I decided to read interesting deep learning papers often and try to summarize them to aid my own understanding of the topics.

This paper, ImageNet Classification with Deep Convolutional Neural Networks demonstrates a record breaking result on the ImageNet LSVRC-2012 competition. The authors participated in the competition under the name SuperVision (which is extremently difficult to Google, especially in the “deep learning” context, which makes Google surface supervised learning related results).

## Data

They trained the model on the ImageNet dataset, which contains about 1.2 million images of 1000 categories.

### Image preprocessing

Since ImageNet images are variable resolution, and the model presented in this paper requires fixed size images, they scaled every image to 256x256 pixels. The scaling like so:

1. Scale a possibly rectangular image so that the shorter side is 256 pixels.
2. Take the middle 256x256 patch as the input image.

### Data augmentation

Actually, the network presented in the paper works with 224x224 images, which are generated by randomly sampling patches of that size from each 256x256 image. This is one of the ways they did data augmentation.

The other way they augmented the dataset involved perturbing the R,G,B values of each input image by a scaled version of the principal components (in RGB space) across the whole training set. If $\mathbf{p_1}$, $\mathbf{p_2}$ and $\mathbf{p_3}$ be the eigenvectors and $\lambda_1$, $\lambda_2$, $\lambda_3$ the corresponding eigenvalues, each pixel in a given image is perturbed such that:

where each $\alpha_1$ is picked from a Gaussian with 0 mean and 0.1 std every time an image is used for training (but the values are shared by all pixels in the image once picked). They claim this captured an invariance of object identity under changes in intensity and color of illumination, which seems to have accounted for a ~ 1% decrease in top-1 error rate.

## Model architecture

The model has 5 convolutional layers and 3 fully-connected layers.

### Choice of nonlinearity

Authors chose the ReLU function (ReLU(x) = max(0,x)) and were I think one of the first to endorse it, and for good reasons. They saw that using the ReLU activation function, as opposed to the conventional tanh or sigmoid, yielded tremendous gains in training speed, owing to the non saturating nature of the function. When pitted against the tanh activation with no other changes, they were able to train their model to 25% error rate on the training set 6 times faster with the ReLU activation.

### Pooling

They use max pooling but with overlapping windows (filter size of 3 and stride of 2). Max pooling is applied after response renormalization, described next.

### Local response renormalization

Activations from a given filter/kernel for a given location $\left(x,y\right)$ are scaled by a term involving the activations on $n$ adjacent kernels for the same location. $n$ and other terms in question are hyperparameters and described very well in the corresponding section in the paper. The inspiration seems to be biological in nature, and in effect causes filters in an $n$ group to “compete” for having larger activations. Of course they probably wouldn’t have kept it had it not given them a significant error rate reduction on their validation set (which it did).

### Dropout, momentum and weight decay

Dropout is used in the first two fully-connected layers (layers 6 and 7 in the overall network). Essentially, a given neuron is switched-off in a forward-backward cycle with probability 0.5. This means that the model is forced to learn redundant representations with shared weights. During testing, they do not switch of neurons, but multiply their outputs by 0.5. No one quite knows exactly why dropout works well, but indeed it has worked extremely well for the authors and for many other models that followed. It has become very commonplace today to incorporate dropout in FC layers to reduce overfitting. Dropout approximately doubled the training time.

Momentum is a technique to speed up training by pushing batch gradient descent updates to move towards the optimum faster. This is especially important when the objective function is like a long ravine leading to the optimum with steep wells (or like a hammock with a heavy person inside, but in higher dimensions).

In such an objective, plain batch gradient descent, if starting on the left side, will oscillate up and down, moving slowly towards the optimum. Momentum uses exponentially weighted averages to take into account previous gradient descent updates. This averaging has an effect of cancelling out the up and down oscillations, thus speeding up the learning significantly. Authors use a momentum parameter of $0.9$, which roughly keeps the past 10 values in the moving average memory.

Weight decay is simply to scale down weights by a factor slightly less than 1 to prevent them from becoming too big. This can be seen as a regularizing technique. In all the gradient descent updates for the model weights look like the following:

## Training

The training is done on two GPUs for parallelism, and the setup is quite interesting. The GPUs used each have 3GB memory. The network is split into halves, as can be seen in the model description figure, across the two GPUs. The interesting bit is that while layer 4 (CONV) operates on input which comes from the layer 3 activations from both GPUs, other conv layers in the network do not have this cross GPU communication going on, and only work with activations from the GPU local half of the network.

Selecting how many layers should have cross-GPU comms going on is a problem for cross-validation, and apparently this scheme seems to have worked well for them. In the following figure, we see the first conv layer weights visualized, the top half being the first GPU, and the bottom half the other. As the authors mention in the paper, the two halves of the network consistently specialized in features, with one focusing on frequency/orientation, and the other in coloured blob detection.

Because of the sheer number of parameters (60 million), the training still takes 5-6 days. The learning rate had to be manually decayed when progress plateaued.

## Conclusion

This work was the first of its kind to have trained deep convolutional networks on GPUs to achieve impressive results on the ImageNet dataset for object detection. To conclude, here’s a Google Trends visualization for the term “deep learning” over time:

(the paper was presented in December 2012 at NIPS).