa1c653620c 464:24f4a8b53fcc
1 \documentclass{article} % For LaTeX2e
2 \usepackage{nips10submit_e,times}
3
4 \usepackage{amsthm,amsmath,amssymb,bbold,bbm}
5 \usepackage{algorithm,algorithmic}
6 \usepackage[utf8]{inputenc}
7 \usepackage{graphicx,subfigure}
8 \usepackage{mlapa}
9
10 \title{Generating and Exploiting Perturbed and Multi-Task Handwritten Training Data for Deep Architectures}
11 \author{The IFT6266 Gang}
12
13 \begin{document}
14
15 %\makeanontitle
16 \maketitle
17
18 \begin{abstract}
19 Recent theoretical and empirical work in statistical machine learning has
20 demonstrated the importance of learning algorithms for deep
21 architectures, i.e., function classes obtained by composing multiple
22 non-linear transformations. In the area of handwriting recognition,
23 deep learning algorithms
24 had been evaluated on rather small datasets with a few tens of thousands
25 of examples. Here we propose a powerful generator of variations
26 of examples for character images based on a pipeline of stochastic
27 transformations that include not only the usual affine transformations
28 but also the addition of slant, local elastic deformations, changes
29 in thickness, background images, color, contrast, occlusion, and
30 various types of pixel and spatially correlated noise.
31 We evaluate a deep learning algorithm (Stacked Denoising Autoencoders)
32 on the task of learning to classify digits and letters transformed
33 with this pipeline, using the hundreds of millions of generated examples
34 and testing on the full 62-class NIST test set.
35 We find that the SDA outperforms its
36 shallow counterpart, an ordinary Multi-Layer Perceptron,
37 and that it is better able to take advantage of the additional
38 generated data, as well as better able to take advantage of
39 the multi-task setting, i.e.,
40 training from more classes than those of interest in the end.
41 In fact, we find that the SDA reaches human performance as
42 estimated by the Amazon Mechanical Turk on the 62-class NIST test characters.
43 \end{abstract}
44
45 \section{Introduction}
46
47 Deep Learning has emerged as a promising new area of research in
48 statistical machine learning (see~\emcite{Bengio-2009} for a review).
49 Learning algorithms for deep architectures are centered on the learning
50 of useful representations of data, which are better suited to the task at hand.
51 This is in great part inspired by observations of the mammalian visual cortex,
52 which consists of a chain of processing elements, each of which is associated with a
53 different representation. In fact,
54 it was found recently that the features learnt in deep architectures resemble
55 those observed in the first two of these stages (in areas V1 and V2
56 of visual cortex)~\cite{HonglakL2008}.
57 Processing images typically involves transforming the raw pixel data into
58 new {\bf representations} that can be used for analysis or classification.
59 For example, a principal component analysis representation linearly projects
60 the input image into a lower-dimensional feature space.
61 Why learn a representation? Current practice in the computer vision
62 literature converts the raw pixels into a hand-crafted representation
63 (e.g.\ SIFT features~\cite{Lowe04}), but deep learning algorithms
64 tend to discover similar features in their first few
65 levels~\cite{HonglakL2008,ranzato-08,Koray-08,VincentPLarochelleH2008-very-small}.
66 Learning increases the
67 ease and practicality of developing representations that are at once
68 tailored to specific tasks, yet are able to borrow statistical strength
69 from other related tasks (e.g., modeling different kinds of objects). Finally, learning the
70 feature representation can lead to higher-level (more abstract, more
71 general) features that are more robust to unanticipated sources of
72 variance extant in real data.
73
74 Whereas a deep architecture can in principle be more powerful than a
75 shallow one in terms of representation, depth appears to render the
76 training problem more difficult in terms of optimization and local minima.
77 It is also only recently that successful algorithms were proposed to
78 overcome some of these difficulties. All are based on unsupervised
79 learning, often in an greedy layer-wise ``unsupervised pre-training''
80 stage~\cite{Bengio-2009}. One of these layer initialization techniques,
81 applied here, is the Denoising
82 Auto-Encoder~(DEA)~\cite{VincentPLarochelleH2008-very-small}, which
83 performed similarly or better than previously proposed Restricted Boltzmann
84 Machines in terms of unsupervised extraction of a hierarchy of features
85 useful for classification. The principle is that each layer starting from
86 the bottom is trained to encode their input (the output of the previous
87 layer) and try to reconstruct it from a corrupted version of it. After this
88 unsupervised initialization, the stack of denoising auto-encoders can be
89 converted into a deep supervised feedforward neural network and trained by
90 stochastic gradient descent.
