id
stringlengths 9
10
| title
stringlengths 10
127
| summary
stringlengths 581
1.92k
| source
stringlengths 30
31
| authors
sequence | categories
sequence | comment
stringclasses 28
values | journal_ref
nullclasses 1
value | primary_category
stringclasses 6
values | published
stringlengths 8
8
| updated
stringlengths 8
8
| content
stringlengths 21.3k
323k
| references
sequence |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
1102.0183 | High-Performance Neural Networks for Visual Object Classification | We present a fast, fully parameterizable GPU implementation of Convolutional
Neural Network variants. Our feature extractors are neither carefully designed
nor pre-wired, but rather learned in a supervised way. Our deep hierarchical
architectures achieve the best published results on benchmarks for object
classification (NORB, CIFAR10) and handwritten digit recognition (MNIST), with
error rates of 2.53%, 19.51%, 0.35%, respectively. Deep nets trained by simple
back-propagation perform better than more shallow ones. Learning is
surprisingly rapid. NORB is completely trained within five epochs. Test error
rates on MNIST drop to 2.42%, 0.97% and 0.48% after 1, 3 and 17 epochs,
respectively. | http://arxiv.org/pdf/1102.0183 | [
"Dan C. Cireşan",
"Ueli Meier",
"Jonathan Masci",
"Luca M. Gambardella",
"Jürgen Schmidhuber"
] | [
"cs.AI",
"cs.NE"
] | 12 pages, 2 figures, 5 tables | null | cs.AI | 20110201 | 20110201 | High-Performance Neural Networks
for Visual Object Classi cation
Dan C. Cire san, Ueli Meier, Jonathan Masci,
Luca M. Gambardella and J urgen Schmidhuber
Technical Report No. IDSIA-01-11
January 2011
IDSIA / USI-SUPSI
Dalle Molle Institute for Arti cial Intelligence
Galleria 2, 6928 Manno, Switzerland
IDSIA is a joint institute of both University of Lugano (USI) and University of Applied Sciences of Southern Switzerland (SUPSI),
and was founded in 1988 by the Dalle Molle Foundation which promoted quality of life.
This work was partially supported by the Swiss Commission for Technology and Innovation (CTI), Project n. 9688.1 IFF:
Intelligent Fill in Form.arXiv:1102.0183v1 [cs.AI] 1 Feb 2011
Technical Report No. IDSIA-01-11 1
High-Performance Neural Networks
for Visual Object Classi cation
Dan C. Cire san, Ueli Meier, Jonathan Masci,
Luca M. Gambardella and J urgen Schmidhuber
January 2011
Abstract
We present a fast, fully parameterizable GPU implementation of Convolutional Neural
Network variants. Our feature extractors are neither carefully designed nor pre-wired, but
rather learned in a supervised way. Our deep hierarchical architectures achieve the best
published results on benchmarks for object classi cation (NORB, CIFAR10) and handwritten
digit recognition (MNIST), with error rates of 2.53%, 19.51%, 0.35%, respectively. Deep
nets trained by simple back-propagation perform better than more shallow ones. Learning
is surprisingly rapid. NORB is completely trained within ve epochs. Test error rates on
MNIST drop to 2.42%, 0.97% and 0.48% after 1, 3 and 17 epochs, respectively.
1 Introduction
The human visual system eciently recognizes and localizes objects within cluttered scenes. For
arti cial systems, however, this is still dicult, due to viewpoint-dependent object variability,
and the high in-class variability of many object types. Deep hierarchical neural models roughly
mimick the nature of mammalian visual cortex, and by community consensus are among the most
promising architectures for such tasks. The most successful hierarchical object recognition systems
all extract localized features from input images, convolving image patches with lters. Filter
responses are then repeatedly sub-sampled and re- ltered, resulting in a deep feed-forward network
architecture whose output feature vectors are eventually classi ed. One of the rst hierarchical
neural systems was the Neocognitron (Fukushima, 1980) which inspired many of the more recent
variants.
Unsupervised learning methods applied to patches of natural images tend to produce localized
lters that resemble o -center-on-surround lters, orientation-sensitive bar detectors, Gabor lters
(Schmidhuber et al. , 1996; Olshausen and Field, 1997; Hoyer and Hyv arinen, 2000). These ndings
in conjunction with experimental studies of the visual cortex justify the use of such lters in the
so-called standard model for object recognition (Riesenhuber and Poggio, 1999; Serre et al. , 2007;
Mutch and Lowe, 2008), whose lters are xed, in contrast to those of Convolutional Neural
Networks (CNNs) (LeCun et al. , 1998; Behnke, 2003; Simard et al. , 2003), whose weights ( lters)
are randomly initialized and changed in a supervised way using back-propagation (BP).
Despite the hardware progress of the past decades, computational speed is still a limiting
factor for CNN architectures characterized by many building blocks typically set by trial and
error. To systematically test the impact of various architectures on classi cation performance,
we present a fast CNN implementation on Graphics Processing Units (GPUs). Previous GPU
implementations of CNNs (Chellapilla et al. , 2006; Uetz and Behnke, 2009) were hard-coded to
satisfy GPU hardware constraints, whereas our implementation is
exible and fully online (i.e.,
Technical Report No. IDSIA-01-11 2
weight updates after each image). It allows for training large CNNs within days instead of months,
such that we can investigate the in
uence of various structural parameters by exploring large
parameter spaces (Pinto et al. , 2009) and performing error analysis on repeated experiments.
We evaluate various networks on the handwritten digit benchmark MNIST (LeCun et al. , 1998)
and two image classi cation benchmarks: NORB (LeCun et al. , 2004) and CIFAR10 (Krizhevsky,
2009).
2 Convolutional neural networks
CNNs are hierarchical neural networks whose convolutional layers alternate with subsampling
layers, reminiscent of simple and complex cells in the primary visual cortex (Wiesel and Hubel,
1959). CNNs vary in how convolutional and subsampling layers are realized and how the nets
are trained. The CNN architecture considered in this study di ers from the one of Simard et al.
(2003) in the sense that after each CNN-layer an optional max-pooling layer (Scherer et al. , 2010)
can be used. Here we give a complete description of this independent implementation (Fig. 1).
2.1 Image processing layer
The image processing layer is an optional pre-processing layer of prede ned lters that are kept
xed during training. Thus additional information besides the raw input image can be provided
to the network, such as edges and gradients. In particular, we nd that a contrast-extracting layer
(Fukushima, 2003) helps to improve the recognition rate for NORB.
2.2 Convolutional layer
A convolutional layer is parametrized by the size and the number of the maps, kernel sizes, skipping
factors, and the connection table. Each layer has Mmaps of equal size ( Mx,My). A kernel (blue
rectangle in Fig 1) of size ( Kx,Ky) is shifted over the valid region of the input image (i.e. the
kernel has to be completely inside the image). The skipping factors SxandSyde ne how many
pixels the lter/kernel skips in x- and y-direction between subsequent convolutions. The size of
the output map is then de ned as:
Mn
x=Mn 1
x Kn
x
Snx+ 1+ 1;Mn
y=Mn 1
y Kn
y
Sny+ 1+ 1 (1)
where index nindicates the layer. Each map in layer Lnis connected to at most Mn 1maps in
layerLn 1. Neurons of a given map share their weights but have di erent receptive elds.
2.3 Max-pooling layer
The biggest architectural di erence between our implementation and the CNN of LeCun et al.
(1998) is the use of a max-pooling layer instead of a sub-sampling layer. No such layer is used
by Simard et al. (2003) who simply skips nearby pixels prior to convolution, instead of pooling
or averaging. Scherer et al. (2010) found that max-pooling can lead to faster convergence, select
superior invariant features, and improve generalization. The output of the max-pooling layer is
given by the maximum activation over non-overlapping rectangular regions of size ( Kx,Ky). Max-
pooling enables position invariance over larger local regions and downsamples the input image by
a factor ofKxandKyalong each direction.
Technical Report No. IDSIA-01-11 3
2.4 Classi cation layer
Kernel sizes of convolutional lters and max-pooling rectangles as well as skipping factors are
chosen such that either the output maps of the last convolutional layer are downsampled to 1
pixel per map, or a fully connected layer combines the outputs of the topmost convolutional layer
into a 1D feature vector. The top layer is always fully connected, with one output unit per class
label.
Figure 1: Architecture of a convolutional neural network. In this case, the convolutional layers
are fully connected. Both convolutional layers use a kernel of 5 x 5 and skipping factors of 1.
3 GPU implementation
The latest generation of NVIDIA GPUs, the 400 and 500 series (we use GTX 480 & GTX 580), has
many advantages over older GPUs, most notably the presence of a R/W L2 global cache for device
memory. This permits faster programs and simpli es writing the code. In fact, the corresponding
transfer of complexity into hardware alleviates many software and optimization problems. Our
experiments show that the CNN program becomes 2-3 times faster just by switching from GTX
285 to GTX 480.
Technical Report No. IDSIA-01-11 4
Manual optimization of CUDA code is very time-consuming and error prone. We optimize for
the new architecture, relying on the L2 cache for many of the device memory accesses, instead of
manually writing code that uses textures and shared memory. Code obtained by this pragmatic
strategy is fast enough. We use the following types of optimization: pre-computed expressions,
unrolled loops within template kernels, strided matrices to obtain coalesced memory accesses and
registers wherever possible. Additional manual optimizations are possible in case future image
classi cation problems will require even more computing power.
3.1 Data structures
Both outputs yand deltasof layerLnare 2D strided. Their original size is MxMM y, but they
are horizontally strided with a pitch of 32
oats (we use this stride for all 2D data), resulting in
coalesced memory accesses. The vertical stride avoids additional bounding tests in CUDA kernels.
All connections between maps of consecutive layers Ln 1andLnare stored in matrix Cn.
Each row of Cncontains all connections that feed into a particular map in layer Ln. Because
we aim for a
exible architecture with partially connected layers, in the rst column we store the
number of previous connections. This index is useful for Forward Propagation (FP) and Adjusting
Weights (AW) CUDA kernels. The second column stores the number of connections, followed by
corresponding indices of maps in Ln 1connected to the current map.
For BP and FP, analogous information about connections is needed. We therefore store back-
ward connections in CBP. AW requires a list of all map connections (see Subsection 3.4), stored
as an array of map index pairs. Dealing with biases in BP kernel requires to know where the
weights of particular connections start; this information is stored in a 2D array WIDX BPof size
MnMn 1.
3.2 Forward propagation
A straightforward way of parallelizing FP is to assign a thread block to each map that has to
be computed. For maps bigger than 1024 neurons, the job is further split into smaller blocks by
assigning a block to each line of the map, because the number of threads per block is limited (1024
for GTX 480). A one to one correspondence between threads and the map's neurons is assumed.
Because of weight sharing, threads inside a block can access data in parallel, in particular the
same weights and inputs from the previous layer. Each thread starts by initializing its sum with
the bias, then loops over all map connections, convolving the appropriate patch of the input map
with the corresponding kernel. The output is obtained by passing the sum through a scaled tanh
activation function, and then written to device memory.
3.3 Backward propagation
BP of deltas can be done in two ways: by pushing or by pulling. Pushing deltas means taking each
delta from the current layer and computing the corresponding deltas for the previous layer. For
an architecture with shared weights this has the disadvantage of being hard to code. Each delta
from the current layer contributes to many deltas in the previous layer, which translates into a lot
of programming. There are two ways of avoiding this: either writing partial deltas to a separated
block of memory and then putting everything together by calling another kernel (slow because
of a tremendous increase in the number of memory accesses, and the need of another kernel), or
using atomic writes (to avoid data hazards) to update deltas (very slow because many writings
are serialized). We implement pulling deltas, which has almost none of the above speed-limiting
drawbacks, but is a bit more complicated.
Technical Report No. IDSIA-01-11 5
The (uni- or bi-dimensional) thread grid assigns a (bi- or uni-dimensional) thread block to each
map in the previous layer and a thread to each neuron in every map. Similar to FP, for maps
with more than 1024 neurons, the 2D grid is further split into smaller 1D blocks by assigning a
2D block to each row of the map. Each thread computes the delta of its corresponding neuron by
pulling deltas from the current layer. For every neuron in the previous layer we have to determine
the list of neurons in the current layer which are connected to it. Let us consider neuron ( i;j)
from a map in layer Ln 1, and then assume that ( x;y) are the coordinates of neurons in maps
ofLnthat contribute to the delta of neuron ( i;j). The (x;y) neuron is connected to kernel size
number neurons ( KxKy) from each connected map in the previous layer. The indices in Ln 1
of the neurons connected through a kernel to the ( x;y) neuron are:
x(Sx+ 1) ix(Sx+ 1) +Kx 1;
y(Sy+ 1) jy(Sy+ 1) +Ky 1:
We can now compute the inequalities for ( x;y):
i Kx+ 1
Sx+ 1xi
Sx+ 1;
j Ky+ 1
Sy+ 1yj
Sy+ 1:
Because (x;y) has to be inside the map, the nal inequalities are:
maxi Kx+ 1
Sx+ 1
;0
xmini
Sx+ 1
;Mx 1
;
maxj Ky+ 1
Sy+ 1
;0
yminj
Sy+ 1
;My 1
:
The above inequalities state that the delta of neuron ( i;j) fromLn 1is computed from deltas of
neurons in a rectangular area in maps of Ln(Fig. 2). After summing up the deltas, each thread
multiplies the result by the derivative of the activation function.
3.4 Adjusting weights
FP and BP have a grid on the list of maps, but the AW thread grid is on the list of kernels ( lters)
between maps of two consecutive layers. The 1D grid has a block for each connection between two
maps. Thread blocks are 2D, with a corresponding thread for every kernel weight. The bias weight
is included as an entire row of threads, thus requiring thread blocks to have ( Kx+1)Kythreads.
Most of the time these additional Kythreads will do nothing, thread (0,0) being activated only
for blocks that have to process the bias.
4 Experiments
We use a system with a Core i7-920 (2.66GHz), 12 GB DDR3 and four graphics cards: 2 x
GTX 480 and 2 x GTX 580. The correctness of the CPU version is checked by comparing the
analytical gradient with its nite di erence approximation. On GPU this is not possible because
all computations are performed with single precision
oating point numbers. Hence the GPU
implementation's correctness is checked by comparing its results to those of a randomly initialized
net after training it for several epochs on the more accurate CPU version. Obtaining identical
results after trillions of operations is a strong indication of correctness.
Technical Report No. IDSIA-01-11 6
Figure 2: Back propagating deltas. A connection between two maps from two consecutive layers
is displayed. The map in Ln 1has 29 x 29 neurons; the map in Lnhas 13 x 13 neurons. They
are linked through a 5 x 5 kernel K. Skipping factors of Sx= 1 andSy= 1 are assumed. Arrows
and colors depict the correspondence between neurons in Ln 1and their sources in Ln.
The implemented CNN's plain feed-forward architecture is trained using on-line gradient de-
scent. All images from the training set are used for training and also for validation. If deformations
are enabled, only the images from the training set will be deformed. Weights are initialized ac-
cording to a uniform random distribution in the range [ 0:05;0:05]. Each neuron's activation
function is a scaled hyperbolic tangent: y(a) = 1:7159 tanh(0 :6666a) (LeCun et al. , 1998).
We pick the trained CNN with the lowest validation error, and evaluate it on the test set
(Test for best Validation - TfbV). The best test error (bT) is also listed for all experiments. The
reported computation times per epoch include training, validation and testing as well as all data
transfers.
4.1 Experiments on MNIST
For the MNIST dataset the networks are trained on deformed images, continually generated in
on-line fashion. Ane (translation, rotation, scaling, horizontal shearing) and elastic deformations
(Simard et al. , 2003) are combined. We use a variable learning rate that shrinks by a multiplicative
constant after each epoch, from 10 3down to 3 10 5after 500 epochs.
Fully connected convolutional layers lead to an exploding number of network connections and
weights, making training of big and deep CNNs for hundreds of epochs impractical even on GPUs.
Partial connectivity alleviates this problem and is also biologically more plausible. We reduce
the number of connections between convolutional layers in a random way. Table 1 lists results
of various networks with 2 to 7 hidden layers with random connections. Additional layers result
in better networks, the best one achieving a test error of 0.35% for best validation and a best
test error of 0.27%. The best previous CNN result on MNIST is 0.40% (Simard et al. , 2003).
A 0.35% error rate was recently also obtained by a big, deep MLP (Cire san et al. , 2010) with
many more free parameters. Deeper nets require more computation time to complete an epoch,
but we observe that they also need fewer epochs to achieve good test errors. The deepest CNN
Technical Report No. IDSIA-01-11 7
Table 1: Error rates on MNIST test set for randomly connected CNNs with 2 to 6 convolutional
layers with M Maps and an optional fully connected layer with N neurons. Various kernel sizes
and skipping factors were used.
#M, #N bT TfbV
in Hidden Layers [%] [%]
20M-60M 0.95 1.02
20M-60M-150N 0.50 0.55
20M-60M-100M-150N 0.33 0.38
20M-40M-60M-80M-100M-120M-150N 0.27 0.35
from Table 1 reaches 2.42%, 0.97% and 0.48% after one, three and seventeen epochs, respectively.
On the other hand, the network with 4 instead of 7 hidden layers reaches 4.71%, 1.58%, 0.68%
after one, three and seventeen epochs, achieving a test error below 0.50% after only 34 epochs.
This shows once more that deep networks, contrary to common belief, can be trained successfully
by back-propagation. Despite the numerous free parameters, deep networks seem to learn faster
(better recognition rates after fewer epochs) than shallow ones.
We consider MNIST an almost solved problem. The remaining errors stem from digits that
are ambiguous or miss parts.
4.2 Experiments on NORB
NORB contains stereo images of 3D objects. Hence there are two maps on the input layer.
Rotation, scaling, shearing and elastic distortions seem to have a negative impact on generalization.
These deformations improve recognition rates for digits that are intrinsically 2D (Cire san et al. ,
2010), but seem inadequate for 3D objects.
