Content uploaded by Nitish Srivastava

Author content

All content in this area was uploaded by Nitish Srivastava on May 12, 2021

Content may be subject to copyright.

Improving neural networks by preventing

co-adaptation of feature detectors

G. E. Hinton∗, N. Srivastava, A. Krizhevsky, I. Sutskever and R. R. Salakhutdinov

Department of Computer Science, University of Toronto,

6 King’s College Rd, Toronto, Ontario M5S 3G4, Canada

∗To whom correspondence should be addressed; E-mail: hinton@cs.toronto.edu

When a large feedforward neural network is trained on a small training set,

it typically performs poorly on held-out test data. This “overﬁtting” is greatly

reduced by randomly omitting half of the feature detectors on each training

case. This prevents complex co-adaptations in which a feature detector is only

helpful in the context of several other speciﬁc feature detectors. Instead, each

neuron learns to detect a feature that is generally helpful for producing the

correct answer given the combinatorially large variety of internal contexts in

which it must operate. Random “dropout” gives big improvements on many

benchmark tasks and sets new records for speech and object recognition.

A feedforward, artiﬁcial neural network uses layers of non-linear “hidden” units between

its inputs and its outputs. By adapting the weights on the incoming connections of these hidden

units it learns feature detectors that enable it to predict the correct output when given an input

vector (1). If the relationship between the input and the correct output is complicated and the

network has enough hidden units to model it accurately, there will typically be many different

settings of the weights that can model the training set almost perfectly, especially if there is

only a limited amount of labeled training data. Each of these weight vectors will make different

predictions on held-out test data and almost all of them will do worse on the test data than on

the training data because the feature detectors have been tuned to work well together on the

training data but not on the test data.

Overﬁtting can be reduced by using “dropout” to prevent complex co-adaptations on the

training data. On each presentation of each training case, each hidden unit is randomly omitted

from the network with a probability of 0.5, so a hidden unit cannot rely on other hidden units

being present. Another way to view the dropout procedure is as a very efﬁcient way of perform-

ing model averaging with neural networks. A good way to reduce the error on the test set is to

average the predictions produced by a very large number of different networks. The standard

1

arXiv:1207.0580v1 [cs.NE] 3 Jul 2012

way to do this is to train many separate networks and then to apply each of these networks to

the test data, but this is computationally expensive during both training and testing. Random

dropout makes it possible to train a huge number of different networks in a reasonable time.

There is almost certainly a different network for each presentation of each training case but all

of these networks share the same weights for the hidden units that are present.

We use the standard, stochastic gradient descent procedure for training the dropout neural

networks on mini-batches of training cases, but we modify the penalty term that is normally

used to prevent the weights from growing too large. Instead of penalizing the squared length

(L2 norm) of the whole weight vector, we set an upper bound on the L2 norm of the incoming

weight vector for each individual hidden unit. If a weight-update violates this constraint, we

renormalize the weights of the hidden unit by division. Using a constraint rather than a penalty

prevents weights from growing very large no matter how large the proposed weight-update is.

This makes it possible to start with a very large learning rate which decays during learning,

thus allowing a far more thorough search of the weight-space than methods that start with small

weights and use a small learning rate.

At test time, we use the “mean network” that contains all of the hidden units but with their

outgoing weights halved to compensate for the fact that twice as many of them are active.

In practice, this gives very similar performance to averaging over a large number of dropout

networks. In networks with a single hidden layer of Nunits and a “softmax” output layer for

computing the probabilities of the class labels, using the mean network is exactly equivalent

to taking the geometric mean of the probability distributions over labels predicted by all 2N

possible networks. Assuming the dropout networks do not all make identical predictions, the

prediction of the mean network is guaranteed to assign a higher log probability to the correct

answer than the mean of the log probabilities assigned by the individual dropout networks (2).

Similarly, for regression with linear output units, the squared error of the mean network is

always better than the average of the squared errors of the dropout networks.

We initially explored the effectiveness of dropout using MNIST, a widely used benchmark

for machine learning algorithms. It contains 60,000 28x28 training images of individual hand

written digits and 10,000 test images. Performance on the test set can be greatly improved by

enhancing the training data with transformed images (3) or by wiring knowledge about spatial

transformations into a convolutional neural network (4) or by using generative pre-training to

extract useful features from the training images without using the labels (5). Without using any

of these tricks, the best published result for a standard feedforward neural network is 160 errors

on the test set. This can be reduced to about 130 errors by using 50% dropout with separate L2

constraints on the incoming weights of each hidden unit and further reduced to about 110 errors

by also dropping out a random 20% of the pixels (see ﬁgure 1).