91
92
93 \section{Perturbation and Transformation of Character Images}
94
95 This section describes the different transformations we used to generate data, in their order.
96 The code for these transformations (mostly python) is available at
97 {\tt http://anonymous.url.net}. All the modules in the pipeline share
98 a global control parameter ($0 \le complexity \le 1$) that allows one to modulate the
99 amount of deformation or noise introduced.
100
101 We can differentiate two important parts in the pipeline. The first one,
102 from slant to pinch, performs transformations of the character. The second
103 part, from blur to contrast, adds noise to the image.
104
105 \subsection{Slant}
106
107 In order to mimic a slant effect, we simply shift each row of the image
108 proportionnaly to its height: $shift = round(slant \times height)$. We
109 round the shift in order to have a discret displacement. We do not use a
110 filter to smooth the result in order to save computing time and also
111 because latter transformations have similar effects.
112
113 The $slant$ coefficient can be negative or positive with equal probability
114 and its value is randomly sampled according to the complexity level. In
115 our case we take uniformly a number in the range $[0,complexity]$, so the
116 maximum displacement for the lowest or highest pixel line is of
117 $round(complexity \times 32)$.
118
119
120 \subsection{Thickness}
121
122 To change the thickness of the characters we used morpholigical operators:
123 dilation and erosion~\cite{Haralick87,Serra82}.
124
125 The basic idea of such transform is, for each pixel, to multiply in the
126 element-wise manner its neighbourhood with a matrix called the structuring
127 element. Then for dilation we remplace the pixel value by the maximum of
128 the result, or the minimum for erosion. This will dilate or erode objects
129 in the image and the strength of the transform only depends on the
130 structuring element.
131
132 We used ten different structural elements with increasing dimensions (the
133 biggest is $5\times5$). for each image, we radomly sample the operator
134 type (dilation or erosion) with equal probability and one structural
135 element from a subset of the $n$ smallest structuring elements where $n$ is
136 $round(10 \times complexity)$ for dilation and $round(6 \times complexity)$
137 for erosion. A neutral element is always present in the set, if it is
138 chosen the transformation is not applied. Erosion allows only the six
139 smallest structural elements because when the character is too thin it may
140 erase it completly.
141
142 \subsection{Affine Transformations}
143
144 We generate an affine transform matrix according to the complexity level,
145 then we apply it directly to the image. The matrix is of size $2 \times
146 3$, so we can represent it by six parameters $(a,b,c,d,e,f)$. Formally,
147 for each pixel $(x,y)$ of the output image, we give the value of the pixel
148 nearest to : $(ax+by+c,dx+ey+f)$, in the input image. This allows to
149 produce scaling, translation, rotation and shearing variances.
150
151 The sampling of the parameters $(a,b,c,d,e,f)$ have been tuned by hand to
152 forbid important rotations (not to confuse classes) but to give good
153 variability of the transformation. For each image we sample uniformly the
154 parameters in the following ranges: $a$ and $d$ in $[1-3 \times
155 complexity,1+3 \times complexity]$, $b$ and $e$ in $[-3 \times complexity,3
156 \times complexity]$ and $c$ and $f$ in $[-4 \times complexity, 4 \times
157 complexity]$.
158
159
160 \subsection{Local Elastic Deformations}
161
162 This filter induces a "wiggly" effect in the image. The description here
163 will be brief, as the algorithm follows precisely what is described in
164 \cite{SimardSP03}.
165
166 The general idea is to generate two "displacements" fields, for horizontal
167 and vertical displacements of pixels. Each of these fields has the same
168 size as the original image.
169
170 When generating the transformed image, we'll loop over the x and y
171 positions in the fields and select, as a value, the value of the pixel in
172 the original image at the (relative) position given by the displacement
173 fields for this x and y. If the position we'd retrieve is outside the
174 borders of the image, we use a 0 value instead.