Initial experiments on NORB show that unlike with MNIST where we use deformations, the
CNN needs only 3 to 6 epochs to reach zero validation error. This allows us to quickly run numerous
repetitive experiments with huge networks with hundreds of maps per layer. We decided to use a
CNN with ve hidden layers: layer1, a convolutional layer with 300 maps, kernel size 6 6 and
skipping factors 1 1; layer2, a max-pooling layer over a 2 2 region; layer3, a convolutional layer
with 500 maps, kernel size 4 4, skipping factors 0 0; layer4, a max-pooling layer over a 4 4
region; layer5, a fully connected layer with 500 neurons. The learning rate is initialized by 0.001
and multiplied by 0.95 after every epoch.
Table 2 summarizes the results of four di erent experiments by switching on/o translation as
well as the xed image processing layer. We report the average error rate as well as the standard
deviation of N independent runs with identical architectures but di erent weight initializations.
For the rst experiment without translation and no image processing (IP), an average test error
rate of 7.86% is obtained. With additional translations of at most 5%, the average error rate drops
to 4.71%, contradicting the common belief that CNNs are translation invariant. These results are
on par or better than others in the literature: 5.90% error rate for a combination of CNNs and
SVMs (LeCun et al. , 2004) and 5.20% error rate for restricted Boltzman machines (Nair and
Hinton, 2009).
The best previously published result on NORB (2.87%) was obtained by a hierarchical neural
network which to every convolutional layer provides a subsampled version plus edge information
of the original image (Uetz and Behnke, 2009). This motivated us to implement a pre-processing
layer with xed lters. We tried simple edge masks (Sobel, Scharr) but obtained best results with a
contrast-extraction layer (Fukushima, 2003) realized by Mexican hat-shaped lters of size 21 21,
one with a concentric on-center receptive eld and one with a concentric o -center receptive eld,
Technical Report No. IDSIA-01-11 8
Table 2: Average error rates and standard deviations of N runs for a ve hidden layer CNN on
the NORB test set (see text for details).
trans. [%] IP TfbV [%] runs time/epoch [s]
0 no 7.86 0.55 50 1141
5 no 4.71 0.57 50 1563
0 yes 3.94 0.48 50 1658
5 yes 2.530.40 100 2080
similar to the lters automatically created by unsupervised Predictability Minimization (Schmid-
huber, 1992) applied to natural images (Schmidhuber et al. , 1996). The rst lter extracts positive
contrast in brightness, whereas the latter extracts negative contrast. Each image from the original
NORB is ltered, consequently the input of the CNN has six maps: the original image plus the
positive and negative contrast for each of the two stereo channels. Using such a pre-processing
layer results in lower average error rates, 3.94% without translation and 2.53% with translation.
This result improves the previous state of the art on NORB (Uetz and Behnke, 2009).
Experience with other image datasets tells us that NORB is unusual. The training set has only
ve instances per class. The resulting poor training set variability makes the nets learn quickly
but generalize badly. NORB is the only dataset that pro ts from a xed pre-processing layer in
a substantial way. For MNIST and CIFAR10 such pre-processing has little or no e ect. It is also
worth noting that NORB's standard error rate deviation is bigger than CIFAR10's (see Tables 2
and 3). Identical nets with di erent initializations do not produce very consistent results. The
best net had an error rate of 1.72%, the worst 3.69%.
4.3 Experiments on CIFAR 10
CIFAR10 is a collection of natural color images of 32x32 pixels. It contains 10 classes, each of them
with 5000 samples in the training set and 1000 in the test set. The images greatly vary inside each
class. They are not necessarily centered, may contain only parts of the object, and have varying
backgrounds. All of this makes CIFAR10 the hardest problem addressed in this paper. The CNN
has three maps, one for each color channel (RGB). The CIFAR10 images are relatively small in
comparison to NORB's, and force us to use small kernels. The tested CNNs di er only in the
number of maps per convolutional and max-pooling layer. All have eight hidden layers: layer1, a
convolutional layer with 3 3 kernels and skipping factor of 0; layer2, a max-pooling layer over a
33 region; layer3, a convolutional layer with 3 3 kernels and skipping factors of 0 0; layer4,
a max-pooling over a 2 2 region; layer5, a convolutional layer with 3 3 kernels and a skipping
factors of 0 0; layer6, a max pooling layer over a 2 2 region; layer7, a fully connected layer
with 300 neurons; layer8, a fully connected layer with 100 neurons.
Like for MNIST, the learning rate is initialized by 0.001 and multiplied by 0.993 after every
epoch. Results in Table 3 show that without translation the error rate does not drop below 28%;
adding edge information does not help at all. Translations have a very positive e ect, decreasing
the error rate to almost 20%. Contrast extraction lters are better than the Sobel/Scharr lters but
still worse than no pre-processing layer at all. Despite some CNN-inherent translation invariance,
additional training image translations cause better generalization; additional image processing
proved useless though.
To see if bigger nets are better, we increase the number of maps per layer from 100 to 200, 300
and 400, respectively (last three rows in Tab. 3). Training time increases exponentially, but the
test error decreases, reaching a minimum for nets with 300 maps per layer. Our 19.51% error rate
is better than the previous state of the art for this dataset, 20.40% (Coates et al. , 2010) and 25.50%
Technical Report No. IDSIA-01-11 9
Table 3: Average error rates and standard deviations for N runs of an eight hidden layer CNN on
the CIFAR10 test set (see text for details). The rst ve nets have 100 maps per convolutional
and max-pooling layer, whereas the sixth, seventh and eighth have 200, 300 and 400 maps per
hidden layer, respectively. IP - image processing layer: edge - 3 3 Sobel and Scharr lters; hat -
1313 positive and negative contrast extraction lters.
trans. [%] maps IP TfbV [%] runs time/epoch [s]
0 100 no 28.87 0.37 11 93
0 100 edge 29.11 0.36 15 104
5 100 no 20.26 0.21 11 111
5 100 edge 21.87 0.57 5 120
5 100 hat 21.44 0.44 4 136
5 200 no 19.90 0.16 5 248
5 300 no 19.51 0.18 5 532
5 400 no 19.54 0.16 5 875
(Yu and Zhang, 2010). Unlike Coates et al. (2010), however, we use the original images without
any particular input normalization. Note that the error rate standard deviations are smaller than
those obtained on NORB, that is, di erent initializations yield consistent results.
4.4 Speedup factor of GPU code
The GPU code scales well with network size. For small nets the speedup is small (but still over 10)
since they t better inside the CPU cache, and GPU resources are underutilized. For huge nets
(ex: Table 2) the GPU implementation is more than 60 times faster than a compiler-optimized
CPU version. Given the
exibility of our GPU version, this is a signi cant speedup. One epoch
takes 35 GPU minutes but more than 35 CPU hours.
5 Conclusion
We presented high-performance GPU-based CNN variants trained by on-line gradient descent, with
sparse random connectivity, computationally more ecient and biologically more plausible than
fully connected CNNs. Principal advantages include state-of-the-art generalization capabilities,
great
exibility and speed. All structural CNN parameters such as input image size, number of
hidden layers, number of maps per layer, kernel sizes, skipping factors and connection tables are
adaptable to any particular application. We applied our networks to benchmark datasets for digit
recognition (MNIST), 3D object recognition (NORB), and natural images (CIFAR10). On MNIST
the best network achieved a recognition test error rate of 0.35%, on NORB 2.53% and on CIFAR10
19.51%. Our results are raising the bars for all three benchmarks. Currently the particular CNN
types discussed in this paper seem to be the best adaptive image recognizers, provided there is a
labeled dataset of sucient size. No unsupervised pretraining is required. Good results require
big and deep but sparsely connected CNNs, computationally prohibitive on CPUs, but feasible on
current GPUs, where our implementation is 10 to 60 times faster than a compiler-optimized CPU
version.
Technical Report No. IDSIA-01-11 10
Acknowledgment
This work was partially funded by the Swiss Commission for Technology and Innovation (CTI),
Project n. 9688.1 IFF: Intelligent Fill in Form.
References
S. Behnke. Hierarchical Neural Networks for Image Interpretation , volume 2766 of Lecture Notes
in Computer Science . Springer, 2003.
K. Chellapilla, S. Puri, and P. Simard. High performance convolutional neural networks for
document processing. In International Workshop on Frontiers in Handwriting Recognition ,
2006.
D. C. Cire san, U. Meier, L. M. Gambardella, and J. Schmidhuber. Deep big simple neural nets
for handwritten digit recogntion. Neural Computation , 22(12):3207{3220, 2010.
A. Coates, H. Lee, and A. Ng. An analysis of single-layer networks in unsupervised feature learning.
InAdvances in Neural Information Processing Systems , 2010.
K. Fukushima. Neocognitron: A self-organizing neural network for a mechanism of pattern recog-
nition una ected by shift in position. Biological Cybernetics , 36(4):193{202, 1980.
K. Fukushima. Neocognitron for handwritten digit recognition. Neurocomputing , 51:161{180,
2003.
P. O. Hoyer and A. Hyv arinen. Independent component analysis applied to feature extraction
from colour and stero images. Network: Computation in Neural Systems , 11(3):191{210, 2000.
A. Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, Computer
Science Department, University of Toronto, 2009.
Y. LeCun, L. Bottou, Y. Bengio, and P. Ha ner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE , 86(11):2278{2324, November 1998.
Y. LeCun, F.-J. Huang, and L. Bottou. Learning methods for generic object recognition with in-
variance to pose and lighting. In Proc. of Computer Vision and Pattern Recognition Conference ,
2004.
J. Mutch and D. G. Lowe. Object class recognition and localization using sparse features with
limited receptive elds. Int. J. Comput. Vision , 56(6):503{511, 2008.
V. Nair and G. E. Hinton. 3d object recognition with deep belief nets. In Advances in Neural
Information Processing Systems , 2009.
B. A. Olshausen and D. J. Field. Sparse coding with an overcomplete basis set: A strategy
employed by v1? Vision Research , 37(23):3311{3325, December 1997.
N. Pinto, D. Doukhan, J. J. DiCarlo, and D. D. Cox. A high-throughput screening approach
to discovering good forms of biologically inspired visual representation. PLoS computational
biology , 5(11):e1000579, November 2009.
M. Riesenhuber and T. Poggio. Hierarchical models of object recognition in cortex. Nat. Neurosci. ,
2(11):1019{1025, 1999.
Technical Report No. IDSIA-01-11 11
D. Scherer, A. M uller, and S. Behnke. Evaluation of pooling operations in convolutional architec-
tures for object recognition. In International Conference on Arti cial Neural Networks , 2010.
J. Schmidhuber, M. Eldracher, and B. Foltin. Semilinear predictability minimization produces
well-known feature detectors. Neural Computation , 8(4):773{786, 1996.
J. Schmidhuber. Learning factorial codes by predictability minimization. Neural Computation ,
4(6):863{879, 1992.
T. Serre, L. Wolf, and T. Poggio. Object recognition with features inspired by visual cortex. In
Proc. of Computer Vision and Pattern Recognition Conference , 2007.
P. Simard, D. Steinkraus, and J. Platt. Best practices for convolutional neural networks applied
to visual document analysis. In Seventh International Conference on Document Analysis and
Recognition , pages 958{963, 2003.
R. Uetz and S. Behnke. Large-scale object recognition with cuda-accelerated hierarchical neural
networks. In IEEE International Converence on Intelligent Computing and Intelligent Systems
(ICIS) , 2009.
D. H. Wiesel and T. N. Hubel. Receptive elds of single neurones in the cat's striate cortex. J.
Physiol. , 148:574{591, 1959.
K. Yu and T. Zhang. Improved local coordinate coding using local tangents. In Proceedings of the
International Conference on Machine Learning , 2010. | [] |
1301.4083 | Knowledge Matters: Importance of Prior Information for Optimization | We explore the effect of introducing prior information into the intermediate
level of neural networks for a learning task on which all the state-of-the-art
machine learning algorithms tested failed to learn. We motivate our work from
the hypothesis that humans learn such intermediate concepts from other
individuals via a form of supervision or guidance using a curriculum. The
experiments we have conducted provide positive evidence in favor of this
hypothesis. In our experiments, a two-tiered MLP architecture is trained on a
dataset with 64x64 binary inputs images, each image with three sprites. The
final task is to decide whether all the sprites are the same or one of them is
different. Sprites are pentomino tetris shapes and they are placed in an image
with different locations using scaling and rotation transformations. The first
part of the two-tiered MLP is pre-trained with intermediate-level targets being
the presence of sprites at each location, while the second part takes the
output of the first part as input and predicts the final task's target binary
event. The two-tiered MLP architecture, with a few tens of thousand examples,
was able to learn the task perfectly, whereas all other algorithms (include
unsupervised pre-training, but also traditional algorithms like SVMs, decision
trees and boosting) all perform no better than chance. We hypothesize that the
optimization difficulty involved when the intermediate pre-training is not
performed is due to the {\em composition} of two highly non-linear tasks. Our
findings are also consistent with hypotheses on cultural learning inspired by
the observations of optimization problems with deep learning, presumably
because of effective local minima. | http://arxiv.org/pdf/1301.4083 | [
"Çağlar Gülçehre",
"Yoshua Bengio"
] | [
"cs.LG",
"cs.CV",
"cs.NE",
"stat.ML"
] | 37 Pages, 5 figures, 5 tables JMLR Special Topics on Representation
Learning Submission | null | cs.LG | 20130117 | 20130713 | Knowledge Matters: Importance of Prior Information for
Optimization
C a glar G ul cehre gulcehrc@iro.umontreal.ca
D epartement d'informatique et de recherche op erationnelle
Universit e de Montr eal, Montr eal, QC, Canada
Yoshua Bengio bengioy@iro.umontreal.ca
D epartement d'informatique et de recherche op erationnelle
Universit e de Montr eal, Montr eal, QC, Canada
Editor: Not Assigned
Abstract
We explore the e ect of introducing prior information into the intermediate level of deep
supervised neural networks for a learning task on which all the black-box state-of-the-art ma-
chine learning algorithms tested have failed to learn. We motivate our work from the hypothesis
that there is an optimization obstacle involved in the nature of such tasks, and that humans
learn useful intermediate concepts from other individuals via a form of supervision or guidance
using a curriculum. The experiments we have conducted provide positive evidence in favor of
this hypothesis. In our experiments, a two-tiered MLP architecture is trained on a dataset for
which each image input contains three sprites, and the binary target class is 1 if all three have
the same shape. Black-box machine learning algorithms only got chance on this task. Standard
deep supervised neural networks also failed. However, using a particular structure and guiding
the learner by providing intermediate targets in the form of intermediate concepts (the pres-
ence of each object) allows to nail the task. Much better than chance but imperfect results are
also obtained by exploring architecture and optimization variants, pointing towards a dicult
optimization task. We hypothesize that the learning diculty is due to the composition of two
highly non-linear tasks. Our ndings are also consistent with hypotheses on cultural learning
inspired by the observations of e ective local minima (possibly due to ill-conditioning and the
training procedure not being able to escape what appears like a local minimum).
Keywords: Deep Learning, Neural Networks, Optimization, Evolution of Culture, Curricu-
lum Learning, Training with Hints
1. Introduction
There is a recent emerging interest in di erent elds of science for cultural learning (Henrich
and McElreath, 2003) and how groups of individuals exchanging information can learn in ways
superior to individual learning. This is also witnessed by the emergence of new research elds
such as "Social Neuroscience". Learning from other agents in an environment by the means
of cultural transmission of knowledge with a peer-to-peer communication is an ecient and
natural way of acquiring or propagating common knowledge. The most popular belief on how
the information is transmitted between individuals is that bits of information are transmitted
by small units, called memes, which share some characteristics of genes, such as self-replication,
mutation and response to selective pressures (Dawkins, 1976).
1arXiv:1301.4083v6 [cs.LG] 13 Jul 2013
This paper is based on the hypothesis (which is further elaborated in Bengio (2013a)) that
human culture and the evolution of ideas have been crucial to counter an optimization issue:
this diculty would otherwise make it dicult for human brains to capture high level knowl-
edge of the world without the help of other educated humans. In this paper machine learning
experiments are used to investigate some elements of this hypothesis by seeking answers for
the following questions: are there machine learning tasks which are intrinsically hard for a
lone learning agent but that may become very easy when intermediate concepts are provided
by another agent as additional intermediate learning cues, in the spirit of Curriculum Learn-
ing (Bengio et al., 2009b)? What makes such learning tasks more dicult? Can speci c initial
values of the neural network parameters yield success when random initialization yield com-
plete failure? Is it possible to verify that the problem being faced is an optimization problem or
with a regularization problem? These are the questions discussed (if not completely addressed)
here, which relate to the following broader question: how can humans (and potentially one day,
machines) learn complex concepts?
In this paper, results of di erent machine learning algorithms on an arti cial learning task
involving binary 64 64 images are presented. In that task, each image in the dataset contains
3 Pentomino tetris sprites (simple shapes). The task is to gure out if all the sprites in the
image are the same or if there are di erent sprite shapes in the image. Several state-of-the-art
machine learning algorithms have been tested and none of them could perform better than
a random predictor on the test set. Nevertheless by providing hints about the intermediate
concepts (the presence and location of particular sprite classes), the problem can easily be solved
where the same-architecture neural network without the intermediate concepts guidance fails.
Surprisingly, our attempts at solving this problem with unsupervised pre-training algorithms
failed solve this problem. However, with speci c variations in the network architecture or
training procedure, it is found that one can make a big dent in the problem. For showing the
impact of intermediate level guidance, we experimented with a two-tiered neural network, with
supervised pre-training of the rst part to recognize the category of sprites independently of
their orientation and scale, at di erent locations, while the second part learns from the output
of the rst part and predicts the binary task of interest.
The objective of this paper is not to propose a novel learning algorithm or architecture,
but rather to re ne our understanding of the learning diculties involved with composed tasks
(here a logical formula composed with the detection of object classes), in particular the training
diculties involved for deep neural networks. The results also bring empirical evidence in favor
of some of the hypotheses from Bengio (2013a), discussed below, as well as introducing a
particular form of curriculum learning (Bengio et al., 2009b).