Dropout can also be combined with generative pre-training, but in this case we use a small

learning rate and no weight constraints to avoid losing the feature detectors discovered by the

pre-training. The publically available, pre-trained deep belief net described in (5) got 118 errors

when it was ﬁne-tuned using standard back-propagation and 92 errors when ﬁne-tuned using

50% dropout of the hidden units. When the publically available code at URL was used to pre-

2

Fig. 1: The error rate on the MNIST test set for a variety of neural network architectures trained

with backpropagation using 50% dropout for all hidden layers. The lower set of lines also

use 20% dropout for the input layer. The best previously published result for this task using

backpropagation without pre-training or weight-sharing or enhancements of the training set is

shown as a horizontal line.

train a deep Boltzmann machine ﬁve times, the unrolled network got 103, 97, 94, 93 and 88

errors when ﬁne-tuned using standard backpropagation and 83, 79, 78, 78 and 77 errors when

using 50% dropout of the hidden units. The mean of 79 errors is a record for methods that do

not use prior knowledge or enhanced training sets (For details see Appendix A).

We then applied dropout to TIMIT, a widely used benchmark for recognition of clean speech

with a small vocabulary. Speech recognition systems use hidden Markov models (HMMs) to

deal with temporal variability and they need an acoustic model that determines how well a frame

of coefﬁcients extracted from the acoustic input ﬁts each possible state of each hidden Markov

model. Recently, deep, pre-trained, feedforward neural networks that map a short sequence of

frames into a probability distribution over HMM states have been shown to outperform tradional

Gaussian mixture models on both TIMIT (6) and a variety of more realistic large vocabulary

tasks (7, 8).

Figure 2 shows the frame classiﬁcation error rate on the core test set of the TIMIT bench-

mark when the central frame of a window is classiﬁed as belonging to the HMM state that is

given the highest probability by the neural net. The input to the net is 21 adjacent frames with an

advance of 10ms per frame. The neural net has 4 fully-connected hidden layers of 4000 units per

3

Fig. 2: The frame classiﬁcation error rate on the core test set of the TIMIT benchmark. Com-

parison of standard and dropout ﬁnetuning for different network architectures. Dropout of 50%

of the hidden units and 20% of the input units improves classiﬁcation.

layer and 185 “softmax” output units that are subsequently merged into the 39 distinct classes

used for the benchmark. Dropout of 50% of the hidden units signiﬁcantly improves classiﬁca-

tion for a variety of different network architectures (see ﬁgure 2). To get the frame recognition

rate, the class probabilities that the neural network outputs for each frame are given to a decoder

which knows about transition probabilities between HMM states and runs the Viterbi algorithm

to infer the single best sequence of HMM states. Without dropout, the recognition rate is 22.7%

and with dropout this improves to 19.7%, which is a record for methods that do not use any

information about speaker identity.

CIFAR-10 is a benchmark task for object recognition. It uses 32x32 downsampled color

images of 10 different object classes that were found by searching the web for the names of the

class (e.g. dog) or its subclasses (e.g. Golden Retriever). These images were labeled by hand

to produce 50,000 training images and 10,000 test images in which there is a single dominant

object that could plausibly be given the class name (9) (see ﬁgure 3). The best published error

rate on the test set, without using transformed data, is 18.5% (10). We achieved an error rate of

16.6% by using a neural network with three convolutional hidden layers interleaved with three

“max-pooling” layers that report the maximum activity in local pools of convolutional units.

These six layers were followed by one locally-connected layer (For details see Appendix D) .

Using dropout in the last hidden layer gives an error rate of 15.6%.

ImageNet is an extremely challenging object recognition dataset consisting of thousands of

high-resolution images of thousands of classes of object (11). In 2010, a subset of 1000 classes

with roughly 1000 examples per class was the basis of an object recognition competition in

4

Fig. 3: Ten examples of the class “bird” from the CIFAR-10 test set illustrating the variety

of types of bird, viewpoint, lighting and background. The neural net gets all but the last two

examples correct.

which the winning entry, which was actually an average of six separate models, achieved an

error rate of 47.2% on the test set. The current state-of-the-art result on this dataset is 45.7%

(12). We achieved comparable performance of 48.6% error using a single neural network with

ﬁve convolutional hidden layers interleaved with “max-pooling” layer followed by two globally

connected layers and a ﬁnal 1000-way softmax layer. All layers had L2 weight constraints on

the incoming weights of each hidden unit. Using 50% dropout in the sixth hidden layer reduces

this to a record 42.4% (For details see Appendix E).

For the speech recognition dataset and both of the object recognition datasets it is necessary

to make a large number of decisions in designing the architecture of the net. We made these

decisions by holding out a separate validation set that was used to evaluate the performance of

a large number of different architectures and we then used the architecture that performed best

with dropout on the validation set to assess the performance of dropout on the real test set.

The Reuters dataset contains documents that have been labeled with a hierarchy of classes.