175
176 To generate a pixel in either field, first a value between -1 and 1 is
177 chosen from a uniform distribution. Then all the pixels, in both fields, is
178 multiplied by a constant $\alpha$ which controls the intensity of the
179 displacements (bigger $\alpha$ translates into larger wiggles).
180
181 As a final step, each field is convoluted with a Gaussian 2D kernel of
182 standard deviation $\sigma$. Visually, this results in a "blur"
183 filter. This has the effect of making values next to each other in the
184 displacement fields similar. In effect, this makes the wiggles more
185 coherent, less noisy.
186
187 As displacement fields were long to compute, 50 pairs of fields were
188 generated per complexity in increments of 0.1 (50 pairs for 0.1, 50 pairs
189 for 0.2, etc.), and afterwards, given a complexity, we selected randomly
190 among the 50 corresponding pairs.
191
192 $\sigma$ and $\alpha$ were linked to complexity through the formulas
193 $\alpha = \sqrt[3]{complexity} \times 10.0$ and $\sigma = 10 - 7 \times
194 \sqrt[3]{complexity}$.
195
196
197 \subsection{Pinch}
198
199 This is another GIMP filter we used. The filter is in fact named "Whirl and
200 pinch", but we don't use the "whirl" part (whirl is set to 0). As described
201 in GIMP, a pinch is "similar to projecting the image onto an elastic
202 surface and pressing or pulling on the center of the surface".
203
204 Mathematically, for a square input image, think of drawing a circle of
205 radius $r$ around a center point $C$. Any point (pixel) $P$ belonging to
206 that disk (region inside circle) will have its value recalculated by taking
207 the value of another "source" pixel in the original image. The position of
208 that source pixel is found on the line thats goes through $C$ and $P$, but
209 at some other distance $d_2$. Define $d_1$ to be the distance between $P$
210 and $C$. $d_2$ is given by $d_2 = sin(\frac{\pi{}d_1}{2r})^{-pinch} \times
211 d_1$, where $pinch$ is a parameter to the filter.
212
213 If the region considered is not square then, before computing $d_2$, the
214 smallest dimension (x or y) is stretched such that we may consider the
215 region as if it was square. Then, after $d_2$ has been computed and
216 corresponding components $d_2\_x$ and $d_2\_y$ have been found, the
217 component corresponding to the stretched dimension is compressed back by an
218 inverse ratio.
219
220 The actual value is given by bilinear interpolation considering the pixels
221 around the (non-integer) source position thus found.
222
223 The value for $pinch$ in our case was given by sampling from an uniform
224 distribution over the range $[-complexity, 0.7 \times complexity]$.
225
226 \subsection{Motion Blur}
227
228 This is a GIMP filter we applied, a "linear motion blur" in GIMP
229 terminology. The description will be brief as it is a well-known filter.
230
231 This algorithm has two input parameters, $length$ and $angle$. The value of
232 a pixel in the final image is the mean value of the $length$ first pixels
233 found by moving in the $angle$ direction. An approximation of this idea is
234 used, as we won't fall onto precise pixels by following that
235 direction. This is done using the Bresenham line algorithm.
236
237 The angle, in our case, is chosen from a uniform distribution over
238 $[0,360]$ degrees. The length, though, depends on the complexity; it's
239 sampled from a Gaussian distribution of mean 0 and standard deviation
240 $\sigma = 3 \times complexity$.
241
242 \subsection{Occlusion}
243
244 This filter selects random parts of other (hereafter "occlusive") letter
245 images and places them over the original letter (hereafter "occluded")
246 image. To be more precise, having selected a subregion of the occlusive
247 image and a desination position in the occluded image, to determine the
248 final value for a given overlapping pixel, it selects whichever pixel is
249 the lightest. As a reminder, the background value is 0, black, so the value
250 nearest to 1 is selected.