Building dicult AI problems has a long history in computer science. Speci cally hard
AI problems have been studied to create CAPTCHA's that are easy to solve for humans, but
hard to solve for machines (Von Ahn et al., 2003). In this paper we are investigating a dicult
problem for the o -the-shelf black-box machine learning algorithms.1
1.1 Curriculum Learning and Cultural Evolution Against E ective Local Minima
What Bengio (2013a) calls an e ective local minimum is a point where iterative training
stalls, either because of an actual local minimum or because the optimization algorithm is
1. You can access the source code of some experiments presented in that paper and their hyperparameters from
here: https://github.com/caglar/kmatters
2
unable (in reasonable time) to nd a descent path (e.g., because of serious ill-conditioning). In
this paper, it is hypothesized that some more abstract learning tasks such as those obtained
by composing simpler tasks are more likely to yield e ective local minima for neural networks,
and are generally hard for general-purpose machine learning algorithms.
The idea that learning can be enhanced by guiding the learner through intermediate easier
tasks is old, starting with animal training by shaping (Skinner, 1958; Peterson, 2004; Krueger
and Dayan, 2009). Bengio et al. (2009b) introduce a computational hypothesis related to a
presumed issue with e ective local minima when directly learning the target task: the good
solutions correspond to hard-to- nd-by-chance e ective local minima, and intermediate tasks
prepare the learner's internal con guration (parameters) in a way similar to continuation meth-
ods in global optimization (which go through a sequence of intermediate optimization problems,
starting with a convex one where local minima are no issue, and gradually morphing into the
target task of interest).
In a related vein, Bengio (2013a) makes the following inferences based on experimental
observations of deep learning and neural network learning:
Point 1: Training deep architectures is easier when some hints are given about the function
that the intermediate levels should compute (Hinton et al., 2006; Weston et al., 2008;
Salakhutdinov and Hinton, 2009; Bengio, 2009). The experiments performed here expand
in particular on this point.
Point 2: It is much easier to train a neural network with supervision (where examples ar
provided to it of when a concept is present and when it is not present in a variety of
examples) than to expect unsupervised learning to discover the concept (which may also
happen but usually leads to poorer renditions of the concept). The poor results obtained
with unsupervised pre-training reinforce that hypothesis .
Point 3: Directly training all the layers of a deep network together not only makes it dicult to
exploit all the extra modeling power of a deeper architecture but in many cases it actually
yields worse results as the number of required layers is increased (Larochelle et al., 2009;
Erhan et al., 2010). The experiments performed here also reinforce that hypothesis.
Point 4: Erhan et al. (2010) observed that no two training trajectories ended up in the
same e ective local minimum, out of hundreds of runs, even when comparing solutions as
functions from input to output, rather than in parameter space (thus eliminating from the
picture the presence of symmetries and multiple local minima due to relabeling and other
reparametrizations). This suggests that the number of di erent e ective local minima
(even when considering them only in function space) must be huge.
Point 5: Unsupervised pre-training, which changes the initial conditions of the descent pro-
cedure, sometimes allows to reach substantially better e ective local minima (in terms
of generalization error!), and these better local minima do not appear to be reachable
by chance alone (Erhan et al., 2010). The experiments performed here provide another
piece of evidence in favor of the hypothesis that where random initialization can yield
rather poor results, speci cally targeted initialization can have a drastic impact, i.e., that
3
e ective local minima are not just numerous but that some small subset of them are much
better and hard to reach by chance.2
Based on the above points, Bengio (2013a) then proposed the following hypotheses regarding
learning of high-level abstractions.
Optimization Hypothesis: When it learns, a biological agent performs an approximate
optimization with respect to some implicit objective function.
Deep Abstractions Hypothesis: Higher level abstractions represented in brains re-
quire deeper computations (involving the composition of more non-linearities).
Local Descent Hypothesis: The brain of a biological agent relies on approximate local
descent and gradually improves itself while learning.
E ective Local Minima Hypothesis: The learning process of a single human learner
(not helped by others) is limited by e ective local minima.
Deeper Harder Hypothesis: E ective local minima are more likely to hamper learning
as the required depth of the architecture increases.
Abstractions Harder Hypothesis: High-level abstractions are unlikely to be discov-
ered by a single human learner by chance, because these abstractions are represented by
a deep subnetwork of the brain, which learns by local descent.
Guided Learning Hypothesis: A human brain can learn high level abstractions if
guided by the signals produced by other agents that act as hints or indirect supervision
for these high-level abstractions.
Memes Divide-and-Conquer Hypothesis: Linguistic exchange, individual learning
and the recombination of memes constitute an ecient evolutionary recombination oper-
ator in the meme-space. This helps human learners to collectively build better internal
representations of their environment, including fairly high-level abstractions.
This paper is focused on \ Point 1 " and testing the \ Guided Learning Hypothesis ", using
machine learning algorithms to provide experimental evidence. The experiments performed
also provide evidence in favor of the \ Deeper Harder Hypothesis " and associated \ Abstractions
Harder Hypothesis ". Machine Learning is still far beyond the current capabilities of humans,
and it is important to tackle the remaining obstacles to approach AI. For this purpose, the
question to be answered is why tasks that humans learn e ortlessly from very few examples,
while machine learning algorithms fail miserably?
2. Recent work showed that rather deep feedforward networks can be very successfully trained when large
quantities of labeled data are available (Ciresan et al., 2010; Glorot et al., 2011a; Krizhevsky et al., 2012).
Nonetheless, the experiments reported here suggest that it all depends on the task being considered, since
even with very large quantities of labeled examples, the deep networks trained here were unsuccessful.
4
2. Culture and Optimization Diculty
As hypothesized in the \ Local Descent Hypothesis ", human brains would rely on a local ap-
proximate descent, just like a Multi-Layer Perceptron trained by a gradient-based iterative op-
timization. The main argument in favor of this hypothesis relies on the biologically-grounded
assumption that although ring patterns in the brain change rapidly, synaptic strengths un-
derlying these neural activities change only gradually, making sure that behaviors are generally
consistent across time. If a learning algorithm is based on a form of local (e.g. gradient-based)
descent, it can be sensitive to e ective local minima (Bengio, 2013a).
When one trains a neural network, at some point in the training phase the evaluation of
error seems to saturate, even if new examples are introduced. In particular Erhan et al. (2010)
nd that early examples have a much larger weight in the nal solution. It looks like the learner
is stuck in or near a local minimum. But since it is dicult to verify if this is near a true local
minimum or simply an e ect of strong ill-conditioning, we call such a \stuck" con guration an
e ective local minimum , whose de nition depends not just on the optimization objective but
also on the limitations of the optimization algorithm.
Erhan et al. (2010) highlighted both the issue of e ective local minima and a regulariza-
tion e ect when initializing a deep network with unsupervised pre-training. Interestingly, as
the network gets deeper the diculty due to e ective local minima seems to be get more pro-
nounced. That might be because of the number of e ective local minima increases (more like
an actual local minima issue), or maybe because the good ones are harder to reach (more like
an ill-conditioning issue) and more work will be needed to clarify this question.
As a result of Point 4 we hypothesize that it is very dicult for an individual's brain to
discover some higher level abstractions by chance only. As mentioned in the \ Guided Learning
Hypothesis " humans get hints from other humans and learn high-level concepts with the guid-
ance of other humans3. Curriculum learning (Bengio et al., 2009a) and incremental learning
(Solomono , 1989), are examples of this. This is done by properly choosing the sequence of
examples seen by the learner, where simpler examples are introduced rst and more complex
examples shown when the learner is ready for them. One of the hypothesis on why curriculum
works states that curriculum learning acts as a continuation method that allows one to discover
a good minimum, by rst nding a good minimum of a smoother error function. Recent ex-
periments on human subjects also indicates that humans teach by using a curriculum strategy
(Khan et al., 2011).
Some parts of the human brain are known to have a hierarchical organization (i.e. visual
cortex) consistent with the deep architecture studied in machine learning papers. As we go from
the sensory level to higher levels of the visual cortex, we nd higher level areas corresponding
to more abstract concepts. This is consistent with the Deep Abstractions Hypothesis .
Training neural networks and machine learning algorithms by decomposing the learning
task into sub-tasks and exploiting prior information about the task is well-established and
in fact constitutes the main approach to solving industrial problems with machine learning.
The contribution of this paper is rather on rendering explicit the e ective local minima issue
and providing evidence on the type of problems for which this diculty arises. This prior
information and hints given to the learner can be viewed as inductive bias for a particular task,
an important ingredient to obtain a good generalization error (Mitchell, 1980). An interesting
3. But some high-level concepts may also be hardwired in the brain, as assumed in the universal grammar
hypothesis (Montague, 1970), or in nature vs nurture discussions in cognitive science.
5
earlier nding in that line of research was done with Explanation Based Neural Networks
(EBNN) in which a neural network transfers knowledge across multiple learning tasks. An
EBNN uses previously learned domain knowledge as an initialization or search bias (i.e. to
constrain the learner in the parameter space) (O'Sullivan, 1996; Mitchell and Thrun, 1993).
Another related work in machine learning is mainly focused on reinforcement learning al-
gorithms, based on incorporating prior knowledge in terms of logical rules to the learning
algorithm as a prior knowledge to speed up and bias learning (Kunapuli et al., 2010; Towell
and Shavlik, 1994).
As discussed in \ Memes Divide and Conquer Hypothesis \ societies can be viewed as a
distributed computational processing systems. In civilized societies knowledge is distributed
across di erent individuals, this yields a space eciency. Moreover computation, i.e. each
individual can specialize on a particular task/topic, is also divided across the individuals in the
society and hence this will yield a computational eciency. Considering the limitations of the
human brain, the whole processing can not be done just by a single agent in an ecient manner.
A recent study in paleoantropology states that there is a substantial decline in endocranial
volume of the brain in the last 30000 years Henneberg (1988). The volume of the brain shrunk
to 1241 ml from 1502 ml (Henneberg and Steyn, 1993). One of the hypothesis on the reduction
of the volume of skull claims that, decline in the volume of the brain might be related to the
functional changes in brain that arose as a result of cultural development and emergence of
societies given that this time period overlaps with the transition from hunter-gatherer lifestyle
to agricultural societies.
3. Experimental Setup
Some tasks, which seem reasonably easy for humans to learn4, are nonetheless appearing almost
impossible to learn for current generic state-of-art machine learning algorithms.
Here we study more closely such a task, which becomes learnable if one provides hints to
the learner about appropriate intermediate concepts. Interestingly, the task we used in our
experiments is not only hard for deep neural networks but also for non-parametric machine
learning algorithms such as SVM's, boosting and decision trees.
The result of the experiments for varying size of dataset with several o -the-shelf black box
machine learning algorithms and some popular deep learning algorithms are provided in Table
1. The detailed explanations about the algorithms and the hyperparameters used for those
algorithms are given in the Appendix Section 5.2. We also provide some explanations about
the methodologies conducted for the experiments at Section 3.2.
3.1 Pentomino Dataset
In order to test our hypothesis, an arti cial dataset for object recognition using 64 64 binary
images is designed5. If the task is two tiered (i.e., with guidance provided), the task in the
rst part is to recognize and locate each Pentomino object class6in the image. The second
4. keeping in mind that humans can exploit prior knowledge, either from previous learning or innate knowledge.
5. The source code for the script that generates the arti cial Pentomino datasets (Arcade-Universe) is available
at:https://github.com/caglar/Arcade-Universe . This implementation is based on Olivier Breuleux's
bugland dataset generator.
6. A human learner does not seem to need to be taught the shape categories of each Pentomino sprite in order
to solve the task. On the other hand, humans have lots of previously learned knowledge about the notion of
shape and how central it is in de ning categories.
6
(a) sprites, not all same type
(b) sprites, all of same type
Figure 1: Left (a): An example image from the dataset which has a di erent sprite type in it.
Right (b): An example image from the dataset that has only one type of Pentomino
object in it, but with di erent orientations and scales.
part/ nal binary classi cation task is to gure out if all the Pentominos in the image are of
the same shape class or not. If a neural network learned to detect the categories of each object
at each location in an image, the remaining task becomes an XOR-like operation between the
detected object categories. The types of Pentomino objects that is used for generating the
dataset are as follows:
Pentomino sprites N, P, F, Y, J, and Q, along with the Pentomino N2 sprite (mirror of
\Pentomino N" sprite), the Pentomino F2 sprite (mirror of \Pentomino F" sprite), and the
Pentomino Y2 sprite (mirror of \Pentomino Y" sprite).
Figure 2: Di erent classes of Pentomino shapes used in our dataset.
As shown in Figures 1(a) and 1(b), the synthesized images are fairly simple and do not
have any texture. Foreground pixels are \1" and background pixels are \0". Images of the
training and test sets are generated iid. For notational convenience, assume that the domain
of raw input images is X, the set of sprites is S, the set of intermediate object categories is Y
for each possible location in the image and the set of nal binary task outcomes is Z=f0;1g.
Two di erent types of rigid body transformation is performed: sprite rotation rot(X;
) where
=f
: (
= 90 )^[( 2N);(0 3)]gand scaling scale (X; ) where 2f1;2gis the
scaling factor. The data generating procedure is summarized below.
Sprite transformations: Before placing the sprites in an empty image, for each image x2X,
a value for z2Zis randomly sampled which is to have (or not) the same three sprite
shapes in the image. Conditioned on the constraint given by z, three sprites are randomly
7
selectedsijfromSwithout replacement. Using a uniform probability distribution over
all possible scales, a scale is chosen and accordingly each sprite image is scaled. Then
rotate each sprite is randomly rotated by a multiple of 90 degrees.
Sprite placement: Upon completion of sprite transformations, a 64 64 uniform grid is gener-
ated which is divided into 8 8 blocks, each block being of size 8 8 pixels, and randomly
select three di erent blocks from the 64=8 8 on the grid and place the transformed
objects into di erent blocks (so they cannot overlap, by construction).
Each sprite is centered in the block in which it is located. Thus there is no object translation
inside the blocks. The only translation invariance is due to the location of the block inside the
image.
A Pentomino sprite is guaranteed to not over
ow the block in which it is located, and there
are no collisions or overlaps between sprites, making the task simpler. The largest possible
Pentomino sprite can be t into an 8 4 mask.
3.2 Learning Algorithms Evaluated
Initially the models are cross-validated by using 5-fold cross-validation. With 40,000 examples,
this gives 32,000 examples for training and 8,000 examples for testing. For neural network
algorithms, stochastic gradient descent (SGD) is used for training. The following standard
learning algorithms were rst evaluated: decision trees, SVMs with Gaussian kernel, ordinary
fully-connected Multi-Layer Perceptrons, Random Forests, k-Nearest Neighbors, Convolutional
Neural Networks, and Stacked Denoising Auto-Encoders with supervised ne-tuning. More
details of the con gurations and hyper-parameters for each of them are given in Appendix
Section 5.2. The only better than chance results were obtained with variations of the Structured
Multi-Layer Perceptron described below.
3.2.1 Structured Multi-Layer Perceptron (SMLP)
The neural network architecture that is used to solve this task is called the SMLP (Structured
Multi-Layer Perceptron), a deep neural network with two parts as illustrated in Figure 5 and
7:
The lower part, P1NN ( Part 1 Neural Network , as it is called in the rest of the paper),
has shared weights and local connectivity, with one identical MLP instance of the P1NN for
each patch of the image, and typically an 11-element output vector per patch (unless otherwise
noted). The idea is that these 11 outputs per patch could represent the detection of the sprite
shape category (or the absence of sprite in the patch). The upper part, P2NN ( Part 2 Neural
Network ) is a fully connected one hidden layer MLP that takes the concatenation of the outputs
of all patch-wise P1NNs as input. Note that the rst layer of P1NN is similar to a convolutional
layer but where the stride equals the kernel size, so that windows do not overlap, i.e., P1NN can
be decomposed into separate networks sharing the same parameters but applied on di erent
patches of the input image, so that each network can actually be trained patch-wise in the case
where a target is provided for the P1NN outputs. The P1NN output for patch piwhich is
extracted from the image xis computed as follows:
f(pi) =g2(Vg1(Upi+b) +c) (1)
8
where pi2Rdis the input patch/receptive eld extracted from location iof a single image.
U2Rdhdis the weight matrix for the rst layer of P1NN and b2Rd
his the vector of biases
for the rst layer of P1NN. g1() is the activation function of the rst layer and g2() is the
activation function of the second layer. In many of the experiments, best results were obtained
withg1() a rectifying non-linearity (a.k.a. as RELU), which is max(0; X) (Jarrett et al.,
2009b; Nair and Hinton, 2010; Glorot et al., 2011a; Krizhevsky et al., 2012). V2Rdhdois the
second layer's weights matrix, such that and c2Rdoare the biases of the second layer of the
P1NN, with doexpected to be smaller than dh.
In this way, g1(Upi+b) is an overcomplete representation of the input patch that can
potentially represent all the possible Pentomino shapes for all factors of variations in the patch
(rotation, scaling and Pentomino shape type). On the other hand, when trained with hints,
f(pi) is expected to be the lower dimensional representation of a Pentomino shape category
invariant to scaling and rotation in the given patch.
In the experiments with SMLP trained with hints (targets at the output of P1NN), the
P1NN is expected to perform classi cation of each 8 8 non-overlapping patches of the original
6464 input image without having any prior knowledge of whether that speci c patch contains
a Pentomino shape or not. P1NN in SMLP without hints just outputs the local activations for
each patch, and gradients on f(pi) are backpropagated from the upper layers. In both cases
P1NN produces the input representation for the Part 2 Neural Net (P2NN). Thus the input
representation of P2NN is the concatenated output of P1NN across all the 64 patch locations:
ho= [f(p0);:::;f(pi);:::;f(pN))] whereNis the number of patches and the ho2Rdi;di=
doN.hois the concatenated output of the P1NN at each patch.
There is a standardization layer on top of the output of P1NN that centers the activations
and performs divisive normalization by dividing by the standard deviation over a minibatch
of the activations of that layer. We denote the standardization function z(). Standardization
makes use of the mean and standard deviation computed for each hidden unit such that each
hidden unit of howill have 0 activation and unit standard deviation on average over the
minibatch. Xis the set of pentomino images in the minibatch, where X2RdinNis a matrix
withNimages.h(i)
o(xj) is the vector of activations of the i-th hidden unit of hidden layer
ho(xj) for thej-th example, with xj2X.
h(i)
o=1
NX
xj2Xh(i)
o(xj) (2)
h(i)
o=sPN
j(h(i)
o(xj) h(i)
o)2
N+ (3)
z(h(i)
o(xj)) =h(i)
o(xj) h(i)
o
max(h(i)
o;)(4)
whereis a very small constant, that is used to prevent numerical under
ows in the standard
deviation. P1NN is trained on each 8 8 patches extracted from the image.
hois standardized for each training and test sample separately. Di erent values of were
used for SMLP-hints and SMLP-nohints.