We created training and test sets each containing 201,369 documents from 50 mutually exclu-

sive classes. Each document was represented by a vector of counts for 2000 common non-stop

words, with each count Cbeing transformed to log(1 + C). A feedforward neural network with

2 fully connected layers of 2000 hidden units trained with backpropagation gets 31.05% error

on the test set. This is reduced to 29.62% by using 50% dropout (Appendix C).

We have tried various dropout probabilities and almost all of them improve the generaliza-

tion performance of the network. For fully connected layers, dropout in all hidden layers works

better than dropout in only one hidden layer and more extreme probabilities tend to be worse,

which is why we have used 0.5 throughout this paper. For the inputs, dropout can also help,

Fig. 4: Some Imagenet test cases with the probabilities of the best 5 labels underneath. Many

of the top 5 labels are quite plausible.

5

though it it often better to retain more than 50% of the inputs. It is also possible to adapt the

individual dropout probability of each hidden or input unit by comparing the average perfor-

mance on a validation set with the average performance when the unit is present. This makes

the method work slightly better. For datasets in which the required input-output mapping has a

number of fairly different regimes, performance can probably be further improved by making

the dropout probabilities be a learned function of the input, thus creating a statistically efﬁcient

“mixture of experts” (13) in which there are combinatorially many experts, but each parameter

gets adapted on a large fraction of the training data.

Dropout is considerably simpler to implement than Bayesian model averaging which weights

each model by its posterior probability given the training data. For complicated model classes,

like feedforward neural networks, Bayesian methods typically use a Markov chain Monte Carlo

method to sample models from the posterior distribution (14). By contrast, dropout with a

probability of 0.5assumes that all the models will eventually be given equal importance in the

combination but the learning of the shared weights takes this into account. At test time, the fact

that the dropout decisions are independent for each unit makes it very easy to approximate the

combined opinions of exponentially many dropout nets by using a single pass through the mean

net. This is far more efﬁcient than averaging the predictions of many separate models.

A popular alternative to Bayesian model averaging is “bagging” in which different models

are trained on different random selections of cases from the training set and all models are given

equal weight in the combination (15). Bagging is most often used with models such as decision

trees because these are very quick to ﬁt to data and very quick at test time (16). Dropout allows

a similar approach to be applied to feedforward neural networks which are much more powerful

models. Dropout can be seen as an extreme form of bagging in which each model is trained on

a single case and each parameter of the model is very strongly regularized by sharing it with

the corresponding parameter in all the other models. This is a much better regularizer than the

standard method of shrinking parameters towards zero.

A familiar and extreme case of dropout is “naive bayes” in which each input feature is

trained separately to predict the class label and then the predictive distributions of all the features

are multiplied together at test time. When there is very little training data, this often works much

better than logistic classiﬁcation which trains each input feature to work well in the context of

all the other features.

Finally, there is an intriguing similarity between dropout and a recent theory of the role of

sex in evolution (17). One possible interpretation of the theory of mixability articulated in (17)

is that sex breaks up sets of co-adapted genes and this means that achieving a function by using

a large set of co-adapted genes is not nearly as robust as achieving the same function, perhaps

less than optimally, in multiple alternative ways, each of which only uses a small number of

co-adapted genes. This allows evolution to avoid dead-ends in which improvements in ﬁtness

require co- ordinated changes to a large number of co-adapted genes. It also reduces the proba-

bility that small changes in the environment will cause large decreases in ﬁtness a phenomenon

which is known as “overﬁtting” in the ﬁeld of machine learning.

6

References and Notes

1. D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 323, 533 (1986).

2. G. E. Hinton, Neural Computation 14, 1771 (2002).

3. L. M. G. D. C. Ciresan, U. Meier, J. Schmidhuber, Neural Computation 22, 3207 (2010).

4. Y. B. Y. Lecun, L. Bottou, P. Haffner, Proceedings of the IEEE 86, 2278 (1998).

5. G. E. Hinton, R. Salakhutdinov, Science 313, 504 (2006).

6. A. Mohamed, G. Dahl, G. Hinton, IEEE Transactions on Audio, Speech, and Language

Processing, 20, 14 (2012).

7. G. Dahl, D. Yu, L. Deng, A. Acero, IEEE Transactions on Audio, Speech, and Language

Processing, 20, 30 (2012).

8. N. Jaitly, P. Nguyen, A. Senior, V. Vanhoucke, An Application OF Pretrained Deep Neu-

ral Networks To Large Vocabulary Conversational Speech Recognition, Tech. Rep. 001,

Department of Computer Science, University of Toronto (2012).

9. A. Krizhevsky, Learning multiple layers of features from tiny images, Tech. Rep. 001, De-

partment of Computer Science, University of Toronto (2009).