251
252 To select a subpart of the occlusive image, four numbers are generated. For
253 compability with the code, we'll call them "haut", "bas", "gauche" and
254 "droite" (respectively meaning top, bottom, left and right). Each of these
255 numbers is selected according to a Gaussian distribution of mean $8 \times
256 complexity$ and standard deviation $2$. This means the largest the
257 complexity is, the biggest the occlusion will be. The absolute value is
258 taken, as the numbers must be positive, and the maximum value is capped at
259 15.
260
261 These four sizes collectively define a window centered on the middle pixel
262 of the occlusive image. This is the part that will be extracted as the
263 occlusion.
264
265 The next step is to select a destination position in the occluded
266 image. Vertical and horizontal displacements $y\_arrivee$ and $x\_arrivee$
267 are selected according to Gaussian distributions of mean 0 and of standard
268 deviations of, respectively, 3 and 2. Then an horizontal placement mode,
269 $place$, is selected to be of three values meaning
270 left, middle or right.
271
272 If $place$ is "middle", the occlusion will be horizontally centered
273 around the horizontal middle of the occluded image, then shifted according
274 to $x\_arrivee$. If $place$ is "left", it will be placed on the left of
275 the occluded image, then displaced right according to $x\_arrivee$. The
276 contrary happens if $place$ is $right$.
277
278 In both the horizontal and vertical positionning, the maximum position in
279 either direction is such that the selected occlusion won't go beyond the
280 borders of the occluded image.
281
282 This filter has a probability of not being applied, at all, of 60\%.
283
284
285 \subsection{Pixel Permutation}
286
287 This filter permuts neighbouring pixels. It selects first
288 $\frac{complexity}{3}$ pixels randomly in the image. Each of them are then
289 sequentially exchanged to one other pixel in its $V4$ neighbourhood. Number
290 of exchanges to the left, right, top, bottom are equal or does not differ
291 from more than 1 if the number of selected pixels is not a multiple of 4.
292
293 It has has a probability of not being applied, at all, of 80\%.
294
295
296 \subsection{Gaussian Noise}
297
298 This filter simply adds, to each pixel of the image independently, a
299 Gaussian noise of mean $0$ and standard deviation $\frac{complexity}{10}$.
300
301 It has has a probability of not being applied, at all, of 70\%.
302
303
304 \subsection{Background Images}
305
306 Following~\cite{Larochelle-jmlr-2009}, this transformation adds a random
307 background behind the letter. The background is chosen by first selecting,
308 at random, an image from a set of images. Then we choose a 32x32 subregion
309 of that image as the background image (by sampling x and y positions
310 uniformly while making sure not to cross image borders).
311
312 To combine the original letter image and the background image, contrast
313 adjustments are made. We first get the maximal values (i.e. maximal
314 intensity) for both the original image and the background image, $maximage$
315 and $maxbg$. We also have a parameter, $contrast$, given by sampling from a
316 uniform distribution over $[complexity, 1]$.
317
318 Once we have all these numbers, we first adjust the values for the
319 background image. Each pixel value is multiplied by $\frac{max(maximage -
320 contrast, 0)}{maxbg}$. Therefore the higher the contrast, the darkest the
321 background will be.
322
323 The final image is found by taking the brightest (i.e. value nearest to 1)
324 pixel from either the background image or the corresponding pixel in the
325 original image.
326
327 \subsection{Salt and Pepper Noise}
328
329 This filter adds noise to the image by randomly selecting a certain number
330 of them and, for those selected pixels, assign a random value according to
331 a uniform distribution over the $[0,1]$ ranges. This last distribution does
332 not change according to complexity. Instead, the number of selected pixels
333 does: the proportion of changed pixels corresponds to $complexity / 5$,
334 which means, as a maximum, 20\% of the pixels will be randomized. On the
335 lowest extreme, no pixel is changed.
336
337 This filter also has a probability of not being applied, at all, of 75\%.
338
339 \subsection{Spatially Gaussian Noise}
340
341 The aim of this transformation is to filter, with a gaussian kernel,
342 different regions of the image. In order to save computing time we decided
343 to convolve the whole image only once with a symmetric gaussian kernel of
344 size and variance choosen uniformly in the ranges: $[12,12 + 20 \times
345 complexity]$ and $[2,2 + 6 \times complexity]$. The result is normalized
346 between $0$ and $1$. We also create a symmetric averaging window, of the
347 kernel size, with maximum value at the center. For each image we sample
348 uniformly from $3$ to $3 + 10 \times complexity$ pixels that will be
349 averaging centers between the original image and the filtered one. We
350 initialize to zero a mask matrix of the image size. For each selected pixel
351 we add to the mask the averaging window centered to it. The final image is
352 computed from the following element-wise operation: $\frac{image + filtered
353 image \times mask}{mask+1}$.