The concatenated output of P1NN is fed as an input to the P2NN. P2NN is a feedforward
MLP with a sigmoid output layer using a single RELU hidden layer. The task of P2NN is to
perform a nonlinear logical operation on the representation provided at the output of P1NN.
9
3.2.2 Structured Multi Layer Perceptron Trained with Hints (SMLP-hints)
The SMLP-hints architecture exploits a hint about the presence and category of Pentomino
objects, specifying a semantics for the P1NN outputs. P1NN is trained with the intermediate
targetY, specifying the type of Pentomino sprite shape present (if any) at each of the 64 patches
(88 non-overlapping blocks) of the image. Because a possible answer at a given location can
be \none of the object types" i.e., an empty patch, yp(for patchp) can take one of the 11
possible values, 1 for rejection and the rest is for the Pentomino shape classes, illustrated in
Figure 2:
yp=(
0 if patch pis empty
s2Sif the patch pcontains a Pentomino sprite :
A similar task has been studied by Fleuret et al. (2011) (at SI appendix Problem 17), who
compared the performance of humans vs computers.
The SMLP-hints architecture takes advantage of dividing the task into two subtasks during
training with prior information about intermediate-level relevant factors. Because the sum of
the training losses decomposes into the loss on each patch, the P1NN can be pre-trained patch-
wise. Each patch-speci c component of the P1NN is a fully connected MLP with 8 8 inputs
and 11 outputs with a softmax output layer. SMLP-hints uses the the standardization given
in Equation 3 but with = 0.
The standardization is a crucial step for training the SMLP on the Pentomino dataset, and
yields much sparser outputs, as seen on Figures 3 and 4. If the standardization is not used,
even SMLP-hints could not solve the Pentomino task. In general, the standardization step
dampens the small activations and augments larger ones(reducing the noise). Centering the
activations of each feature detector in a neural network has been studied in (Raiko et al., 2012)
and (Vatanen et al., 2013). They proposed that transforming the outputs of each hidden neuron
in a multi-layer perceptron network to have zero output and zero slope on average makes rst
order optimization methods closer to the second order techniques.
By default, the SMLP uses recti er hidden units as activation function, we found a sig-
ni cant boost by using recti cation compared to hyperbolic tangent and sigmoid activation
functions. The P1NN has a highly overcomplete architecture with 1024 hidden units per patch,
and L1 and L2 weight decay regularization coecients on the weights (not the biases) are re-
spectively 1e-6 and 1e-5. The learning rate for the P1NN is 0.75. 1 training epoch was enough
for the P1NN to learn the features of Pentomino shapes perfectly on the 40000 training ex-
amples. The P2NN has 2048 hidden units. L1 and L2 penalty coecients for the P2NN are
1e-6, and the learning rate is 0.1. These were selected by trial and error based on validation
set error. Both P1NN (for each patch) and P2NN are fully-connected neural networks, even
though P1NN globally is a special kind of convolutional neural network.
Filters of the rst layer of SMLP are shown in Figure 6. These are the examples of the
lters obtained with the SLMP-hints trained with 40k examples, whose results are given in
Table 1. Those lters look very noisy but they work perfectly on the Pentomino task.
3.2.3 Deep and Structured Supervised MLP without Hints (SMLP-nohints)
SMLP-nohints uses the same connectivity pattern (and deep architecture) that is also used
in the SMLP-hints architecture, but without using the intermediate targets ( Y). It directly
predicts the nal outcome of the task ( Z), using the same number of hidden units, the same
10
Figure 3: Bar chart of concatenated softmax output activations hoof P1NN (1164=704 out-
puts) in SMLP-hints before standardization, for a selected example. There are very
large spikes at each location for one of the possible 11 outcome (1 of K representa-
tion).
11
Figure 4: Softmax output activations hoof P1NN at SMLP-hints before standardization. There
are positive spiked outputs at the locations where there is a Pentomino shape. Posi-
tive and negative spikes arise because most of the outputs are near an average value.
Activations are higher at the locations where there is a pentomino shape.
12
Structured MLP
Architecture with
Hints
Final Binary task
labels
Intermediate level
targets .
Second
Level
Neural
Network
First
Level
Neural
Network
Figure 5: Structured MLP architecture, used with hints (trained in two phases, rst P1NN,
bottom two layers, then P2NN, top two layers). In SMLP-hints, P1NN is trained
on each 8x8 patch extracted from the image and the softmax output probabilities of
all 64 patches are concatenated into a 64 11 vector that forms the input of P2NN.
OnlyUandVare learned in the P1NN and its output on each patch is fed into
P2NN. The rst level and the second level neural networks are trained separately,
not jointly.
Figure 6: Filters of Structured MLP architecture, trained with hints on 40k examples.
13
connectivity and the same activation function for the hidden units as SMLP-hints. 120 hy-
perparameter values have been evaluated by randomly selecting the number of hidden units
from [64;128;256;512;1024;1200;2048] and randomly sampling 20 learning rates uniformly in
the log-domain within the interval of [0 :008;0:8]. Two fully connected hidden layers with 1024
hidden units (same as P1NN) per patch is used and 2048 (same as P2NN) for the last hid-
den layer, with twenty training epochs. For this network the best results are obtained with a
learning rate of 0.05.7
Structured MLP
Architecture without
Hints
Final Binary task
labels
Second
Level
Neural
Network
First
Level
Neural
Network
Figure 7: Structured MLP architecture, used without hints (SMLP-nohints). It is the same
architecture as SMLP-hints (Figure 5) but with both parts (P1NN and P2NN) trained
jointly with respect to the nal binary classi cation task.
We chose to experiment with various SMLP-nohint architectures and optimization proce-
dures, trying unsuccessfully to achieve as good results with SMLP-nohint as with SMLP-hints.
Recti er Non-Linearity A recti er nonlinearity is used for the activations of MLP hidden
layers. We observed that using piecewise linear nonlinearity activation function such as the
recti er can make the optimization more tractable.
7. The source code of the structured MLP is available at the github repository: https://github.com/caglar/
structured_mlp
14
Figure 8: First layer lters learned by the Structured MLP architecture, trained without us-
ing hints on 447600 examples with online SGD and a sigmoid intermediate layer
activation.
Intermediate Layer The output of the P1NN is considered as an intermediate layer of the
SMLP. For the SMLP-hints, only softmax output activations have been tried at the intermediate
layer, and that suced to learn the task. Since things did not work nearly as well with
the SMLP-nohints, several di erent activation functions have been tried: softmax( ), tanh(),
sigmoid() and linear activation functions.
Standardization Layer Normalization at the last layer of the convolutional neural networks
has been used occasionaly to encourage the competition between the hidden units. (Jarrett
et al., 2009a) used a local contrast normalization layer in their architecture which performs
subtractive and divisive normalization. A local contrast normalization layer enforces a local
competition between adjacent features in the feature map and between features at the same
spatial location in di erent feature maps. Similarly (Krizhevsky et al., 2012) observed that
using a local response layer that enjoys the bene t of using local normalization scheme aids
generalization.
Standardization has been observed to be crucial for both SMLP trained with or with-
out hints. In both SMLP-hints and SMLP-nohints experiments, the neural network was not
able to generalize or even learn the training set without using standardization in the SMLP
intermediate layer, doing just chance performance. More speci cally, in the SMLP-nohints ar-
chitecture, standardization is part of the computational graph, hence the gradients are being
backpropagated through it. The mean and the standard deviation is computed for each hidden
unit separately at the intermediate layer as in Equation 4. But in order to prevent numerical
under
ows or over
ows during the backpropagation we have used = 1e 8 (Equation 3).
The bene t of having sparse activations may be speci cally important for the ill-conditioned
problems, for the following reasons. When a hidden unit is \o ", its gradient (the derivative of
the loss with respect to its output) is usually close to 0 as well, as seen here. That means that all
o -diagonal second derivatives involving that hidden unit (e.g. its input weights) are also near 0.
This is basically like removing some columns and rows from the Hessian matrix associated with
a particular example. It has been observed that the condition number of the Hessian matrix
(speci cally, its largest eigenvalue) increases as the size of the network increases (Dauphin and
Bengio, 2013), making training considerably slower and inecient (Dauphin and Bengio, 2013).
Hence one would expect that as sparsity of the gradients (obtained because of sparsity of the
activations) increases, training would become more ecient, as if we were training a smaller
sub-network for each example, with shared weights across examples, as in dropouts (Hinton
et al., 2012).
In Figure 9, the activation of each hidden unit in a bar chart is shown: the e ect of
standardization is signi cant, making the activations sparser.
15
(a) Before standardization.
(b) After standardization.
Figure 9: Activations of the intermediate-level hidden units of an SLMP-nohints for a particular
examples (x-axis: hidden unit number, y-axis: activation value). Left (a): before
standardization. Right (b): after standardization.
In Figure 10, one can see the activation histogram of the SMLP-nohints intermediate layer,
showing the distribution of activation values, before and after standardization. Again the
sparsifying e ect of standardization is very apparent.
(a) Before standardization.
(b) After standardization.
Figure 10: Distribution histogram of activation values of SMLP-nohints intermediate layer.
Left (a): before standardization. Right (b): after standardization.
In Figures 10 and 9, the intermediate level activations of SMLP-nohints are shown before
and after standardization. These are for the same SMLP-nohints architecture whose results
are presented on Table 1. For that same SMLP, the Adadelta (Zeiler, 2012) adaptive learning
16
rate scheme has been used, with 512 hidden units for the hidden layer of P1NN and recti er
activation function. For the output of the P1NN, 11 sigmoidal units have been used while
P2NN had 1200 hidden units with recti er activation function. The output nonlinearity of the
P2NN is a sigmoid and the training objective is the binary crossentropy.
Adaptive Learning Rates We have experimented with several di erent adaptive learning
rate algorithms. We tried rmsprop8, Adadelta (Zeiler, 2012), Adagrad (Duchi et al., 2010) and
a linearly (1/t) decaying learning rate (Bengio, 2013b). For the SMLP-nohints with sigmoid
activation function we have found Adadelta(Zeiler, 2012) converging faster to an e ective local
minima and usually yielding better generalization error compared to the others.
3.2.4 Deep and Structured MLP with Unsupervised Pre-Training
Several experiments have been conducted using an architecture similar to the SMLP-nohints,
but by using unsupervised pre-training of P1NN, with Denoising Auto-Encoder (DAE) and/or
Contractive Auto-Encoders (CAE). Supervised ne-tuning proceeds as in the deep and struc-
tured MLP without hints. Because an unsupervised learner may not focus the representation
just on the shapes, a larger number of intermediate-level units at the output of P1NN has been
explored: previous work on unsupervised pre-training generally found that larger hidden layers
were optimal when using unsupervised pre-training, because not all unsupervised features will
be relevant to the task at hand. Instead of limiting to 11 units per patch, we experimented
with networks with up to 20 hidden (i.e., code) units per patch in the second-layer patch-wise
auto-encoder.
In Appendix 5.1 we also provided the result of some experiments with binary-binary RBMs
trained on 88 patches from the 40k training dataset.
In unsupervised pretraining experiments in this paper, both contractive auto-encoder (CAE)
with sigmoid nonlinearity and binary cross entropy cost function and denoising auto-encoder (DAE)
have been used. In the second layer, experiments were performed with a DAE with recti er
hidden units utilizing L1 sparsity and weight decay on the weights of the auto-encoder. Greedy
layerwise unsupervised training procedure is used to train the deep auto-encoder architecture
(Bengio et al., 2007). In unsupervised pretraining experiments, tied weights have been used.
Di erent combinations of CAE and DAE for unsupervised pretraining have been tested, but
none of the con gurations tested managed to learn the Pentomino task, as shown in Table 1.
3.3 Experiments with 1 of K representation
To explore the e ect of changing the complexity of the input representation on the diculty
of the task, a set of experiments have been designed with symbolic representations of the
information in each patch. In all cases an empty patch is represented with a 0 vector. These
representation can be seen as an alternative input for a P2NN-like network, i.e., they were fed
as input to an MLP or another black-box classi er.
The following four experiments have been conducted, each one using one using a di erent
input representation for each patch:
8. This is learning rate scaling method that is discussed by G. Hinton in his Video Lecture 6.5 - rmsprop: Divide
the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine
Learning, 2012.
17
Algorithm 20k dataset 40k dataset 80k dataset
Training Test Training Test Training Test
Error Error Error Error Error Error
SVM RBF 26.2 50.2 28.2 50.2 30.2 49.6
K Nearest Neighbors 24.7 50.0 25.3 49.5 25.6 49.0
Decision Tree 5.8 48.6 6.3 49.4 6.9 49.9
Randomized Trees 3.2 49.8 3.4 50.5 3.5 49.1
MLP 26.5 49.3 33.2 49.9 27.2 50.1
Convnet/Lenet5 50.6 49.8 49.4 49.8 50.2 49.8
Maxout Convnet 14.5 49.5 0.0 50.1 0.0 44.6
2 layer sDA 49.4 50.3 50.2 50.3 49.7 50.3
Struct. Supervised MLP w/o hints 0.0 48.6 0.0 36.0 0.0 12.4
Struct. MLP+CAE Supervised Finetuning 50.5 49.7 49.8 49.7 50.3 49.7
Struct. MLP+CAE+DAE, Supervised Finetuning 49.1 49.7 49.4 49.7 50.1 49.7
Struct. MLP+DAE+DAE, Supervised Finetuning 49.5 50.3 49.7 49.8 50.3 49.7
Struct. MLP with Hints 0.21 30.7 0 3.1 0 0.01
Table 1: The error percentages with di erent learning algorithms on Pentomino dataset with
di erent number of training examples.
Experiment 1-Onehot representation without transformations: In this experiment sev-
eral trials have been done with a 10-input one-hot vector per patch. Each input corre-
sponds to an object category given in clear, i.e., the ideal input for P2NN if a supervised
P1NN perfectly did its job.
Experiment 2-Disentangled representations: In this experiment, we did trials with 16
binary inputs per patch, 10 one-hot bits for representing each object category, 4 for
rotations and 2 for scaling, i.e., the whole information about the input is given, but it is
perfectly disentangled. This would be the ideal input for P2NN if an unsupervised P1NN
perfectly did its job.
Experiment 3-Onehot representation with transformations: For each of the ten object
types there are 8 = 4 2 possible transformations. Two objects in two di erent patches
are the considered \the same" (for the nal task) if their category is the same regardless
of the transformations. The one-hot representation of a patch corresponds to the cross-
product between the 10 object shape classes and the 4 2 transformations, i.e., one out of
80=1042 possibilities represented in an 80-bit one-hot vector. This also contains all
the information about the input image patch, but spread out in a kind of non-parametric
and non-informative (not disentangled) way, like a perfect memory-based unsupervised
learner (like clustering) could produce. Nevertheless, the shape class would be easier to
read out from this representation than from the image representation (it would be an OR
over 8 of the bits).
Experiment 4-Onehot representation with 80 choices: This representation has the same
1 of 80 one-hot representation per patch but the target task is de ned di erently. Two ob-
jects in two di erent patches are considered the same i they have exactly the same 80-bit
onehot representation (i.e., are of the same object category with the same transformation
applied).
The rst experiment is a sanity check. It was conducted with single hidden-layered MLP's
with recti er and tanh nonlinearity, and the task was learned perfectly (0 error on both training
and test dataset) with very few training epochs.
18
0 100 200 300 400 500 600 700 800 9000.00.10.20.30.40.50.6
Training Error Rate
Test Error Rate(a) Training and Test Errors for Experiment 4
0 100 200 300 400 500 600 700 8000.00.10.20.30.40.50.6
Training Error Rate
Test Error Rate (b) Training and Test Errors for Experiment 3
Figure 11: Tanh MLP training curves. Left (a): The training and test errors of Experiment
3 over 800 training epochs with 100k training examples using Tanh MLP. Right
(b):The training and test errors of Experiment 4 over 700 training epochs with 100k
training examples using Tanh MLP.
The results of Experiment 2 are given in Table 2. To improve results, we experimented
with the Maxout non-linearity in a feedforward MLP (Goodfellow et al., 2013) with two hidden
layers. Unlike the typical Maxout network mentioned in the original paper, regularizers have
been deliberately avoided in order to focus on the optimization issue, i.e: no weight decay, norm
constraint on the weights, or dropout. Although learning from a disentangled representation is
more dicult than learning from perfect object detectors, it is feasible with some architectures
such as the Maxout network. Note that this representation is the kind of representation that
one could hope an unsupervised learning algorithm could discover, at best, as argued in Bengio
et al. (2012).
The only results obtained on the validation set for Experiment 3 and Experiment 4 are
shown respectively in Table 3 and Table 4. In these experiments a tanh MLP with two hidden
layers have been tested with the same hyperparameters. In experiment 3 the complexity of the
problem comes from the transformations (8=4 2) and the number of object types.
But in experiment 4, the only source of complexity of the task comes from the number of
di erent object types. These results are in between the complete failure and complete success
observed with other experiments, suggesting that the task could become solvable with better
training or more training examples. Figure 11 illustrates the progress of training a tanh MLP,
on both the training and test error, for Experiments 3 and 4. Clearly, something has been
learned, but the task is not nailed yet. On experiment 3 for both maxout and tanh the maxout
there was a long plateau where the training error and objective stays almost same. Maxout
did just chance on the experiment for about 120 iterations on the training and the test set.
But after 120th iteration the training and test error started decline and eventually it was able
to solve the task. Moreover as seen from the curves in Figure 11(a) and 11(b), the training
and test error curves are almost the same for both tasks. This implies that for onehot inputs,
whether you increase the number of possible transformations for each object or the number of
19
Learning Algorithm Training Error Test Error
SVM 0.0 35.6
Random Forests 1.29 40.475
Tanh MLP 0.0 0.0
Maxout MLP 0.0 0.0
Table 2: Performance of di erent learning algorithms on disentangled representation in Exper-
iment 2.