10. A. Coates, A. Y. Ng, ICML (2011), pp. 921–928.

11. J. Deng, et al.,CVPR09 (2009).

12. J. Sanchez, F. Perronnin, CVPR11 (2011).

13. S. J. N. R. A. Jacobs, M. I. Jordan, G. E. Hinton, Neural Computation 3, 79 (1991).

14. R. M. Neal, Bayesian Learning for Neural Networks, Lecture Notes in Statistics No. 118

(Springer-Verlag, New York, 1996).

15. L. Breiman, Machine Learning 24, 123 (1996).

16. L. Breiman, Machine Learning 45, 5 (2001).

17. J. D. A. Livnat, C. Papadimitriou, M. W. Feldman, PNAS 105, 19803 (2008).

18. R. R. Salakhutdinov, G. E. Hinton, Artiﬁcial Intelligence and Statistics (2009).

19. D. D. Lewis, T. G. R. Y. Yang, Journal of Machine Learning 5, 361 (2004).

20. We thank N. Jaitly for help with TIMIT, H. Larochelle, R. Neal, K. Swersky and C.K.I.

Williams for helpful discussions, and NSERC, Google and Microsoft Research for funding.

GEH and RRS are members of the Canadian Institute for Advanced Research.

7

A Experiments on MNIST

A.1 Details for dropout training

The MNIST dataset consists of 28 ×28 digit images - 60,000 for training and 10,000 for testing.

The objective is to classify the digit images into their correct digit class. We experimented with

neural nets of different architectures (different number of hidden units and layers) to evaluate

the sensitivity of the dropout method to these choices. We show results for 4 nets (784-800-

800-10, 784-1200-1200-10, 784-2000-2000-10, 784-1200-1200-1200-10). For each of these

architectures we use the same dropout rates - 50% dropout for all hidden units and 20% dropout

for visible units. We use stochastic gradient descent with 100-sized minibatches and a cross-

entropy objective function. An exponentially decaying learning rate is used that starts at the

value of 10.0 (applied to the average gradient in each minibatch). The learning rate is multiplied

by 0.998 after each epoch of training. The incoming weight vector corresponding to each hidden

unit is constrained to have a maximum squared length of l. If, as a result of an update, the

squared length exceeds l, the vector is scaled down so as to make it have a squared length

of l. Using cross validation we found that l= 15 gave best results. Weights are initialzed

to small random values drawn from a zero-mean normal distribution with standard deviation

0.01. Momentum is used to speed up learning. The momentum starts off at a value of 0.5 and

is increased linearly to 0.99 over the ﬁrst 500 epochs, after which it stays at 0.99. Also, the

learning rate is multiplied by a factor of (1-momentum). No weight decay is used. Weights

were updated at the end of each minibatch. Training was done for 3000 epochs. The weight

update takes the following form:

∆wt=pt∆wt−1−(1 −pt)th∇wLi

wt=wt−1+ ∆wt,

where,

t=0ft

pt=(t

Tpi+ (1 −t

T)pft < T

pft≥T

with 0= 10.0,f= 0.998,pi= 0.5,pf= 0.99,T= 500. While using a constant learning

rate also gives improvements over standard backpropagation, starting with a high learning rate

and decaying it provided a signiﬁcant boost in performance. Constraining input vectors to

have a ﬁxed length prevents weights from increasing arbitrarily in magnitude irrespective of the

learning rate. This gives the network a lot of opportunity to search for a good conﬁguration in

the weight space. As the learning rate decays, the algorithm is able to take smaller steps and

ﬁnds the right step size at which it can make learning progress. Using a high ﬁnal momentum

distributes gradient information over a large number of updates making learning stable in this

scenario where each gradient computation is for a different stochastic network.

8

A.2 Details for dropout ﬁnetuning

Apart from training a neural network starting from random weights, dropout can also be used

to ﬁnetune pretrained models. We found that ﬁnetuning a model using dropout with a small

learning rate can give much better performace than standard backpropagation ﬁnetuning.

Deep Belief Nets - We took a neural network pretrained using a Deep Belief Network (5).

It had a 784-500-500-2000 architecture and was trained using greedy layer-wise Contrastive

Divergence learning 1. Instead of ﬁne-tuning it with the usual backpropagation algorithm, we

used the dropout version of it. Dropout rate was same as before : 50% for hidden units and 20%

for visible units. A constant small learning rate of 1.0 was used. No constraint was imposed on

the length of incoming weight vectors. No weight decay was used. All other hyper-parameters

were set to be the same as before. The model was trained for 1000 epochs with stochstic gradient

descent using minibatches of size 100. While standard back propagation gave about 118 errors,

dropout decreased the errors to about 92.

Deep Boltzmann Machines - We also took a pretrained Deep Boltzmann Machine (18)2

(784-500-1000-10) and ﬁnetuned it using dropout-backpropagation. The model uses a 1784 -

500 - 1000 - 10 architecture (The extra 1000 input units come from the mean-ﬁeld activations

of the second layer of hidden units in the DBM, See (18) for details). All ﬁnetuning hyper-

parameters were set to be the same as the ones used for a Deep Belief Network. We were able

to get a mean of about 79 errors with dropout whereas usual ﬁnetuning gives about 94 errors.