354
355 This filter has a probability of not being applied, at all, of 75\%.
356
357 \subsection{Scratches}
358
359 The scratches module places line-like white patches on the image. The
360 lines are in fact heavily transformed images of the digit "1" (one), chosen
361 at random among five thousands such start images of this digit.
362
363 Once the image is selected, the transformation begins by finding the first
364 $top$, $bottom$, $right$ and $left$ non-zero pixels in the image. It is
365 then cropped to the region thus delimited, then this cropped version is
366 expanded to $32\times32$ again. It is then rotated by a random angle having a
367 Gaussian distribution of mean 90 and standard deviation $100 \times
368 complexity$ (in degrees). The rotation is done with bicubic interpolation.
369
370 The rotated image is then resized to $50\times50$, with anti-aliasing. In
371 that image, we crop the image again by selecting a region delimited
372 horizontally to $left$ to $left+32$ and vertically by $top$ to $top+32$.
373
374 Once this is done, two passes of a greyscale morphological erosion filter
375 are applied. Put briefly, this erosion filter reduces the width of the line
376 by a certain $smoothing$ amount. For small complexities (< 0.5),
377 $smoothing$ is 6, so the line is very small. For complexities ranging from
378 0.25 to 0.5, $smoothing$ is 5. It is 4 for complexities 0.5 to 0.75, and 3
379 for higher complexities.
380
381 To compensate for border effects, the image is then cropped to 28x28 by
382 removing two pixels everywhere on the borders, then expanded to 32x32
383 again. The pixel values are then linearly expanded such that the minimum
384 value is 0 and the maximal one is 1. Then, 50\% of the time, the image is
385 vertically flipped.
386
387 This filter is only applied only 15\% of the time. When it is applied, 50\%
388 of the time, only one patch image is generated and applied. In 30\% of
389 cases, two patches are generated, and otherwise three patches are
390 generated. The patch is applied by taking the maximal value on any given
391 patch or the original image, for each of the 32x32 pixel locations.
392
393 \subsection{Color and Contrast Changes}
394
395 This filter changes the constrast and may invert the image polarity (white
396 on black to black on white). The contrast $C$ is defined here as the
397 difference between the maximum and the minimum pixel value of the image. A
398 contrast value is sampled uniformly between $1$ and $1-0.85 \times
399 complexity$ (this insure a minimum constrast of $0.15$). We then simply
400 normalize the image to the range $[\frac{1-C}{2},1-\frac{1-C}{2}]$. The
401 polarity is inverted with $0.5$ probability.
402
403
404 \begin{figure}[h]
405 \resizebox{.99\textwidth}{!}{\includegraphics{images/example_t.png}}\\
406 \caption{Illustration of the pipeline of stochastic
407 transformations applied to the image of a lower-case t
408 (the upper left image). Each image in the pipeline (going from
409 left to right, first top line, then bottom line) shows the result
410 of applying one of the modules in the pipeline. The last image
411 (bottom right) is used as training example.}
412 \label{fig:pipeline}
413 \end{figure}
414
415
416 \section{Experimental Setup}
417
418 \subsection{Training Datasets}
419
420 \subsubsection{Data Sources}
421
422 \begin{itemize}
423 \item {\bf NIST}
424 The NIST Special Database 19 (NIST19) is a very widely used dataset for training and testing OCR systems.
425 The dataset is composed with 8????? digits and characters (upper and lower cases), with hand checked classifications,
426 extracted from handwritten sample forms of 3600 writers. The characters are labelled by one of the 62 classes
427 corresponding to "0"-"9","A"-"Z" and "a"-"z". The dataset contains 8 series of different complexity.