Learning Algorithm Training Error Test Error
SVM 11.212 32.37
Random Forests 24.839 48.915
Tanh MLP 0.0 22.475
Maxout MLP 0.0 0.0
Table 3: Performance of di erent learning algorithms using a dataset with onehot vector and
80 inputs as discussed for Experiment 3.
object categories, as soon as the number of possible con gurations is same, the complexity of
the problem is almost the same for the MLP.
3.4 Does the E ect Persist with Larger Training Set Sizes?
The results shown in this section indicate that the problem in the Pentomino task clearly
is not just a regularization problem, but rather basically hinges on an optimization problem.
Otherwise, we would expect test error to decrease as the number of training examples increases.
This is shown rst by studying the online case and then by studying the ordinary training
case with a xed size training set but considering increasing training set sizes. In the online
minibatch setting, parameter updates are performed as follows:
t+1=t t (5)
t=PN
irtL(xt;t)
N(6)
whereL(xt;t) is the loss incurred on example xtwith parameters t, wheret2Z+and
is the learning rate.
Ordinary batch algorithms converge linearly to the optimum , however the noisy gradient
estimates in the online SGD will cause parameter to
uctuate near the local optima. However,
online SGD directly optimizes the expected risk, because the examples are drawn iid from the
ground-truth distribution (Bottou, 2010). Thus:
L1=E[L(x;)] =Z
xL(x;)p(x)dx (7)
20
Learning Algorithm Training Error Test Error
SVM 4.346 40.545
Random Forests 23.456 47.345
Tanh MLP 0 25.8
Table 4: Performance of di erent algorithms using a dataset with onehot vector and 80 binary
inputs as discussed in Experiment 4.
whereL1is the generalization error. Therefore online SGD is trying to minimize the
expected risk with noisy updates. Those noisy updates have the e ect of regularizer:
t=PN
irtL(xt;t)
N=rtL(x;t) +t (8)
wherertL(x;t) is the true gradient and tis the zero-mean stochastic gradient \noise"
due to computing the gradient over a nite-size minibatch sample.
We would like to know if the problem with the Pentomino dataset is more a regularization
or an optimization problem. An SMLP-nohints model was trained by online SGD with the ran-
domly generated online Pentomino stream. The learning rate was adaptive, with the Adadelta
procedure (Zeiler, 2012) on minibatches of 100 examples. In the online SGD experiments, two
SMLP-nohints that is trained with and without standardization at the intermediate layer with
exactly the same hyperparameters are tested. The SMLP-nohints P1NN patch-wise submodel
has 2048 hidden units and the SMLP intermediate layer has 1152 = 64 18 hidden units. The
nonlinearity that is used for the intermediate layer is the sigmoid. P2NN has 2048 hidden units.
SMLP-nohints has been trained either with or without standardization on top of the output
units of the P1NN. The experiments illustrated in Figures 12 and 13 are with the same SMLP
without hints architecture for which results are given in Table 1. In those graphs only the results
for the training on the randomly generated 545400 Pentomino samples have been presented. As
shown in the plots SMLP-nohints was not able to generalize without standardization. Although
without standardization the training loss seems to decrease initially, it eventually gets stuck in
a plateau where training loss doesn't change much.
Training of SMLP-nohints online minibatch SGD is performed using standardization in the
intermediate layer and Adadelta learning rate adaptation, on 1046000 training examples from
the randomly generated Pentomino stream. At the end of the training, test error is down to
27.5%, which is much better than chance but from from the score obtained with SMLP-hints
of near 0 error.
In another SMLP-nohints experiment without standardization the model is trained with the
1580000 Pentomino examples using online minibatch SGD. P1NN has 2048 hidden units and 16
sigmoidal outputs per patch. for the P1NN hidden layer. P2NN has 1024 hidden units for the
hidden layer. Adadelta is used to adapt the learning rate. At the end of training this SMLP,
the test error remained stuck, at 50.1%.
21
0 100 200 300 400 500 600 700 800
Batch no0.200.250.300.350.400.450.500.55Test Error
SMLP with standardization
SMLP without standardizationFigure 12: Test errors of SMLP-nohints with and without standardization in the intermediate
layer. Sigmoid as an intermediate layer activation has been used. Each tick (batch
no) in the x-axis represents 400 examples.
22
0 100 200 300 400 500 600 700
Batch no01234567Training LossSMLP with standardization
SMLP without standardizationFigure 13: Training errors of SMLP-nohints with and without standardization in the interme-
diate layer. Sigmoid nonlinearity has been used as an intermediate layer activation
function. The x-axis is in units of blocks of 400 examples in the training set.
23
3.4.1 Experiments with Increased Training Set Size
Here we consider the e ect of training di erent learners with di erent numbers of training
examples. For the experimental results shown in Table 1, 3 training set sizes (20k, 40k and 80k
examples) had been used. Each dataset was generated with di erent random seeds (so they
do not overlap). Figure 14 also shows the error bars for an ordinary MLP with three hidden
layers, for a larger range of training set sizes, between 40k and 320k examples. The number of
training epochs is 8 (more did not help), and there are three hidden layers with 2048 feature
detectors. The learning rate we used in our experiments is 0.01. The activation function of the
MLP is a tanh nonlinearity, while the L1, L2 penalty coecients are both 1e-6.
Table 1 shows that, without guiding hints, none of the state-of-art learning algorithms could
perform noticeably better than a random predictor on the test set. This shows the importance
of intermediate hints introduced in the SMLP. The decision trees and SVMs can over t the
training set but they could not generalize on the test set. Note that the numbers reported in the
table are for hyper-parameters selected based on validation set error, hence lower training errors
are possible if avoiding all regularization and taking large enough models. On the training set,
the MLP with two large hidden layers (several thousands) could reach nearly 0% training error,
but still did not manage to achieve good test error.
In the experiment results shown in Figure 14, we evaluate the impact of adding more training
data for the fully-connected MLP. As mentioned before for these experiments we have used a
MLP with three hidden layers where each layer has 2048 hidden units. The tanh( ) activation
function is used with 0.05 learning rate and minibatches of size 200.
As can be seen from the gure, adding more training examples did not help either training
or test error (both are near 50%, with training error slightly lower and test error slightly
higher), reinforcing the hypothesis that the dicult encountered is one of optimization, not of
regularization.
Figure 14: Training and test error bar charts for a regular MLP with 3 hidden layers. There
is no signi cant improvement on the generalization error of the MLP as the new
training examples are introduced.
24
3.5 Experiments on E ect of Initializing with Hints
Initialization of the parameters in a neural network can have a big impact on the learning
and generalization (Glorot and Bengio, 2010). Previously Erhan et al. (2010) showed that
initializing the parameters of a neural network with unsupervised pretraining guides the learn-
ing towards basins of attraction of local minima that provides better generalization from the
training dataset. In this section we analyze the e ect of initializing the SMLP with hints and
then continuing without hints at the rest of the training. For experimental analysis of hints
based initialization, SMLP is trained for 1 training epoch using the hints and for 60 epochs it
is trained without hints on the 40k examples training set. We also compared the same archi-
tecture with the same hyperparameters, against to SMLP-nohints trained for 61 iterations on
the same dataset. After one iteration of hint-based training SMLP obtained 9% training error
and 39% test error. Following the hint based training, SMLP is trained without hints for 60
epochs, but at epoch 18, it already got 0% training and 0% test error. The hyperparameters
for this experiment and the experiment that the results shown for the SMLP-hints in Table 1
are the same. The test results for initialization with and without hints are shown on Figure 15.
This gure suggests that initializing with hints can give the same generalization performance
but training takes longer.
Figure 15: Plots showing the test error of SMLP with random initialization vs initializing with
hint based training.
3.5.1 Further Experiments on Optimization for Pentomino Dataset
With extensive hyperparameter optimization and using standardization in the intermediate
level of the SMLP with softmax nonlinearity, SMLP-nohints was able to get 5.3% training and
25
6.7% test error on the 80k Pentomino training dataset. We used the 2050 hidden units for the
hidden layer of P1NN and 11 softmax output per patch. For the P2NN, we used 1024 hidden
units with sigmoid and learning rate 0.1 without using any adaptive learning rate method. This
SMLP uses a recti er nonlinearity for hidden layers of both P1NN and P2NN. Considering that
architecture uses softmax as the intermediate activation function of SMLP-nohints. It is very
likely that P1NN is trying to learn the presence of speci c Pentomino shape in a given patch.
This architecture has a very large capacity in the P1NN, that probably provides it enough
capacity to learn the presence of Pentomino shapes at each patch e ortlessly.
An MLP with 2 hidden layers, each 1024 recti er units, was trained using LBFGS (the
implementation from the scipy.optimize library) on 40k training examples, with gradients com-
puted on batches of 10000 examples at each iteration. However, after convergence of training,
the MLP was still doing chance on the test dataset.
We also observed that using linear units for the intermediate layer yields better general-
ization error without standardization compared to using activation functions such as sigmoid,
tanh and RELU for the intermediate layer. SMLP-nohints was able to get 25% generaliza-
tion error with linear units without standardization whereas all the other activation functions
that has been tested failed to generalize with the same number of training iterations without
standardization and hints. This suggests that using non-linear intermediate-level activation
functions without standardization introduces an optimization diculty for the SMLP-nohints,
maybe because the intermediate level acts like a bottleneck in this architecture.
4. Conclusion and Discussion
In this paper we have shown an example of task which seems almost impossible to solve by
standard black-box machine learning algorithms, but can be almost perfectly solved when
one encourages a semantics for the intermediate-level representation that is guided by prior
knowledge. The task has the particularity that it is de ned by the composition of two non-
linear sub-tasks (object detection on one hand, and a non-linear logical operation similar to
XOR on the other hand).
What is interesting is that in the case of the neural network, we can compare two networks
with exactly the same architecture but a di erent pre-training, one of which uses the known
intermediate concepts to teach an intermediate representation to the network. With enough
capacity and training time they can over t but did not not capture the essence of the task, as
seen by test set performance.
We know that a structured deep network can learn the task, if it is initialized in the right
place, and do it from very few training examples. Furthermore we have shown that if one
pre-trains SMLP with hints for only one epoch, it can nail the task. But the exactly same
architecture which started training from random initialization, failed to generalize.
Consider the fact that even SMLP-nohints with standardization after being trained using
online SGD on 1046000 generated examples and still gets 27.5% test error. This is an indication
that the problem is not a regularization problem but possibly an inability to nd a good e ective
local minima of generalization error .
What we hypothesize is that for most initializations and architectures (in particular the
fully-connected ones), although it is possible to nd a good e ective local minimum of training
error when enough capacity is provided, it is dicult (without the proper initialization) to nd a
good local minimum of generalization error. On the other hand, when the network architecture
26
is constrained enough but still allows it to represent a good solution (such as the structured
MLP of our experiments), it seems that the optimization problem can still be dicult and even
training error remains stuck high if the standardization isn't used. Standardization obviously
makes the training objective of the SMLP easier to optimize and helps it to nd at least a
better e ective local minimum of training error . This nding suggests that by using speci c
architectural constraints and sometimes domain speci c knowledge about the problem, one can
alleviate the optimization diculty that generic neural network architectures face.
It could be that the combination of the network architecture and training procedure pro-
duces a training dynamics that tends to yield into these minima that are poor from the point of
view of generalization error, even when they manage to nail training error by providing enough
capacity. Of course, as the number of examples increases, we would expect this discrepancy to
decrease, but then the optimization problem could still make the task unfeasible in practice.
Note however that our preliminary experiments with increasing the training set size (8-fold) for
MLPs did not reveal signs of potential improvements in test error yet, as shown in Figure 14.
Even using online training on 545400 Pentomino examples, the SMLP-nohints architecture was
still doing far from perfect in terms of generalization error (Figure 12).
These ndings bring supporting evidence to the \Guided Learning Hypothesis" and \Deeper
Harder Hypothesis" from Bengio (2013a): higher level abstractions, which are expressed by
composing simpler concepts, are more dicult to learn (with the learner often getting in an
e ective local minimum ), but that diculty can be overcome if another agent provides hints
of the importance of learning other, intermediate-level abstractions which are relevant to the
task.
Many interesting questions remain open. Would a network without any guiding hint even-
tually nd the solution with a enough training time and/or with alternate parametrizations?
To what extent is ill-conditioning a core issue? The results with LBFGS were disappointing but
changes in the architectures (such as standardization of the intermediate level) seem to make
training much easier. Clearly, one can reach good solutions from an appropriate initialization,
pointing in the direction of an issue with local minima, but it may be that good solutions are
also reachable from other initializations, albeit going through a tortuous ill-conditioned path
in parameter space. Why did our attempts at learning the intermediate concepts in an unsu-
pervised way fail? Are these results speci c to the task we are testing or a limitation of the
unsupervised feature learning algorithm tested? Trying with many more unsupervised vari-
ants and exploring explanatory hypotheses for the observed failures could help us answer that.
Finally, and most ambitious, can we solve these kinds of problems if we allow a community
of learners to collaborate and collectively discover and combine partial solutions in order to
obtain solutions to more abstract tasks like the one presented here? Indeed, we would like to
discover learning algorithms that can solve such tasks without the use of prior knowledge as
speci c and strong as the one used in the SMLP here. These experiments could be inspired by
and inform us about potential mechanisms for collective learning through cultural evolutions
in human societies.
Acknowledgments
We would like to thank to the ICLR 2013 reviewers for their insightful comments, and NSERC,
CIFAR, Compute Canada and Canada Research Chairs for funding.
27
References
A. Ben-Hur and J. Weston. A user's guide to support vector machines. Methods in Molecular
Biology , 609:223{239, 2010.
Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep
networks. In NIPS'2006 , 2007.
Yoshua Bengio. Learning deep architectures for AI. Foundations and Trends in Machine
Learning , 2(1):1{127, 2009. Also published as a book. Now Publishers, 2009.
Yoshua Bengio. Evolving culture vs local minima. In Growing Adaptive Machines: Integrating
Development and Learning in Arti cial Neural Networks , number also as ArXiv 1203.2990v1,
pages T. Kowaliw, N. Bredeche & R. Doursat, eds. Springer-Verlag, March 2013a. URL
http://arxiv.org/abs/1203.2990 .
Yoshua Bengio. Practical recommendations for gradient-based training of deep architectures.
In K.-R. M uller, G. Montavon, and G. B. Orr, editors, Neural Networks: Tricks of the Trade .
Springer, 2013b.
Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning.
In L eon Bottou and Michael Littman, editors, Proceedings of the Twenty-sixth International
Conference on Machine Learning (ICML'09) . ACM, 2009a.
Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning.
InICML'09 , 2009b.
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Unsupervised feature learning and deep
learning: A review and new perspectives. Technical Report arXiv:1206.5538, U. Montreal,
2012. URL http://arxiv.org/abs/1206.5538 .
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Unsupervised feature learning and deep
learning: A review and new perspectives. IEEE Trans. Pattern Analysis and Machine Intel-
ligence (PAMI) , 2013.
James Bergstra, Olivier Breuleux, Fr ed eric Bastien, Pascal Lamblin, Razvan Pascanu, Guil-
laume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU
and GPU math expression compiler. In Proceedings of the Python for Scienti c Computing
Conference (SciPy) , 2010.
L eon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of
COMPSTAT'2010 , pages 177{186. Springer, 2010.
Leo Breiman. Random forests. Machine Learning , 45(1):5{32, 2001.
D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber. Deep big simple neural nets
for handwritten digit recognition. Neural Computation , 22:1{14, 2010.
Yann Dauphin and Yoshua Bengio. Big neural networks waste capacity. Technical Report
arXiv:1301.3583, Universite de Montreal, 2013.
Richard Dawkins. The Sel sh Gene . Oxford University Press, 1976.
28
J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research , 12:2121{2159, 2010.
Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent,
and Samy Bengio. Why does unsupervised pre-training help deep learning? Journal of
Machine Learning Research , 11:625{660, February 2010.
Fran cois Fleuret, Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, and Donald
Geman. Comparing machines and humans on a visual categorization test. Proceedings of the
National Academy of Sciences , 108(43):17621{17625, 2011.
X. Glorot, A. Bordes, and Y. Bengio. Deep sparse recti er neural networks. In AISTATS ,
2011a.
Xavier Glorot and Yoshua Bengio. Understanding the diculty of training deep feedforward
neural networks. In JMLR W&CP: Proceedings of the Thirteenth International Conference
on Arti cial Intelligence and Statistics (AISTATS 2010) , volume 9, pages 249{256, May
2010.
Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse recti er neural networks. In
JMLR W&CP: Proceedings of the Fourteenth International Conference on Arti cial Intelli-
gence and Statistics (AISTATS 2011) , April 2011b.
Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio.
Maxout networks. In ICML , 2013.
Maciej Henneberg. Decrease of human skull size in the holocene. Human biology , pages 395{405,
1988.
Maciej Henneberg and Maryna Steyn. Trends in cranial capacity and cranial index in subsa-
haran africa during the holocene. American journal of human biology , 5(4):473{479, 1993.
J. Henrich and R. McElreath. The evolution of cultural evolution. Evolutionary Anthropology:
Issues, News, and Reviews , 12(3):123{135, 2003.
Geo rey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep
belief nets. Neural Computation , 18:1527{1554, 2006.
Geo rey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi-
nov. Improving neural networks by preventing co-adaptation of feature detectors. Technical
report, arXiv:1207.0580, 2012.
C.W. Hsu, C.C. Chang, C.J. Lin, et al. A practical guide to support vector classi cation, 2003.
Kevin Jarrett, Koray Kavukcuoglu, Marc'Aurelio Ranzato, and Yann LeCun. What is the
best multi-stage architecture for object recognition? In Proc. International Conference on
Computer Vision (ICCV'09) , pages 2146{2153. IEEE, 2009a.
Kevin Jarrett, Koray Kavukcuoglu, Marc'Aurelio Ranzato, and Yann LeCun. What is the best
multi-stage architecture for object recognition? In ICCV'09 , 2009b.
29
Faisal Khan, Xiaojin Zhu, and Bilge Mutlu. How do humans teach: On curriculum learning and
teaching dimension. In Advances in Neural Information Processing Systems 24 (NIPS'11) ,
pages 1449{1457, 2011.