A.3 Effect on features

One reason why dropout gives major improvements over backpropagation is that it encourages

each individual hidden unit to learn a useful feature without relying on speciﬁc other hidden

units to correct its mistakes. In order to verify this and better understand the effect of dropout on

feature learning, we look at the ﬁrst level of features learned by a 784-500-500 neural network

without any generative pre-training. The features are shown in Figure 5. Each panel shows 100

random features learned by each network. The features that dropout learns are simpler and look

like strokes, whereas the ones learned by standard backpropagation are difﬁcult to interpret.

This conﬁrms that dropout indeed forces the discriminative model to learn good features which

are less co-adapted and leads to better generalization.

B Experiments on TIMIT

The TIMIT Acoustic-Phonetic Continuous Speech Corpus is a standard dataset used for eval-

uation of automatic speech recognition systems. It consists of recordings of 630 speakers of 8

dialects of American English each reading 10 phonetically-rich sentences. It also comes with

1For code see http://www.cs.toronto.edu/˜hinton/MatlabForSciencePaper.html

2For code see http://www.utstat.toronto.edu/˜rsalakhu/DBM.html

9

(a) (b)

Fig. 5: Visualization of features learned by ﬁrst layer hidden units for (a) backprop and (b)

dropout on the MNIST dataset.

the word and phone-level transcriptions of the speech. The objective is to convert a given speech

signal into a transcription sequence of phones. This data needs to be pre-processed to extract

input features and output targets. We used Kaldi, an open source code library for speech 3, to

pre-process the dataset so that our results can be reproduced exactly. The inputs to our networks

are log ﬁlter bank responses. They are extracted for 25 ms speech windows with strides of 10

ms.

Each dimension of the input representation was normalized to have mean 0 and variance 1.

Minibatches of size 100 were used for both pretraining and dropout ﬁnetuning. We tried several

network architectures by varying the number of input frames (15 and 31), number of layers in

the neural network (3, 4 and 5) and the number of hidden units in each layer (2000 and 4000).

Figure 6 shows the validation error curves for a number of these combinations. Using dropout

consistently leads to lower error rates.

B.1 Pretraining

For all our experiments on TIMIT, we pretrain the neural network with a Deep Belief Net-

work (5). Since the inputs are real-valued, the ﬁrst layer was pre-trained as a Gaussian RBM.

3http://kaldi.sourceforge.net

10

(a)

(b)

Fig. 6: Frame classiﬁcation error and cross-entropy on the (a) Training and (b) Validation set

as learning progresses. The training error is computed using the stochastic nets.

11

Visible biases were initialized to zero and weights to random numbers sampled from a zero-

mean normal distribution with standard deviation 0.01. The variance of each visible unit was

set to 1.0 and not learned. Learning was done by minimizing Contrastive Divergence. Momen-

tum was used to speed up learning. Momentum started at 0.5 and was increased linearly to 0.9

over 20 epochs. A learning rate of 0.001 on the average gradient was used (which was then

multiplied by 1-momentum). An L2 weight decay of 0.001 was used. The model was trained

for 100 epochs.

All subsequent layers were trained as binary RBMs. A learning rate of 0.01 was used. The

visible bias of each unit was initialized to log(p/(1 −p)) where pwas the mean activation of

that unit in the dataset. All other hyper-parameters were set to be the same as those we used for

the Gaussian RBM. Each layer was trained for 50 epochs.

B.2 Dropout Finetuning

The pretrained RBMs were used to initialize the weights in a neural network. The network

was then ﬁnetuned with dropout-backpropagation. Momentum was increased from 0.5 to 0.9

linearly over 10 epochs. A small constant learning rate of 1.0 was used (applied to the average

gradient on a minibatch). All other hyperparameters are the same as for MNIST dropout ﬁne-

tuning. The model needs to be run for about 200 epochs to converge. The same network was

also ﬁnetuned with standard backpropagation using a smaller learning rate of 0.1, keeping all

other hyperparameters

Figure 6 shows the frame classiﬁcation error and cross-entropy objective value on the train-

ing and validation sets. We compare the performance of dropout and standard backpropaga-

tion on several network architectures and input representations. Dropout consistently achieves

lower error and cross-entropy. It signiﬁcantly controls overﬁtting, making the method robust

to choices of network architecture. It allows much larger nets to be trained and removes the

need for early stopping. We also observed that the ﬁnal error obtained by the model is not very

sensitive to the choice of learning rate and momentum.