428 The fourth series, $hsf_4$, experimentally recognized to be the most difficult one for classification task is recommended
429 by NIST as testing set and is used in our work for that purpose. It contains 82600 examples,
430 while the training and validation sets (which have the same distribution) contain XXXXX and
431 XXXXX examples respectively.
432 The performances reported by previous work on that dataset mostly use only the digits.
433 Here we use all the classes both in the training and testing phase. This is especially
434 useful to estimate the effect of a multi-task setting.
435 Note that the distribution of the classes in the NIST training and test sets differs
436 substantially, with relatively many more digits in the test set, and uniform distribution
437 of letters in the test set, not in the training set (more like the natural distribution
438 of letters in text).
439
440 \item {\bf Fonts} TODO!!!
441
442 \item {\bf Captchas}
443 The Captcha data source is an adaptation of the \emph{pycaptcha} library (a python based captcha generator library) for
444 generating characters of the same format as the NIST dataset. The core of this data source is composed with a random character
445 generator and various kinds of tranformations similar to those described in the previous sections.
446 In order to increase the variability of the data generated, different fonts are used for generating the characters.
447 Transformations (slant, distorsions, rotation, translation) are applied to each randomly generated character with a complexity
448 depending on the value of the complexity parameter provided by the user of the data source. Two levels of complexity are
449 allowed and can be controlled via an easy to use facade class.
450 \item {\bf OCR data}
451 \end{itemize}
452
453 \subsubsection{Data Sets}
454 \begin{itemize}
455 \item {\bf NIST} This is the raw NIST special database 19.
456 \item {\bf P07}
457 The dataset P07 is sampled with our transformation pipeline with a complexity parameter of $0.7$.
458 For each new exemple to generate, we choose one source with the following probability: $0.1$ for the fonts,
459 $0.25$ for the captchas, $0.25$ for OCR data and $0.4$ for NIST. We apply all the transformations in their order
460 and for each of them we sample uniformly a complexity in the range $[0,0.7]$.
461 \item {\bf NISTP} NISTP is equivalent to P07 (complexity parameter of $0.7$ with the same sources proportion)
462 except that we only apply
463 transformations from slant to pinch. Therefore, the character is
464 transformed but no additionnal noise is added to the image, giving images
465 closer to the NIST dataset.
466 \end{itemize}
467
468 \subsection{Models and their Hyperparameters}
469
470 \subsubsection{Multi-Layer Perceptrons (MLP)}
471
472 An MLP is a family of functions that are described by stacking layers of of a function similar to
473 $$g(x) = \tanh(b+Wx)$$
474 The input, $x$, is a $d$-dimension vector.
475 The output, $g(x)$, is a $m$-dimension vector.
476 The parameter $W$ is a $m\times d$ matrix and is called the weight matrix.
477 The parameter $b$ is a $m$-vector and is called the bias vector.
478 The non-linearity (here $\tanh$) is applied element-wise to the output vector.
479 Usually the input is referred to a input layer and similarly for the output.
480 You can of course chain several such functions to obtain a more complex one.
481 Here is a common example
482 $$f(x) = c + V\tanh(b+Wx)$$
483 In this case the intermediate layer corresponding to $\tanh(b+Wx)$ is called a hidden layer.
484 Here the output layer does not have the same non-linearity as the hidden layer.
485 This is a common case where some specialized non-linearity is applied to the output layer only depending on the task at hand.
486
487 If you put 3 or more hidden layers in such a network you obtain what is called a deep MLP.
488 The parameters to adapt are the weight matrix and the bias vector for each layer.
489
490 \subsubsection{Stacked Denoising Auto-Encoders (SDAE)}
491 \label{SdA}
492
493 Auto-encoders are essentially a way to initialize the weights of the network to enable better generalization.
494 This is essentially unsupervised training where the layer is made to reconstruct its input through and encoding and decoding phase.
495 Denoising auto-encoders are a variant where the input is corrupted with random noise but the target is the uncorrupted input.
496 The principle behind these initialization methods is that the network will learn the inherent relation between portions of the data and be able to represent them thus helping with whatever task we want to perform.
497
498 An auto-encoder unit is formed of two MLP layers with the bottom one called the encoding layer and the top one the decoding layer.