Alex Krizhevsky, Ilya Sutskever, and Geo rey Hinton. ImageNet classi cation with deep
convolutional neural networks. In Advances in Neural Information Processing Systems 25
(NIPS'2012) . 2012.
Kai A. Krueger and Peter Dayan. Flexible shaping: how learning in small steps helps. Cognition ,
110:380{394, 2009.
G. Kunapuli, K.P. Bennett, R. Maclin, and J.W. Shavlik. The adviceptron: Giving advice to
the perceptron. Proceedings of the Conference on Arti cial Neural Networks In Engineering
(ANNIE 2010) , 2010.
Hugo Larochelle, Yoshua Bengio, Jerome Louradour, and Pascal Lamblin. Exploring strategies
for training deep neural networks. Journal of Machine Learning Research , 10:1{40, 2009.
Y. LeCun, L. Bottou, Y. Bengio, and P. Ha ner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE , 86(11):2278{2324, 1998.
T.M. Mitchell. The need for biases in learning generalizations . Department of Computer
Science, Laboratory for Computer Science Research, Rutgers Univ., 1980.
T.M. Mitchell and S.B. Thrun. Explanation-based neural network learning for robot control.
Advances in Neural information processing systems , pages 287{287, 1993.
R. Montague. Universal grammar. Theoria , 36(3):373{398, 1970.
V. Nair and G. E Hinton. Recti ed linear units improve restricted Boltzmann machines. In
ICML'10 , 2010.
L.B.J.H.F.R.A. Olshen and C.J. Stone. Classi cation and regression trees. Belmont, Calif.:
Wadsworth , 1984.
Joseph O'Sullivan. Integrating initialization bias and search bias in neural network learning,
1996.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in python. The
Journal of Machine Learning Research , 12:2825{2830, 2011.
Gail B. Peterson. A day of great illumination: B. F. Skinner's discovery of shaping. Journal of
the Experimental Analysis of Behavior , 82(3):317{328, 2004.
Tapani Raiko, Harri Valpola, and Yann LeCun. Deep learning made easier by linear transfor-
mations in perceptrons. In International Conference on Arti cial Intelligence and Statistics ,
pages 924{932, 2012.
Salah Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and Yoshua Bengio. Contractive
auto-encoders: Explicit invariance during feature extraction. In ICML'2011 , 2011.
30
Salah Rifai, Yoshua Bengio, Yann Dauphin, and Pascal Vincent. A generative process for sam-
pling contractive auto-encoders. In Proceedings of the Twenty-nine International Conference
on Machine Learning (ICML'12) . ACM, 2012. URL http://icml.cc/discuss/2012/590.
html .
R. Salakhutdinov and G.E. Hinton. Deep Boltzmann machines. In Proceedings of the Twelfth
International Conference on Arti cial Intelligence and Statistics (AISTATS 2009) , volume 8,
2009.
Burrhus F. Skinner. Reinforcement today. American Psychologist , 13:94{99, 1958.
R.J. Solomono . A system for incremental learning based on algorithmic probability. In Proceed-
ings of the Sixth Israeli Conference on Arti cial Intelligence, Computer Vision and Pattern
Recognition , pages 515{527. Citeseer, 1989.
G.G. Towell and J.W. Shavlik. Knowledge-based arti cial neural networks. Arti cial intelli-
gence , 70(1):119{165, 1994.
Tommi Vatanen, Tapani Raiko, Harri Valpola, and Yann LeCun. Pushing stochastic gradient
towards second-order methods{backpropagation learning with transformations in nonlinear-
ities. arXiv preprint arXiv:1301.3476 , 2013.
Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol.
Stacked denoising autoencoders: Learning useful representations in a deep network with a
local denoising criterion. Journal of Machine Learning Research , 11:3371{3408, December
2010.
Luis Von Ahn, Manuel Blum, Nicholas J Hopper, and John Langford. Captcha: Using hard
ai problems for security. In Advances in CryptologyEUROCRYPT 2003 , pages 294{311.
Springer, 2003.
Jason Weston, Fr ed eric Ratle, and Ronan Collobert. Deep learning via semi-supervised em-
bedding. In William W. Cohen, Andrew McCallum, and Sam T. Roweis, editors, Pro-
ceedings of the Twenty- fth International Conference on Machine Learning (ICML'08) ,
pages 1168{1175, New York, NY, USA, 2008. ACM. ISBN 978-1-60558-205-4. doi:
10.1145/1390156.1390303.
Matthew D Zeiler. Adadelta: An adaptive learning rate method. arXiv preprint
arXiv:1212.5701 , 2012.
5. Appendix
5.1 Binary-Binary RBMs on Pentomino Dataset
We trained binary-binary RBMs (both visible and hidden are binary) on 8 8 patches extracted
from the Pentomino Dataset using PCD (stochastic maximum likelihood), a weight decay of
.0001 and a sparsity penalty9. We used 256 hidden units and trained by SGD with a batch
size of 32 and a annealing learning rate (Bengio, 2013b) starting from 1e-3 with annealing rate
9. implemented as TorontoSparsity in pylearn2, see the yaml le in the repository for more details
31
1.000015. The RBM is trained with momentum starting from 0.5. The biases are initialized to
-2 in order to get a sparse representation. The RBM is trained for 120 epochs (approximately
50 million updates).
After pretraining the RBM, its parameters are used to initialize the rst layer of an SMLP-
nohints network. As in the usual architecture of the SMLP-nohints on top of P1NN, there is
an intermediate layer. Both P1NN and the intermediate layer have a sigmoid nonlinearity, and
the intermediate layer has 11 units per location. This SMLP-nohints is trained with Adadelta
and standardization at the intermediate layer10.
0 5 10 15 20 25 30 35 40 45
Epoch0.00.10.20.30.40.5Error percentageRBM Test and Training Errors
Training Error
Test Error
Figure 16: Training and test errors of an SMLP-nohints network whose rst layer is pre-trained
as an RBM. Training error reduces to 0% at epoch 42, but test error is still chance.
5.2 Experimental Setup and Hyper-parameters
5.2.1 Decision Trees
We used the decision tree implementation in the scikit-learn (Pedregosa et al., 2011) python
package which is an implementation of the CART (Regression Trees) algorithm. The CART
algorithm constructs the decision tree recursively and partitions the input space such that the
samples belonging to the same category are grouped together (Olshen and Stone, 1984). We
used The Gini index as the impurity criteria. We evaluated the hyper-parameter con gura-
tions with a grid-search. We cross-validated the maximum depth ( maxdepth ) of the tree (for
preventing the algorithm to severely over t the training set) and minimum number of samples
10. In our auto-encoder experiments we directly fed features to P2NN without standardization and Adadelta.
32
Figure 17: Filters learned by the binary-binary RBM after training on the 40k examples. The
RBM did learn the edge structure of Pentomino shapes.
Figure 18: 100 samples generated from trained RBM. All the generated samples are valid Pen-
tomino shapes.
33
required to create a split ( minsplit). 20 di erent con gurations of hyper-parameter values were
evaluated. We obtained the best validation error with maxdepth = 300 andminsplit = 8.
5.2.2 Support Vector Machines
We used the \Support Vector Classi er (SVC)" implementation from the scikit-learn package
which in turn uses the libsvm's Support Vector Machine (SVM) implementation. Kernel-
based SVMs are non-parametric models that map the data into a high dimensional space and
separate di erent classes with hyperplane(s) such that the support vectors for each category
will be separated by a large margin. We cross-validated three hyper-parameters of the model
using grid-search: C,
and the type of kernel( kerneltype).Cis the penalty term (weight
decay) for the SVM and
is a hyper-parameter that controls the width of the Gaussian for
the RBF kernel. For the polynomial kernel,
controls the
exibility of the classi er (degree
of the polynomial) as the number of parameters increases (Hsu et al., 2003; Ben-Hur and
Weston, 2010). We evaluated forty-two hyper-parameter con gurations. That includes, two
kernel types:fRBF; Polynomial g; three gammas:f1e 2;1e 3;1e 4gfor the RBF kernel,
f1;2;5gfor the polynomial kernel, and seven Cvalues among:f0:1;1;2;4;8;10;16g. As a
result of the grid search and cross-validation, we have obtained the best test error by using the
RBF kernel, with C= 2 and
= 1.
5.2.3 Multi Layer Perceptron
We have our own implementation of Multi Layer Perceptron based on the Theano (Bergstra
et al., 2010) machine learning libraries. We have selected 2 hidden layers, the recti er activation
function, and 2048 hidden units per layer. We cross-validated three hyper-parameters of the
model using random-search, sampling the learning rates in log-domain, and selecting L1
andL2 regularization penalty coecients in sets of xed values, evaluating 64 hyperparameter
values. The range of the hyperparameter values are 2[0:0001;1],L12f0:;1e 6;1e 5;1e 4g
andL22f0;1e 6;1e 5g. As a result, the following were selected: L1 = 1e 6,L2 = 1e 5
and= 0:05.
5.2.4 Random Forests
We used scikit-learn's implementation of \Random Forests" decision tree learning. The Ran-
dom Forests algorithm creates an ensemble of decision trees by randomly selecting for each tree
a subset of features and applying bagging to combine the individual decision trees (Breiman,
2001). We have used grid-search and cross-validated the maxdepth ,minsplit, and number
of trees (nestimators ). We have done the grid-search on the following hyperparameter val-
ues,nestimators2 f5;10;15;25;50g,maxdepth2 f100;300;600;900g, andminsplits2
f1;4;16g. We obtained the best validation error with maxdepth = 300,minsplit = 4 and
nestimators = 10.
5.2.5 k-Nearest Neighbors
We used scikit-learn's implementation of k-Nearest Neighbors (k-NN). k-NN is an instance-
based, lazy learning algorithm that selects the training examples closest in Euclidean distance
to the input query. It assigns a class label to the test example based on the categories of the
kclosest neighbors. The hyper-parameters we have evaluated in the cross-validation are the
number of neighbors ( k) andweights . Theweights hyper-parameter can be either \uniform" or
34
\distance". With \uniform", the value assigned to the query point is computed by the majority
vote of the nearest neighbors. With \distance", each value assigned to the query point is
computed by weighted majority votes where the weights are computed with the inverse distance
between the query point and the neighbors. We have used nneighbours2f1;2;4;6;8;12gand
weights2f"uniform ";"distance "gfor hyper-parameter search. As a result of cross-validation
and grid search, we obtained the best validation error with k= 2 andweights =\uniform".
5.2.6 Convolutional Neural Nets
We used a Theano (Bergstra et al., 2010) implementation of Convolutional Neural Networks
(CNN) from the deep learning tutorial at deeplearning.net , which is based on a vanilla
version of a CNN LeCun et al. (1998). Our CNN has two convolutional layers. Following
each convolutional layer, we have a max-pooling layer. On top of the convolution-pooling-
convolution-pooling layers there is an MLP with one hidden layer. In the cross-validation we
have sampled 36 learning rates in log-domain in the range [0 :0001;1] and the number of lters
from the range [10 ;20;30;40;50;60] uniformly. For the rst convolutional layer we used 9 9
receptive elds in order to guarantee that each object ts inside the receptive eld. As a result
of random hyperparameter search and doing manual hyperparameter search on the validation
dataset, the following values were selected:
The number of features used for the rst layer is 30 and the second layer is 60.
For the second convolutional layer, 7 7 receptive elds. The stride for both convolutional
layers is 1.
Convolved images are downsampled by a factor of 2 2 at each pooling operation.
The learning rate for CNN is 0.01 and it was trained for 8 epochs.
5.2.7 Maxout Convolutional Neural Nets
We used the pylearn2 ( https://github.com/lisa-lab/pylearn2 ) implementation of maxout
convolutional networks (Goodfellow et al., 2013). There are two convolutional layers in the
selected architecture, without any pooling. In the last convolutional layer, there is a maxout
non-linearity. The following were selected by cross-validation: learning rate, number of channels
for the both convolution layers, number of kernels for the second layer and number of units and
pieces per maxout unit in the last layer, a linearly decaying learning rate, momentum starting
from 0.5 and saturating to 0.8 at the 200'th epoch. Random search for the hyperparameters
was used to evaluate 48 di erent hyperparameter con gurations on the validation dataset. For
the rst convolutional layer, 8 8 kernels were selected to make sure that each Pentomino shape
ts into the kernel. Early stopping was used and test error on the model that has the best
validation error is reported. Using norm constraint on the fan-in of the nal softmax units
yields slightly better result on the validation dataset.
As a result of cross-validation and manually tuning the hyperparameters we used the fol-
lowing hyperparameters:
16 channels per convolutional layer. 600 hidden units for the maxout layer.
6x6 kernels for the second convolutional layer.
35
5 pieces for the convolution layers and 4 pieces for the maxout layer per maxout units.
We decayed the learning rate by the factor of 0.001 and the initial learning rate is 0.026367.
But we scaled the learning rate of the second convolutional layer by a constant factor of
0.6.
The norm constraint (on the incoming weights of each unit) is 1.9365.
Figure 19 shows the rst layer lters of the maxout convolutional net, after being trained
on the 80k training set for 85 epochs.
Figure 19: Maxout convolutional net rst layer lters. Most of the lters were able to learn
the basic edge structure of the Pentomino shapes.
5.2.8 Stacked Denoising Auto-Encoders
Denoising Auto-Encoders (DAE) are a form of regularized auto-encoder (Bengio et al., 2013).
The DAE forces the hidden layer to discover more robust features and prevents it from simply
learning the identity by reconstructing the input from a corrupted version of it (Vincent et al.,
2010). Two DAEs were stacked, resulting in an unsupervised transformation with two hidden
layers of 1024 units each. Parameters of all layers are then ne-tuned with supervised ne-
tuning using logistic regression as the classi er and SGD as the gradient-based optimization
algorithm. The stochastic corruption process is binomial (0 or 1 replacing each input value,
with probability 0.2). The selected learning rate is 0= 0:01 for the DAe and 1= 0:1 for
supervised ne-tuning. Both L1 and L2 penalty for the DAEs and for the logistic regression
layer are set to 1e-6.
CAE+MLP with Supervised Finetuning: A regularized auto-encoder which sometimes
outperforms the DAE is the Contractive Auto-Encoder (CAE), (Rifai et al., 2012), which
penalizes the Frobenius norm of the Jacobian matrix of derivatives of the hidden units with
respect to the CAE's inputs. The CAE serves as pre-training for an MLP, and in the supervised
ne-tuning state, the Adagrad method was used to automatically tune the learning rate (Duchi
et al., 2010).
After training a CAE with 100 sigmoidal units patch-wise, the features extracted on each
patch are concatenated and fed as input to an MLP. The selected Jacobian penalty coecient
is 2, the learning rate for pre-training is 0.082 with batch size of 200 and 200 epochs of un-
supervised learning are performed on the training set. For supervised netuning, the learning
rate is 0.12 over 100 epochs, L1 and L2 regularization penalty terms respectively are 1e-4 and
1e-6, and the top-level MLP has 6400 hidden units.
36
Greedy Layerwise CAE+DAE Supervised Finetuning: For this experiment we stack a
CAE with sigmoid non-linearities and then a DAE with recti er non-linearities during the pre-
training phase. As recommended by Glorot et al. (2011b) we have used a softplus nonlinearity
for reconstruction, softplus (x) =log(1 +ex). We used an L1 penalty on the recti er outputs
to obtain a sparser representation with recti er non-linearity and L2 regularization to keep the
non-zero weights small.
The main di erence between the DAE and CAE is that the DAE yields more robust recon-
struction whereas the CAE obtains more robust features (Rifai et al., 2011).
As seen on Figure 7 the weights U and V are shared on each patch and we concatenate the
outputs of the last auto-encoder on each patch to feed it as an input to an MLP with a large
hidden layer.
We used 400 hidden units for the CAE and 100 hidden units for DAE. The learning rate
used for the CAE is 0.82 and for DAE it is 9*1e-3. The corruption level for the DAE (binomial
noise) is 0.25 and the contraction level for the CAE is 2.0. The L1 regularization penalty for
the DAE is 2.25*1e-4 and the L2 penalty is 9.5*1e-5. For the supervised netuning phase the
learning rate used is 4*1e-4 with L1 and L2 penalties respectively 1e-5 and 1e-6. The top-level
MLP has 6400 hidden units. The auto-encoders are each trained for 150 epochs while the whole
MLP is ne-tuned for 50 epochs.
Greedy Layerwise DAE+DAE Supervised Finetuning: For this architecture, we have
trained two layers of denoising auto-encoders greedily and performed supervised netuning
after unsupervised pre-training. The motivation for using two denoising auto-encoders is the
fact that recti er nonlinearities work well with the deep networks but it is dicult to train
CAEs with the recti er non-linearity. We have used the same type of denoising auto-encoder
that is used for the greedy layerwise CAE+DAE supervised netuning experiment.
In this experiment we have used 400 hidden units for the rst layer DAE and 100 hidden
units for the second layer DAE. The other hyperparameters for DAE and supervised netuning
are the same as with the CAE+DAE MLP Supervised Finetuning experiment.
37 | [] |
1312.4659 | DeepPose: Human Pose Estimation via Deep Neural Networks | We propose a method for human pose estimation based on Deep Neural Networks
(DNNs). The pose estimation is formulated as a DNN-based regression problem
towards body joints. We present a cascade of such DNN regressors which results
in high precision pose estimates. The approach has the advantage of reasoning
about pose in a holistic fashion and has a simple but yet powerful formulation
which capitalizes on recent advances in Deep Learning. We present a detailed
empirical analysis with state-of-art or better performance on four academic
benchmarks of diverse real-world images. | http://arxiv.org/pdf/1312.4659 | [
"Alexander Toshev",
"Christian Szegedy"
] | [
"cs.CV"
] | IEEE Conference on Computer Vision and Pattern Recognition, 2014 | null | cs.CV | 20131217 | 20140820 | DeepPose: Human Pose Estimation via Deep Neural Networks
Alexander Toshev Christian Szegedy
Google
1600 Amphitheatre Pkwy
Mountain View, CA 94043
toshev,szegedy@google.com
Figure 1. Besides extreme variability in articulations, many of the
joints are barely visible. We can guess the location of the right
arm in the left image only because we see the rest of the pose and
anticipate the motion or activity of the person. Similarly, the left
body half of the person on the right is not visible at all. These
are examples of the need for holistic reasoning . We believe that
DNNs can naturally provide such type of reasoning.