C Experiments on Reuters

Reuters Corpus Volume I (RCV1-v2) (19) is an archive of 804,414 newswire stories that have

been manually categorized into 103 topics4. The corpus covers four major groups: corpo-

rate/industrial, economics, government/social, and markets. Sample topics include Energy Mar-

kets, Accounts/Earnings, Government Borrowings, Disasters and Accidents, Interbank Markets,

Legal/Judicial, Production/Services, etc. The topic classes form a tree which is typically of

depth three.

We took the dataset and split it into 63 classes based on the the 63 categories at the second-

level of the category tree. We removed 11 categories that did not have any data and one category

4The corpus is available at http://www.ai.mit.edu/projects/jmlr/papers/volume5/lewis04a/lyrl2004 rcv1v2 README.htm

12

that had only 4 training examples. We also removed one category that covered a huge chunk

(25%) of the examples. This left us with 50 classes and 402,738 documents. We divided the

documents into equal-sized training and test sets randomly. Each document was represented

using the 2000 most frequent non-stopwords in the dataset.

(a) (b)

Fig. 7: Classiﬁcation error rate on the (a) training and (b) validation sets of the Reuters dataset

as learning progresses. The training error is computed using the stochastic nets.

We trained a neural network using dropout-backpropagation and compared it with standard

backpropagation. We used a 2000-2000-1000-50 architecture. The training hyperparameters are

same as that in MNIST dropout training (Appendix A.1). Training was done for 500 epochs.

Figure 7 shows the training and test set errors as learning progresses. We show two nets

- one with a 2000-2000-1000-50 and another with a 2000-1000-1000-50 architecture trained

with and without dropout. As in all previous datasets discussed so far, we obtain signiﬁcant

improvements here too. The learning not only results in better generalization, but also proceeds

smoothly, without the need for early stopping.

D Tiny Images and CIFAR-10

The Tiny Images dataset contains 80 million 32 ×32 color images collected from the web. The

images were found by searching various image search engines for English nouns, so each image

comes with a very unreliable label, which is the noun that was used to ﬁnd it. The CIFAR-10

dataset is a subset of the Tiny Images dataset which contains 60000 images divided among ten

classes5. Each class contains 5000 training images and 1000 testing images. The classes are

5The CIFAR dataset is available at http://www.cs.toronto.edu/∼kriz/cifar.html.

13

airplane,automobile,bird,cat,deer,dog,frog,horse,ship, and truck. The CIFAR-10 dataset

was obtained by ﬁltering the Tiny Images dataset to remove images with incorrect labels. The

CIFAR-10 images are highly varied, and there is no canonical viewpoint or scale at which

the objects appear. The only criteria for including an image were that the image contain one

dominant instance of a CIFAR-10 class, and that the object in the image be easily identiﬁable

as belonging to the class indicated by the image label.

E ImageNet

ImageNet is a dataset of millions of labeled images in thousands of categories. The images

were collected from the web and labelled by human labellers using Amazon’s Mechanical Turk

crowd-sourcing tool. In 2010, a subset of roughly 1000 images in each of 1000 classes was the

basis of an object recognition competition, a part of the Pascal Visual Object Challenge. This

is the version of ImageNet on which we performed our experiments. In all, there are roughly

1.3 million training images, 50000 validation images, and 150000 testing images. This dataset

is similar in spirit to the CIFAR-10, but on a much bigger scale. The images are full-resolution,

and there are 1000 categories instead of ten. Another difference is that the ImageNet images

often contain multiple instances of ImageNet objects, simply due to the sheer number of object

classes. For this reason, even a human would have difﬁculty approaching perfect accuracy on

this dataset. For our experiments we resized all images to 256 ×256 pixels.

F Convolutional Neural Networks

Our models for CIFAR-10 and ImageNet are deep, feed-forward convolutional neural networks

(CNNs). Feed-forward neural networks are models which consist of several layers of “neurons”,

where each neuron in a given layer applies a linear ﬁlter to the outputs of the neurons in the

previous layer. Typically, a scalar bias is added to the ﬁlter output and a nonlinear activation

function is applied to the result before the neuron’s output is passed to the next layer. The linear

ﬁlters and biases are referred to as weights, and these are the parameters of the network that are

learned from the training data.

CNNs differ from ordinary neural networks in several ways. First, neurons in a CNN are

organized topographically into a bank that reﬂects the organization of dimensions in the input

data. So for images, the neurons are laid out on a 2D grid. Second, neurons in a CNN apply ﬁl-

ters which are local in extent and which are centered at the neuron’s location in the topographic

organization. This is reasonable for datasets where we expect the dependence of input dimen-

sions to be a decreasing function of distance, which is the case for pixels in natural images.