499 Usually the top and bottom weight matrices are the transpose of each other and are fixed this way.
500 The network is trained as such and, when sufficiently trained, the MLP layer is initialized with the parameters of the encoding layer.
501 The other parameters are discarded.
502
503 The stacked version is an adaptation to deep MLPs where you initialize each layer with a denoising auto-encoder starting from the bottom.
504 During the initialization, which is usually called pre-training, the bottom layer is treated as if it were an isolated auto-encoder.
505 The second and following layers receive the same treatment except that they take as input the encoded version of the data that has gone through the layers before it.
506 For additional details see \cite{vincent:icml08}.
507
508 \section{Experimental Results}
509
510 \subsection{SDA vs MLP vs Humans}
511
512 We compare here the best MLP (according to validation set error) that we found against
513 the best SDA (again according to validation set error), along with a precise estimate
514 of human performance obtained via Amazon's Mechanical Turk (AMT)
515 service\footnote{http://mturk.com}. AMT users are paid small amounts
516 of money to perform tasks for which human intelligence is required.
517 Mechanical Turk has been used extensively in natural language
518 processing \cite{SnowEtAl2008} and vision
519 \cite{SorokinAndForsyth2008,whitehill09}. AMT users where presented
520 with 10 character images and asked to type 10 corresponding ascii
521 characters. Hence they were forced to make a hard choice among the
522 62 character classes. Three users classified each image, allowing
523 to estimate inter-human variability (shown as +/- in parenthesis below).
524
525 \begin{table}
526 \caption{Overall comparison of error rates ($\pm$ std.err.) on 62 character classes (10 digits +
527 26 lower + 26 upper), except for last columns -- digits only, between deep architecture with pre-training
528 (SDA=Stacked Denoising Autoencoder) and ordinary shallow architecture
529 (MLP=Multi-Layer Perceptron). The models shown are all trained using perturbed data (NISTP or P07)
530 and using a validation set to select hyper-parameters and other training choices.
531 \{SDA,MLP\}0 are trained on NIST,
532 \{SDA,MLP\}1 are trained on NISTP, and \{SDA,MLP\}2 are trained on P07.
533 The human error rate on digits is a lower bound because it does not count digits that were
534 recognized as letters. For comparison, the results found in the literature
535 on NIST digits classification using the same test set are included.}
536 \label{tab:sda-vs-mlp-vs-humans}
537 \begin{center}
538 \begin{tabular}{|l|r|r|r|r|} \hline
539 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
540 Humans& 18.2\% $\pm$.1\% & 39.4\%$\pm$.1\% & 46.9\%$\pm$.1\% & $>1.1\%$ \\ \hline
541 SDA0 & 23.7\% $\pm$.14\% & 65.2\%$\pm$.34\% & 97.45\%$\pm$.06\% & 2.7\% $\pm$.14\%\\ \hline
542 SDA1 & 17.1\% $\pm$.13\% & 29.7\%$\pm$.3\% & 29.7\%$\pm$.3\% & 1.4\% $\pm$.1\%\\ \hline
543 SDA2 & 18.7\% $\pm$.13\% & 33.6\%$\pm$.3\% & 39.9\%$\pm$.17\% & 1.7\% $\pm$.1\%\\ \hline
544 MLP0 & 24.2\% $\pm$.15\% & 68.8\%$\pm$.33\% & 78.70\%$\pm$.14\% & 3.45\% $\pm$.15\% \\ \hline
545 MLP1 & 23.0\% $\pm$.15\% & 41.8\%$\pm$.35\% & 90.4\%$\pm$.1\% & 3.85\% $\pm$.16\% \\ \hline
546 MLP2 & 24.3\% $\pm$.15\% & 46.0\%$\pm$.35\% & 54.7\%$\pm$.17\% & 4.85\% $\pm$.18\% \\ \hline
547 [5] & & & & 4.95\% $\pm$.18\% \\ \hline
548 [2] & & & & 3.71\% $\pm$.16\% \\ \hline
549 [3] & & & & 2.4\% $\pm$.13\% \\ \hline
550 [4] & & & & 2.1\% $\pm$.12\% \\ \hline
551 \end{tabular}
552 \end{center}
553 \end{table}
554
555 \subsection{Perturbed Training Data More Helpful for SDAE}
556
557 \begin{table}
558 \caption{Relative change in error rates due to the use of perturbed training data,
559 either using NISTP, for the MLP1/SDA1 models, or using P07, for the MLP2/SDA2 models.