Abstract
We propose a method for human pose estimation based
on Deep Neural Networks (DNNs). The pose estimation
is formulated as a DNN-based regression problem towards
body joints. We present a cascade of such DNN regres-
sors which results in high precision pose estimates. The
approach has the advantage of reasoning about pose in a
holistic fashion and has a simple but yet powerful formula-
tion which capitalizes on recent advances in Deep Learn-
ing. We present a detailed empirical analysis with state-of-
art or better performance on four academic benchmarks of
diverse real-world images.
1. Introduction
The problem of human pose estimation, defined as the
problem of localization of human joints, has enjoyed sub-
stantial attention in the computer vision community. In
Fig. 1, one can see some of the challenges of this prob-
lem – strong articulations, small and barely visible joints,
occlusions and the need to capture the context.
The main stream of work in this field has been motivatedmainly by the first challenge, the need to search in the large
space of all possible articulated poses. Part-based models
lend themselves naturally to model articulations ([16, 8])
and in the recent years a variety of models with efficient
inference have been proposed ([6, 19]).
The above efficiency, however, is achieved at the cost of
limited expressiveness – the use of local detectors, which
reason in many cases about a single part, and most impor-
tantly by modeling only a small subset of all interactions
between body parts. These limitations, as exemplified in
Fig. 1, have been recognized and methods reasoning about
pose in a holistic manner have been proposed [15, 21] but
with limited success in real-world problems.
In this work we ascribe to this holistic view of human
pose estimation. We capitalize on recent developments of
deep learning and propose a novel algorithm based on a
Deep Neural Network (DNN). DNNs have shown outstand-
ing performance on visual classification tasks [14] and more
recently on object localization [23, 9]. However, the ques-
tion of applying DNNs for precise localization of articulated
objects has largely remained unanswered. In this paper we
attempt to cast a light on this question and present a simple
and yet powerful formulation of holistic human pose esti-
mation as a DNN.
We formulate the pose estimation as a joint regression
problem and show how to successfully cast it in DNN set-
tings. The location of each body joint is regressed to using
as an input the full image and a 7-layered generic convolu-
tional DNN. There are two advantages of this formulation.
First, the DNN is capable of capturing the full context of
each body joint – each joint regressor uses the full image
as a signal. Second, the approach is substantially simpler
to formulate than methods based on graphical models – no
need to explicitly design feature representations and detec-
tors for parts; no need to explicitly design a model topology
and interactions between joints. Instead, we show that a
generic convolutional DNN can be learned for this problem.
Further, we propose a cascade of DNN-based pose pre-
dictors. Such a cascade allows for increased precision of
1arXiv:1312.4659v3 [cs.CV] 20 Aug 2014
joint localization. Starting with an initial pose estimation,
based on the full image, we learn DNN-based regressors
which refines the joint predictions by using higher resolu-
tion sub-images.
We show state-of-art results or better than state-of-art on
four widely used benchmarks against all reported results.
We show that our approach performs well on images of peo-
ple which exhibit strong variation in appearance as well as
articulations. Finally, we show generalization performance
by cross-dataset evaluation.
2. Related Work
The idea of representing articulated objects in general,
and human pose in particular, as a graph of parts has been
advocated from the early days of computer vision [16]. The
so called Pictorial Strictures (PSs), introduced by Fishler
and Elschlager [8], were made tractable and practical by
Felzenszwalb and Huttenlocher [6] using the distance trans-
form trick. As a result, a wide variety of PS-based models
with practical significance were subsequently developed.
The above tractability, however, comes with the limita-
tion of having a tree-based pose models with simple binary
potential not depending on image data. As a result, research
has focused on enriching the representational power of the
models while maintaining tractability. Earlier attempts to
achieve this were based on richer part detectors [19, 1, 4].
More recently, a wide variety of models expressing complex
joint relationships were proposed. Yang and Ramanan [27]
use a mixture model of parts. Mixture models on the full
model scale, by having mixture of PSs, have been studied
by Johnson and Everingham [13]. Richer higher-order spa-
tial relationships were captured in a hierarchical model by
Tian et al. [25]. A different approach to capture higher-
order relationship is through image-dependent PS models,
which can be estimated via a global classifier [26, 20, 18].
Approaches which ascribe to our philosophy of reason-
ing about pose in a holistic manner have shown limited
practicality. Mori and Malik [15] try to find for each test
image the closest exemplar from a set of labeled images
and transfer the joint locations. A similar nearest neighbor
setup is employed by Shakhnarovich et al. [21], who how-
ever use locality sensitive hashing. More recently, Gkioxari
et al. [10] propose a semi-global classifier for part config-
uration. This formulation has shown very good results on
real-world data, however, it is based on linear classifiers
with less expressive representation than ours and is tested
on arms only. Finally, the idea of pose regression has been
employed by Ionescu et al. [11], however they reason about
3D pose.
The closest work to ours uses convolution NNs together
with Neighborhood Component Analysis to regress toward
a point in an embedding representing pose [24]. However,
this work does not employ a cascade of networks. Cascadesof DNN regressors have been used for localization, however
of facial points [22]. On the related problem of face pose
estimation, Osadchy et al. [17] employ a NN-based pose
embedding trained with a contrastive loss.
3. Deep Learning Model for Pose Estimation
We use the following notation. To express a pose, we en-
code the locations of all kbody joints in pose vector defined
asy= (:::;yT
i;:::)T;i2f1;:::;kg, where yicontains
thexandycoordinates of the ithjoint. A labeled image is
denoted by (x;y)wherexstands for the image data and y
is the ground truth pose vector.
Further, since the joint coordinates are in absolute image
coordinates, it proves beneficial to normalize them w. r. t. a
boxbbounding the human body or parts of it. In a trivial
case, the box can denote the full image. Such a box is de-
fined by its center bc2R2as well as width bwand height
bh:b= (bc;bw;bh). Then the joint yican be translated by
the box center and scaled by the box size which we refer to
as normalization by b:
N(yi;b) =1=bw 0
0 1=bh
(yi bc) (1)
Further, we can apply the same normalization to the ele-
ments of pose vector N(y;b) = (:::;N (yi;b)T;:::)Tre-
sulting in a normalized pose vector . Finally, with a slight
abuse of notation, we use N(x;b)to denote a crop of the
imagexby the bounding box b, which de facto normalizes
the image by the box. For brevity we denote by N()nor-
malization with bbeing the full image box.
3.1. Pose Estimation as DNN-based Regression
In this work, we treat the problem of pose estimation as
regression, where the we train and use a function (x;)2
R2kwhich for an image xregresses to a normalized pose
vector, where denotes the parameters of the model. Thus,
using the normalization transformation from Eq. (1) the
pose prediction yin absolute image coordinates reads
y=N 1( (N(x);)) (2)
Despite its simple formulation, the power and complex-
ity of the method is in , which is based on a convolutional
Deep Neural Network (DNN). Such a convolutional net-
work consists of several layers – each being a linear trans-
formation followed by a non-linear one. The first layer takes
as input an image of predefined size and has a size equal to
the number of pixels times three color channels. The last
layer outputs the target values of the regression, in our case
2kjoint coordinates.
We base the architecture of the on the work by
Krizhevsky et al. [14] for image classification since it has
shown outstanding results on object localization as well
(xi, yi)
(x(s-1)i, y (s-1) i)xsi - x(s-1)iysi - y(s-1)iInitial stageStage s
send refined values to next stage220 x 220
DNN-based regressor
27 x 27 x 128
13 x 13 x 192
13 x 13 x192
13 x 13 x192
4096
4096
55 x 55 x 48xiyi...DNN-based refiner
27 x 27 x 128
13 x 13 x 192
13 x 13 x192
13 x 13 x192
4096
4096
55 x 55 x 48Figure 2. Left: schematic view of the DNN-based pose regression. We visualize the network layers with their corresponding dimensions,
where convolutional layers are in blue, while fully connected ones are in green. We do not show the parameter free layers. Right: at stage
s, a refining regressor is applied on a sub image to refine a prediction from the previous stage.
[23]. In a nutshell, the network consists of 7layers (see
Fig. 2 left). Denote by Ca convolutional layer, by LRN
a local response normalization layer, Pa pooling layer
and byFa fully connected layer. Only CandFlayers
contain learnable parameters, while the rest are parame-
ter free. Both CandFlayers consist of a linear trans-
formation followed by a nonlinear one, which in our case
is a rectified linear unit. For Clayers, the size is de-
fined as widthheightdepth, where the first two di-
mensions have a spatial meaning while the depth defines
the number of filters. If we write the size of each layer in
parentheses, then the network can be described concisely
asC(555596) LRN P C(2727256)
LRN P C(1313384) C(1313384)
C(1313256) P F(4096) F(4096) . The filter
size for the first two Clayers is 1111and55and for
the remaining three is 33. Pooling is applied after three
layers and contributes to increased performance despite the
reduction of resolution. The input to the net is an image
of220220which via stride of 4is fed into the network.
The total number of parameters in the above model is about
40M. For further details, we refer the reader to [14].
The use of a generic DNN architecture is motivated by
its outstanding results on both classification and localization
problems. In the experimental section we show that such a
generic architecture can be used to learn a model resulting
in state-of-art or better performance on pose estimation as
well. Further, such a model is a truly holistic one — the
final joint location estimate is based on a complex nonlinear
transformation of the full image.
Additionally, the use of a DNN obviates the need to de-
sign a domain specific pose model. Instead such a model
and the features are learned from the data. Although the re-
gression loss does not model explicit interactions between
joints, such are implicitly captured by all of the 7hidden
layers – all the internal features are shared by all joint re-
gressors.
Training The difference to [14] is the loss. Instead of a
classification loss, we train a linear regression on top of thelast network layer to predict a pose vector by minimizing
L2distance between the prediction and the true pose vec-
tor. Since the ground truth pose vector is defined in abso-
lute image coordinates and poses vary in size from image to
image, we normalize our training set Dusing the normal-
ization from Eq. (1):
DN=f(N(x);N(y))j(x;y)2Dg (3)
Then theL2loss for obtaining optimal network parameters
reads:
arg min
X
(x;y)2DNkX
i=1jjyi i(x;)jj2
2 (4)
For clarity we write out the optimization over individual
joints. It should be noted, that the above objective can
be used even if for some images not all joints are labeled.
In this case, the corresponding terms in the sum would be
omitted.
The above parameters are optimized for using Back-
propagation in a distributed online implementation. For
each mini-batch of size 128, adaptive gradient updates are
computed [3]. The learning rate, as the most important pa-
rameter, is set to 0:0005 . Since the model has large number
of parameters and the used datasets are of relatively small
size, we augment the data using large number of randomly
translated image crops (see Sec. 3.2), left/right flips as well
as DropOut regularization for the Flayers set to 0:6.
3.2. Cascade of Pose Regressors
The pose formulation from the previous section has the
advantage that the joint estimation is based on the full im-
age and thus relies on context. However, due to its fixed
input size of 220220, the network has limited capacity
to look at detail – it learns filters capturing pose properties
at coarse scale. These are necessary to estimate rough pose
but insufficient to always precisely localize the body joints.
Note that we cannot easily increase the input size since this
will increase the already large number of parameters. In or-
der to achieve better precision, we propose to train a cascade
of pose regressors. At the first stage, the cascade starts off
by estimating an initial pose as outlined in the previous sec-
tion. At subsequent stages, additional DNN regressors are
trained to predict a displacement of the joint locations from
previous stage to the true location. Thus, each subsequent
stage can be thought of as a refinement of the currently pre-
dicted pose, as shown in Fig. 2.
Further, each subsequent stage uses the predicted joint
locations to focus on the relevant parts of the image – sub-
images are cropped around the predicted joint location from
previous stage and the pose displacement regressor for this
joint is applied on this sub-image. In this way, subsequent
pose regressors see higher resolution images and thus learn
features for finer scales which ultimately leads to higher
precision.
We use the same network architecture for all stages of
the cascade but learn different network parameters. For
stages2 f1;:::;Sgof totalScascade stages, we de-
note bysthe learned network parameters. Thus, the
pose displacement regressor reads (x;s). To refine a
given joint location yiwe will consider a joint bounding
boxbicapturing the sub-image around yi:bi(y;) =
(yi;diam(y);diam(y))having as center the i-th joint
and as dimension the pose diameter scaled by . The diam-
eter diam (y)of the pose is defined as the distance between
opposing joints on the human torso, such as left shoulder
and right hip, and depends on the concrete pose definition
and dataset.
Using the above notation, at the stage s= 1we start with
a bounding box b0which either encloses the full image or
is obtained by a person detector. We obtain an initial pose:
Stage 1 : y1 N 1( (N(x;b0);1);b0)(5)
At each subsequent stage s2, for all joints i2f1;:::;kg
we regress first towards a refinement displacement ys
i
y(s 1)
i by applying a regressor on the sub image defined
byb(s 1)
i from previous stage (s 1). Then, we estimate
new joint boxes bs
i:
Stages:ys
i y(s 1)
i +N 1( i(N(x;b);s);b)(6)
forb=b(s 1)
i
bs
i (ys
i;diam(ys);diam(ys)) (7)
We apply the cascade for a fixed number of stages S,
which is determined as explained in Sec. 4.1.
Training The network parameters 1are trained as
outlined in Sec. 3.1, Eq. (4). At subsequent stages
s2, the training is done identically with one im-
portant difference. Each joint ifrom a training exam-
ple(x;y)is normalized using a different bounding box(y(s 1)
i;diam(y(s 1));diam(y(s 1)))– the one cen-
tered at the prediction for the same joint obtained from pre-
vious stage – so that we condition the training of the stage
based on the model from previous stage.
Since deep learning methods have large capacity, we
augment the training data by using multiple normalizations
for each image and joint. Instead of using the prediction
from previous stage only, we generate simulated predic-
tions. This is done by randomly displacing the ground truth
location for joint iby a vector sampled at random from a
2-dimensional Normal distribution N(s 1)
i with mean and
variance equal to the mean and variance of the observed dis-
placements (y(s 1)
i yi)across all examples in the train-
ing data. The full augmented training data can be defined
by first sampling an example and a joint from the original
data at uniform and then generating a simulated prediction
based on a sampled displacement fromN(s 1)
i :
Ds
A=f(N(x;b);N(yi;b))j
(x;yi)D;N(s 1)
i;
b= (yi+;diam(y))g
The training objective for cascade stage sis done as in
Eq. (4) by taking extra care to use the correct normalization
for each joint:
s= arg min
X
(x;yi)2Ds
Ajjyi i(x;)jj2
2 (8)
4. Empirical Evaluation
4.1. Setup
Datasets There is a wide variety of benchmarks for hu-
man pose estimation. In this work we use datasets, which
have large number of training examples sufficient to train a
large model such as the proposed DNN, as well as are real-
istic and challenging.
The first dataset we use is Frames Labeled In Cinema
(FLIC), introduced by [20], which consists of 4000 train-
ing and 1000 test images obtained from popular Hollywood
movies. The images contain people in diverse poses and es-
pecially diverse clothing. For each labeled human, 10upper
body joints are labeled.
The second dataset we use is Leeds Sports Dataset [12]
and its extension [13], which we will jointly denote by LSP.
Combined they contain 11000 training and 1000 testing im-
ages. These are images from sports activities and as such
are quite challenging in terms of appearance and especially
articulations. In addition, the majority of people have 150
pixel height which makes the pose estimation even more
challenging. In this dataset, for each person the full body is
labeled with total 14joints.
For all of the above datasets, we define the diameter of a
poseyto be the distance between a shoulder and hip from
opposing sides and denote it by diam (y). It should be noted,
that the joints in all datasets are arranged in a tree kinemat-
ically mimicking the human body. This allows for a defini-
tion of a limb being a pair of neighboring joints in the pose
tree.
Metrics In order to be able to compare with published re-
sults we will use two widely accepted evaluation metrics.
Percentage of Correct Parts (PCP) measures detection rate
of limbs, where a limb is considered detected if the distance
between the two predicted joint locations and the true limb
joint locations is at most half of the limb length [5]. PCP
was the initially preferred metric for evaluation, however it
has the drawback of penalizing shorter limbs, such as lower
arms, which are usually harder to detect.
To address this drawback, recently detection rates of
joints are being reported using a different detection crite-
rion – a joint is considered detected if the distance between
the predicted and the true joint is within a certain fraction of
the torso diameter. By varying this fraction, detection rates
are obtained for varying degrees of localization precision.
This metric alleviates the drawback of PCP since the de-
tection criteria for all joints are based on the same distance
threshold. We refer to this metric as Percent of Detected
Joints (PDJ).
Experimental Details For all the experiments we use the
same network architecture. Inspired by [7], we use a body
detector on FLIC to obtain initially a rough estimate of the
human body bounding box. It is based on a face detector –
the detected face rectangle is enlarged by a fixed scaler. This
scaler is determined on the training data such that it contains
all labeled joints. This face-based body detector results in
a rough estimate, which however presents a good starting
point for our approach. For LSP we use the full image as
initial bounding box since the humans are relatively tightly
cropped by design.
Using a small held-out set of 50 images for both datasets
to determine the algorithm hyperparameters. To measure
optimality of the parameters we used average over PDJ at
0:2across all joints. The scaler , which defines the size of
the refinement joint bounding box as a fraction of the pose
size, is determined as follows: for FLIC we chose = 1:0
after exploring values f0:8;1:0;1:2g, for LSP we use =
2:0after tryingf1:5;1:7;2:0;2:3g. The number of cascade
stagesSis determined by training stages until the algorithm
stopped improving on the held-out set. For both FLIC and
LSP we arrived at S= 3.
To improve generalization, for each cascade stage start-
ing ats= 2 we augment the training data by sampling 40
randomly translated crop boxes for each joint as explained
in Sec. 3.2. Thus, for LSP with 14joints and after mirror-
ing the images and sampling the number training examples
is1100040214 = 12M, which is essential for
training a large network as ours.The presented algorithm allows for an efficient imple-
mentation. The running time is approx. 0:1sper image,
as measured on a 12core CPU. This compares favorably
to other approaches, as some of the current state-of-art ap-
proaches have higher complexity: [20] runs in approx. 4s,
while [27] runs in 1:5s. The training complexity, however,
is higher. The initial stage was trained within 3days on
approx. 100 workers, most of the final performance was
achieved after 12hours though. Each refinement stage was
trained for 7days since the amount of data was 40larger
than the one for the initial stage due to the data augmenta-
tion in Sec. 3.2. Note that using more data led to increased
performance.