In particular, we expect that useful clues to the identity of the object in an input image can be

found by examining small local neighborhoods of the image. Third, all neurons in a bank apply

the same ﬁlter, but as just mentioned, they apply it at different locations in the input image. This

is reasonable for datasets with roughly stationary statistics, such as natural images. We expect

14

that the same kinds of structures can appear at all positions in an input image, so it is reasonable

to treat all positions equally by ﬁltering them in the same way. In this way, a bank of neurons

in a CNN applies a convolution operation to its input. A single layer in a CNN typically has

multiple banks of neurons, each performing a convolution with a different ﬁlter. These banks of

neurons become distinct input channels into the next layer. The distance, in pixels, between the

boundaries of the receptive ﬁelds of neighboring neurons in a convolutional bank determines

the stride with which the convolution operation is applied. Larger strides imply fewer neurons

per bank. Our models use a stride of one pixel unless otherwise noted.

One important consequence of this convolutional shared-ﬁlter architecture is a drastic re-

duction in the number of parameters relative to a neural net in which all neurons apply different

ﬁlters. This reduces the net’s representational capacity, but it also reduces its capacity to overﬁt,

so dropout is far less advantageous in convolutional layers.

F.1 Pooling

CNNs typically also feature “pooling” layers which summarize the activities of local patches

of neurons in convolutional layers. Essentially, a pooling layer takes as input the output of a

convolutional layer and subsamples it. A pooling layer consists of pooling units which are laid

out topographically and connected to a local neighborhood of convolutional unit outputs from

the same bank. Each pooling unit then computes some function of the bank’s output in that

neighborhood. Typical functions are maximum and average. Pooling layers with such units

are called max-pooling and average-pooling layers, respectively. The pooling units are usually

spaced at least several pixels apart, so that there are fewer total pooling units than there are

convolutional unit outputs in the previous layer. Making this spacing smaller than the size of

the neighborhood that the pooling units summarize produces overlapping pooling. This variant

makes the pooling layer produce a coarse coding of the convolutional unit outputs, which we

have found to aid generalization in our experiments. We refer to this spacing as the stride

between pooling units, analogously to the stride between convolutional units. Pooling layers

introduce a level of local translation invariance to the network, which improves generalization.

They are the analogues of complex cells in the mammalian visual cortex, which pool activities

of multiple simple cells. These cells are known to exhibit similar phase-invariance properties.

F.2 Local response normalization

Our networks also include response normalization layers. This type of layer encourages com-

petition for large activations among neurons belonging to different banks. In particular, the

activity ai

x,y of a neuron in bank iat position (x, y)in the topographic organization is divided

by

1 + α

i+N/2

X

j=i−N/2

(aj

x,y)2

β

15

where the sum runs over N“adjacent” banks of neurons at the same position in the topographic

organization. The ordering of the banks is of course arbitrary and determined before training

begins. Response normalization layers implement a form of lateral inhibition found in real

neurons. The constants N, α, and βare hyper-parameters whose values are determined using a

validation set.

F.3 Neuron nonlinearities

All of the neurons in our networks utilize the max-with-zero nonlinearity. That is, their output

is computed as ai

x,y = max(0, zi

x,y)where zi

x,y is the total input to the neuron (equivalently,

the output of the neuron’s linear ﬁlter added to the bias). This nonlinearity has several advan-

tages over traditional saturating neuron models, including a signiﬁcant reduction in the training

time required to reach a given error rate. This nonlinearity also reduces the need for contrast-

normalization and similar data pre-processing schemes, because neurons with this nonlinearity

do not saturate – their activities simply scale up when presented with unusually large input

values. Consequently, the only data pre-processing step which we take is to subtract the mean

activity from each pixel, so that the data is centered. So we train our networks on the (centered)

raw RGB values of the pixels.

F.4 Objective function

Our networks maximize the multinomial logistic regression objective, which is equivalent to

minimizing the average across training cases of the cross-entropy between the true label distri-

bution and the model’s predicted label distribution.

F.5 Weight initialization

We initialize the weights in our model from a zero-mean normal distribution with a variance set

high enough to produce positive inputs into the neurons in each layer. This is a slightly tricky

point when using the max-with-zero nonlinearity. If the input to a neuron is always negative,

no learning will take place because its output will be uniformly zero, as will the derivative

of its output with respect to its input. Therefore it’s important to initialize the weights from

a distribution with a sufﬁciently large variance such that all neurons are likely to get positive

inputs at least occasionally. In practice, we simply try different variances until we ﬁnd an

initialization that works. It usually only takes a few attempts. We also ﬁnd that initializing the

biases of the neurons in the hidden layers with some positive constant (1 in our case) helps get

learning off the ground, for the same reason.

16

F.6 Training

We train our models using stochastic gradient descent with a batch size of 128 examples and

momentum of 0.9. Therefore the update rule for weight wis

vi+1 = 0.9vi+ < ∂E

∂wi

>i

wi+1 =wi+vi+1

where iis the iteration index, vis a momentum variable, is the learning rate, and <∂E

∂wi>i

is the average over the ith batch of the derivative of the objective with respect to wi. We use

the publicly available cuda-convnet package to train all of our models on a single NVIDIA

GTX 580 GPU. Training on CIFAR-10 takes roughly 90 minutes. Training on ImageNet takes

roughly four days with dropout and two days without.