560 A positive value indicates that training on the perturbed data helped for the
561 given test set (the first 3 columns on the 62-class tasks and the last one is
562 on the clean 10-class digits). Clearly, the deep learning models did benefit more
563 from perturbed training data, even when testing on clean data, whereas the MLP
564 trained on perturbed data performed worse on the clean digits and about the same
565 on the clean characters. }
566 \label{tab:sda-vs-mlp-vs-humans}
567 \begin{center}
568 \begin{tabular}{|l|r|r|r|r|} \hline
569 & NIST test & NISTP test & P07 test & NIST test digits \\ \hline
570 SDA0/SDA1-1 & 38\% & 84\% & 228\% & 93\% \\ \hline
571 SDA0/SDA2-1 & 27\% & 94\% & 144\% & 59\% \\ \hline
572 MLP0/MLP1-1 & 5.2\% & 65\% & -13\% & -10\% \\ \hline
573 MLP0/MLP2-1 & -0.4\% & 49\% & 44\% & -29\% \\ \hline
574 \end{tabular}
575 \end{center}
576 \end{table}
577
578
579 \subsection{Multi-Task Learning Effects}
580
581 As previously seen, the SDA is better able to benefit from the
582 transformations applied to the data than the MLP. In this experiment we
583 define three tasks: recognizing digits (knowing that the input is a digit),
584 recognizing upper case characters (knowing that the input is one), and
585 recognizing lower case characters (knowing that the input is one). We
586 consider the digit classification task as the target task and we want to
587 evaluate whether training with the other tasks can help or hurt, and
588 whether the effect is different for MLPs versus SDAs. The goal is to find
589 out if deep learning can benefit more (or less) from multiple related tasks
590 (i.e. the multi-task setting) compared to a corresponding purely supervised
591 shallow learner.
592
593 We use a single hidden layer MLP with 1000 hidden units, and a SDA
594 with 3 hidden layers (1000 hidden units per layer), pre-trained and
595 fine-tuned on NIST.
596
597 Our results show that the MLP benefits marginally from the multi-task setting
598 in the case of digits (5\% relative improvement) but is actually hurt in the case
599 of characters (respectively 3\% and 4\% worse for lower and upper class characters).
600 On the other hand the SDA benefitted from the multi-task setting, with relative
601 error rate improvements of 27\%, 15\% and 13\% respectively for digits,
602 lower and upper case characters, as shown in Table~\ref{tab:multi-task}.
603
604 \begin{table}
605 \caption{Test error rates and relative change in error rates due to the use of
606 a multi-task setting, i.e., training on each task in isolation vs training
607 for all three tasks together, for MLPs vs SDAs. The SDA benefits much
608 more from the multi-task setting. All experiments on only on the
609 unperturbed NIST data, using validation error for model selection.
610 Relative improvement is 1 - single-task error / multi-task error.}
611 \label{tab:multi-task}
612 \begin{center}
613 \begin{tabular}{|l|r|r|r|} \hline
614 & single-task & multi-task & relative \\
615 & setting & setting & improvement \\ \hline
616 MLP-digits & 3.77\% & 3.99\% & 5.6\% \\ \hline
617 MLP-lower & 17.4\% & 16.8\% & -4.1\% \\ \hline
618 MLP-upper & 7.84\% & 7.54\% & -3.6\% \\ \hline
619 SDA-digits & 2.6\% & 3.56\% & 27\% \\ \hline
620 SDA-lower & 12.3\% & 14.4\% & 15\% \\ \hline
621 SDA-upper & 5.93\% & 6.78\% & 13\% \\ \hline
622 \end{tabular}
623 \end{center}
624 \end{table}
625
626 \section{Conclusions}
627
628 \bibliography{strings,ml,aigaion,specials}
629 \bibliographystyle{mlapa}
630 %\bibliographystyle{apalike}
631
632 \end{document}