4.2. Results and Discussion
Comparisons We present comparative results to other ap-
proaches. We compare on LSP using PCP metric in Fig. 1.
We show results for the four most challenging limbs – lower
and upper arms and legs – as well as the average value
across these limbs for all compared algorithms. We clearly
outperform all other approaches, especially achieving bet-
ter estimation for legs. For example, for upper legs we ob-
tain0:78up from 0:74for the next best performing method.
It is worth noting that while the other approaches exhibit
strengths for particular limbs, none of the other dataset con-
sistently dominates across all limbs. In contrary, DeepPose
shows strong results for all challenging limbs.
Using the PDJ metric allows us to vary the threshold for
the distance between prediction and ground truth, which de-
fines a detection. This threshold can be thought of as a
localization precision at which detection rates are plotted.
Thus one could compare approaches across different de-
sired precisions. We present results on FLIC in Fig. 3 com-
paring against additional four methods as well is on LSP in
Fig. 4. For each dataset we train and test according the pro-
tocol for each dataset. Similarly to previous experiment we
outperform all five algorithms. Our gains are bigger in the
low precision domain, in the cases where we detect rough
pose without precisely localizing the joints. On FLIC, at
normalized distance 0:2we obtain a an increase of detection
rates by 0:15and0:2for elbow and wrists against the next
best performing method. On LSP, at normalized distance
0:5we get an absolute increase of 0:1. At low precision
regime of normalized distance of 0:2for LSP we show com-
parable performance for legs and slightly worse arms. This
can be attributed to the fact that the DNN-based approach
computes joint coordinates using 7 layers of transformation,
some of which contain max pooling.
Another observation is that our approach works well for
both appearance heavy movie data as well as string articu-
lation such as the sports images in LSP.
Effects of cascade-based refinement A single DNN-
based joint regressor gives rough joint location. However,
0 0.05 0.1 0.15 0.20.10.30.50.70.9Elbows
Normalized distance to true jointDetection rate
DeepPoseMODECEichner et al.Yang et al.Sapp et al.
0 0.05 0.1 0.15 0.20.10.30.50.70.9Wrists
Normalized distance to true jointDetection rateFigure 3. Percentage of detected joints (PDJ) on FLIC for two joints: elbow and wrist. We compare DeepPose, after two cascade stages,
with four other approaches.
MethodArm LegAve.Upper Lower Upper Lower
DeepPose-st1 0.5 0.27 0.74 0.65 0.54
DeepPose-st2 0.56 0.36 0.78 0.70 0.60
DeepPose-st3 0.56 0.38 0.77 0.71 0.61
Dantone et al. [2] 0.45 0.25 0.65 0.61 0.49
Tian et al. [25]0.52 0.33 0.70 0.60 0.56
Johnson et al. [13] 0.54 0.38 0.75 0.66 0.58
Wang et al. [26]0.565 0.37 0.76 0.68 0.59
Pishchulin [18]0.49 0.32 0.74 0.70 0.56
Table 1. Percentage of Correct Parts (PCP) at 0.5 on LSP for Deep-
Pose as well as five state-of-art approaches.The authors use a
slightly looser version of PCP in which for each limb the average
distance from predicted limb joints belonging to the correspond-
ing true joints is used to determine whether the limb is correctly
detected.The authors use person-centric joint annotations.
0 0.1 0.2 0.3 0.4 0.50.10.30.50.70.9Arms
Normalized distance to true jointDetection rate
DeepPose − wristsDeepPose − elbowsDantone et al. − wristsDantone et al. − elbows
0 0.1 0.2 0.3 0.4 0.50.10.30.50.70.9Legs
Normalized distance to true jointDetection rate
DeepPose − ankleDeepPose − kneeDantone et al. − ankleDantone et al. − knee
Figure 4. Percentage of detected joints (PDJ) on LSP for four
limbs for DeepPose and Dantone et al. [2] over an extended range
of distances to true joint: [0;0:5]of the torso diameter. Results of
DeepPose are plotted with solid lines while all the results by [2]
are plotted in dashed lines. Results for the same joint from both
algorithms are colored with same color.
to obtain higher precision the subsequent stages of the cas-
cade, which serve as a refinement of the initial prediction,
are of paramount importance. To see this, in Fig. 5 we
present the joint detections at different precisions for the ini-
tial prediction as well as two subsequent cascade stages. As
expected, we can see that the major gains of the refinement
00.05 0.10.15 0.20.25 0.30.35 0.40.10.30.50.70.9Wrists
Normalized distance to true jointDetection rate
DeepPose − initial stage 1DeepPose − stage 2DeepPose − stage 3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.40.10.30.50.70.9Elbows
Normalized distance to true jointDetection rateFigure 5. Percent of detected joints (PDJ) on FLIC or the first three
stages of the DNN cascade. We present results over larger spec-
trum of normalized distances between prediction and ground truth.
procedure are at high-precision regime of at normalized dis-
tances of [0:15;0:2]. Further, the major gains are achieved
after one stage of refinement. The reason being that subse-
quent stages end up using smaller sub-images around each
joint. And although the subsequent stages look at higher
resolution inputs, they have more limited context.
Examples of cases, where refinement helps, are visual-
ized in Fig. 6. The initial stage is usually successful at es-
timating a roughly correct pose, however, this pose is not
”snapped” to the correct one. For example, in row three the
pose has the right shape but incorrect scale. In the second
row, the predicted pose is translated north from the ideal
one. In most cases, the second stage of the cascade resolves
this snapping problem and better aligns the joints. In more
rare cases, such as in first row, further facade stages improve
on individual joints.
Cross-dataset Generalization To evaluate the general-
ization properties of our algorithm, we used the trained
models on LSP and FLIC on two related datasets. The full-
body model trained on LSP is tested on the test portion of
the Image Parse dataset [19] with results presented in Ta-
ble 2. The ImageParse dataset is similar to LSP as it con-
tains people doing sports, however it contains a lot of peo-
ple from personal photo collections involved in other activ-
ities. Further, the upper-body model trained on FLIC was
applied on the whole Buffy dataset [7]. We can see that our
Initial stage 1stage 2stage 3Figure 6. Predicted poses in red and ground truth poses in green
for the first three stages of a cascade for three examples.
0 0.05 0.1 0.15 0.20.10.30.50.70.9Elbows
Normalized distance to true jointDetection rate
Eichner et al.Yang et al.Sapp et al.MODECDeepPose
0 0.05 0.1 0.15 0.20.10.30.50.70.9Wrists
Normalized distance to true jointDetection rate
Figure 7. Percentage of detected joints (PDJ) on Buffy dataset
for two joints: elbow and wrist. The models have been trained on
FLIC. We compare DeepPose, after two cascade stages, with four
other approaches.
MethodArm LegAve.Upper Lower Upper Lower
DeepPose 0.8 0.75 0.71 0.5 0.69
Pishchulin [18] 0.80 0.70 0.59 037 0.62
Johnson et al. [13] 0.75 0.67 0.67 0.46 0.64
Yang et al. [27] 0.69 0.64 0.55 0.35 0.56
Table 2. Percentage of Correct Parts (PCP) at 0.5 on Image Parse
dataset for DeepPose as well as two state-of-art approaches on Im-
age Parse dataset. Results obtained from [18].
approach can retain state-of-art performance compared to
other approaches. This shows good generalization abilities.
Example poses To get a better idea of the performance of
our algorithm, we visualize a sample of estimated poses on
images from LSP in Fig. 8. We can see that our algorithm is
able to get correct pose for most of the joints under variety
of conditions: upside-down people (row 1, column 1), se-
vere foreshortening (row1, column 3), unusual poses (row
3, column 5), occluded limbs as the occluded arms in row
3, columns 2 and 6, unusual illumination conditions (row 3,
column 3). In most of the cases, when the estimated pose isnot precise, it still has a correct shape. For example, in the
last row some of the predicted limbs are not aligned with
the true locations, however the overall shape of the pose is
correct. A common failure mode is confusing left with right
side when the person was photographed from the back (row
6, column 6). Results on FLIC (see Fig. 9) are usually better
with occasional visible mistakes on lower arms.
5. Conclusion
We present, to our knowledge, the first application of
Deep Neural Networks (DNNs) to human pose estimation.
Our formulation of the problem as DNN-based regression to
joint coordinates and the presented cascade of such regres-
sors has the advantage of capturing context and reasoning
about pose in a holistic manner. As a result, we are able to
achieve state-of-art or better results on several challenging
academic datasets.
Further, we show that using a generic convolutional neu-
ral network, which was originally designed for classifica-
tion tasks, can be applied to the different task of localiza-
tion. In future, we plan to investigate novel architectures
which could be potentially better tailored towards localiza-
tion problems in general, and in pose estimation in particu-
lar.
Acknowledgements I would like to thank Luca Bertelli,
Ben Sapp and Tianli Yu for assistance with data and fruitful
discussions.
References
[1] M. Andriluka, S. Roth, and B. Schiele. Pictorial structures
revisited: People detection and articulated pose estimation.
InCVPR , 2009.
[2] M. Dantone, J. Gall, C. Leistner, and L. Van Gool. Human
pose estimation using body parts dependent joint regressors.
InCVPR , 2013.
[3] J. Duchi, E. Hazan, and Y . Singer. Adaptive subgradient
methods for online learning and stochastic optimization. In
COLT . ACL, 2010.
[4] M. Eichner and V . Ferrari. Better appearance models for
pictorial structures. 2009.
[5] M. Eichner, M. Marin-Jimenez, A. Zisserman, and V . Ferrari.
Articulated human pose estimation and search in (almost)
unconstrained still images. ETH Zurich, D-ITET, BIWI,
Technical Report No , 272, 2010.
[6] P. F. Felzenszwalb and D. P. Huttenlocher. Pictorial struc-
tures for object recognition. International Journal of Com-
puter Vision , 61(1):55–79, 2005.
[7] V . Ferrari, M. Marin-Jimenez, and A. Zisserman. Progressive
search space reduction for human pose estimation. In CVPR ,
2008.
[8] M. A. Fischler and R. A. Elschlager. The representation and
matching of pictorial structures. Computers, IEEE Transac-
tions on , 100(1):67–92, 1973.
Figure 8. Visualization of pose results on images from LSP. Each pose is represented as a stick figure, inferred from predicted joints.
Different limbs in the same image are colored differently, same limb across different images has the same color.
Figure 9. Visualization of pose results on images from FLIC. Meaning of stick figures is the same as in Fig. 8 above.
[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich fea-
ture hierarchies for accurate object detection and semantic
segmentation. In CVPR , 2014.
[10] G. Gkioxari, P. Arbel ´aez, L. Bourdev, and J. Malik. Articu-
lated pose estimation using discriminative armlet classifiers.
InCVPR , 2013.
[11] C. Ionescu, F. Li, and C. Sminchisescu. Latent structured
models for human pose estimation. In ICCV , 2011.
[12] S. Johnson and M. Everingham. Clustered pose and nonlin-
ear appearance models for human pose estimation. In BMVC ,
2010.
[13] S. Johnson and M. Everingham. Learning effective human
pose estimation from inaccurate annotation. In CVPR , 2011.
[14] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet clas-
sification with deep convolutional neural networks. In NIPS ,
2012.
[15] G. Mori and J. Malik. Estimating human body configurations
using shape context matching. In ECCV , 2002.[16] R. Nevatia and T. O. Binford. Description and recognition of
curved objects. Artificial Intelligence , 8(1):77–98, 1977.
[17] M. Osadchy, Y . LeCun, and M. L. Miller. Synergistic face
detection and pose estimation with energy-based models.
The Journal of Machine Learning Research , 8:1197–1215,
2007.
[18] L. Pishchulin, M. Andriluka, P. Gehler, and B. Schiele. Pose-
let conditioned pictorial structures. In CVPR , 2013.
[19] D. Ramanan. Learning to parse images of articulated bodies.
InNIPS , 2006.
[20] B. Sapp and B. Taskar. Modec: Multimodal decomposable
models for human pose estimation. In CVPR , 2013.
[21] G. Shakhnarovich, P. Viola, and T. Darrell. Fast pose estima-
tion with parameter-sensitive hashing. In CVPR , 2003.
[22] Y . Sun, X. Wang, and X. Tang. Deep convolutional net-
work cascade for facial point detection. In Computer Vision
and Pattern Recognition (CVPR), 2013 IEEE Conference on ,
pages 3476–3483. IEEE, 2013.
[23] C. Szegedy, A. Toshev, and D. Erhan. Object detection via
deep neural networks. In NIPS 26 , 2013.
[24] G. W. Taylor, R. Fergus, G. Williams, I. Spiro, and C. Bre-
gler. Pose-sensitive embedding by nonlinear nca regression.
InNIPS , 2010.
[25] Y . Tian, C. L. Zitnick, and S. G. Narasimhan. Exploring the
spatial hierarchy of mixture models for human pose estima-
tion. In ECCV , 2012.
[26] F. Wang and Y . Li. Beyond physical connections: Tree mod-
els in human pose estimation. In CVPR , 2013.
[27] Y . Yang and D. Ramanan. Articulated pose estimation with
flexible mixtures-of-parts. In CVPR , 2011. | [] |
1405.3531 | Return of the Devil in the Details: Delving Deep into Convolutional Nets | "The latest generation of Convolutional Neural Networks (CNN) have achieved\nimpressive results in c(...TRUNCATED) | http://arxiv.org/pdf/1405.3531 | [
"Ken Chatfield",
"Karen Simonyan",
"Andrea Vedaldi",
"Andrew Zisserman"
] | [
"cs.CV"
] | Published in proceedings of BMVC 2014 | null | cs.CV | 20140514 | 20141105 | "1\nReturn of the Devil in the Details:\nDelving Deep into Convolutional Nets\nKen Chatfield, Karen(...TRUNCATED) | [] |
1406.1078 | Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation | "In this paper, we propose a novel neural network model called RNN\nEncoder-Decoder that consists of(...TRUNCATED) | http://arxiv.org/pdf/1406.1078 | ["Kyunghyun Cho","Bart van Merrienboer","Caglar Gulcehre","Dzmitry Bahdanau","Fethi Bougares","Holge(...TRUNCATED) | [
"cs.CL",
"cs.LG",
"cs.NE",
"stat.ML"
] | EMNLP 2014 | null | cs.CL | 20140603 | 20140903 | "Learning Phrase Representations using RNN Encoder–Decoder\nfor Statistical Machine Translation\nK(...TRUNCATED) | [] |
1406.4729 | Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition | "Existing deep convolutional neural networks (CNNs) require a fixed-size\n(e.g., 224x224) input imag(...TRUNCATED) | http://arxiv.org/pdf/1406.4729 | [
"Kaiming He",
"Xiangyu Zhang",
"Shaoqing Ren",
"Jian Sun"
] | [
"cs.CV"
] | "This manuscript is the accepted version for IEEE Transactions on\n Pattern Analysis and Machine In(...TRUNCATED) | null | cs.CV | 20140618 | 20150423 | "1\nSpatial Pyramid Pooling in Deep Convolutional\nNetworks for Visual Recognition\nKaiming He, Xian(...TRUNCATED) | [] |
1407.2538 | Learning Deep Structured Models | "Many problems in real-world applications involve predicting several random\nvariables which are sta(...TRUNCATED) | http://arxiv.org/pdf/1407.2538 | [
"Liang-Chieh Chen",
"Alexander G. Schwing",
"Alan L. Yuille",
"Raquel Urtasun"
] | [
"cs.LG"
] | 11 pages including reference | null | cs.LG | 20140709 | 20150427 | "Accepted as a workshop contribution at ICLR 2015\nLEARNING DEEPSTRUCTURED MODELS\nLiang-Chieh Chen\(...TRUNCATED) | [] |
1409.0473 | Neural Machine Translation by Jointly Learning to Align and Translate | "Neural machine translation is a recently proposed approach to machine\ntranslation. Unlike the trad(...TRUNCATED) | http://arxiv.org/pdf/1409.0473 | [
"Dzmitry Bahdanau",
"Kyunghyun Cho",
"Yoshua Bengio"
] | [
"cs.CL",
"cs.LG",
"cs.NE",
"stat.ML"
] | Accepted at ICLR 2015 as oral presentation | null | cs.CL | 20140901 | 20160519 | "Published as a conference paper at ICLR 2015\nNEURAL MACHINE TRANSLATION\nBYJOINTLY LEARNING TO ALI(...TRUNCATED) | [] |
1409.1556 | Very Deep Convolutional Networks for Large-Scale Image Recognition | "In this work we investigate the effect of the convolutional network depth on\nits accuracy in the l(...TRUNCATED) | http://arxiv.org/pdf/1409.1556 | [
"Karen Simonyan",
"Andrew Zisserman"
] | [
"cs.CV"
] | null | null | cs.CV | 20140904 | 20150410 | "arXiv:1409.1556v6 [cs.CV] 10 Apr 2015Publishedasa conferencepaperat ICLR2015\nVERYDEEPCONVOLUTION(...TRUNCATED) | [{"id":"1412.0623","title":"Material Recognition in the Wild with the Materials in Context Database"(...TRUNCATED) |
1411.4038 | Fully Convolutional Networks for Semantic Segmentation | "Convolutional networks are powerful visual models that yield hierarchies of\nfeatures. We show that(...TRUNCATED) | http://arxiv.org/pdf/1411.4038 | [
"Jonathan Long",
"Evan Shelhamer",
"Trevor Darrell"
] | [
"cs.CV"
] | to appear in CVPR (2015) | null | cs.CV | 20141114 | 20150308 | "Fully Convolutional Networks for Semantic Segmentation\nJonathan Long\u0003Evan Shelhamer\u0003Trev(...TRUNCATED) | [] |
End of preview. Expand
in Dataset Viewer.
This dataset contains papers related to (and including) the Llama 2 research paper. Related papers were identified by following a trail of references, extracting those papers with the arxiv-bot package, and repeating.
- Downloads last month
- 64