F.7 Learning rates

We use an equal learning rate for each layer, whose value we determine heuristically as the

largest power of ten that produces reductions in the objective function. In practice it is typically

of the order 10−2or 10−3. We reduce the learning rate twice by a factor of ten shortly before

terminating training.

G Models for CIFAR-10

Our model for CIFAR-10 without dropout is a CNN with three convolutional layers. Pooling

layers follow all three. All of the pooling layers summarize a 3×3neighborhood and use a stride

of 2. The pooling layer which follows the ﬁrst convolutional layer performs max-pooling, while

the remaining pooling layers perform average-pooling. Response normalization layers follow

the ﬁrst two pooling layers, with N= 9,α= 0.001, and β= 0.75. The upper-most pooling

layer is connected to a ten-unit softmax layer which outputs a probability distribution over class

labels. All convolutional layers have 64 ﬁlter banks and use a ﬁlter size of 5×5(times the

number of channels in the preceding layer).

Our model for CIFAR-10 with dropout is similar, but because dropout imposes a strong

regularization on the network, we are able to use more parameters. Therefore we add a fourth

weight layer, which takes its input from the third pooling layer. This weight layer is locally-

connected but not convolutional. It is like a convolutional layer in which ﬁlters in the same bank

do not share weights. This layer contains 16 banks of ﬁlters of size 3×3. This is the layer in

which we use 50% dropout. The softmax layer takes its input from this fourth weight layer.

17

H Models for ImageNet

Our model for ImageNet with dropout is a CNN which is trained on 224×224 patches randomly

extracted from the 256 ×256 images, as well as their horizontal reﬂections. This is a form of

data augmentation that reduces the network’s capacity to overﬁt the training data and helps

generalization. The network contains seven weight layers. The ﬁrst ﬁve are convolutional,

while the last two are globally-connected. Max-pooling layers follow the ﬁrst, second, and

ﬁfth convolutional layers. All of the pooling layers summarize a 3×3neighborhood and use a

stride of 2. Response-normalization layers follow the ﬁrst and second pooling layers. The ﬁrst

convolutional layer has 64 ﬁlter banks with 11 ×11 ﬁlters which it applies with a stride of 4

pixels (this is the distance between neighboring neurons in a bank). The second convolutional

layer has 256 ﬁlter banks with 5×5ﬁlters. This layer takes two inputs. The ﬁrst input to

this layer is the (pooled and response-normalized) output of the ﬁrst convolutional layer. The

256 banks in this layer are divided arbitrarily into groups of 64, and each group connects to a

unique random 16 channels from the ﬁrst convolutional layer. The second input to this layer

is a subsampled version of the original image (56 ×56), which is ﬁltered by this layer with a

stride of 2 pixels. The two maps resulting from ﬁltering the two inputs are summed element-

wise (they have exactly the same dimensions) and a max-with-zero nonlinearity is applied to

the sum in the usual way. The third, fourth, and ﬁfth convolutional layers are connected to

one another without any intervening pooling or normalization layers, but the max-with-zero

nonlinearity is applied at each layer after linear ﬁltering. The third convolutional layer has 512

ﬁlter banks divided into groups of 32, each group connecting to a unique random subset of 16

channels produced by the (pooled, normalized) outputs of the second convolutional layer. The

fourth and ﬁfth convolutional layers similarly have 512 ﬁlter banks divided into groups of 32,

each group connecting to a unique random subset of 32 channels produced by the layer below.

The next two weight layers are globally-connected, with 4096 neurons each. In these last two

layers we use 50% dropout. Finally, the output of the last globally-connected layer is fed to a

1000-way softmax which produces a distribution over the 1000 class labels. We test our model

by averaging the prediction of the net on ten 224 ×224 patches of the 256 ×256 input image:

the center patch, the four corner patches, and their horizontal reﬂections. Even though we make

ten passes of each image at test time, we are able to run our system in real-time.

Our model for ImageNet without dropout is similar, but without the two globally-connected

layers which create serious overﬁtting when used without dropout.

In order to achieve state-of-the-art performance on the validation set, we found it necessary

to use the very complicated network architecture described above. Fortunately, the complexity

of this architecture is not the main point of our paper. What we wanted to demonstrate is that

dropout is a signiﬁcant help even for the very complex neural nets that have been developed

by the joint efforts of many groups over many years to be really good at object recognition.

This is clearly demonstrated by the fact that using non-convolutional higher layers with a lot of

parameters leads to a big improvement with dropout but makes things worse without dropout.

18