ReNet
lastmod 2016-03-10created 2016-03-101 edits
ReNet: A Recurrent Neural Network Based Alternative to Convolutional Networks
Abstract
In this paper, we propose a deep neural network architecture for object recognition based on recurrent neural networks. The proposed network, called ReNet, replaces the ubiquitous convolution+pooling layer of the deep convolutional neural network with four recurrent neural networks that sweep horizontally and vertically in both directions across the image. We evaluate the proposed ReNet on three widely-used benchmark datasets; MNIST, CIFAR-10 and SVHN. The result suggests that ReNet is a viable alternative to the deep convolutional neural network, and that further investigation is needed.
Summary
ReNet replaces the existing convolution and pooling layers with RNNs so that the image is scanned horizontally and vertically. These RNN layers are then stacked. The authors showed that the ReNet architecture can be a viable alternative to CNNs (convolutional neural nets). In this paper ReNet does not outperform CNNs, but performance is expected to improve with further optimization and hyperparameter tuning.
- The image is split into patches, and one patch is fed into the RNN at each time step. Each layer has 4 RNNs, one for each of the four horizontal and vertical directions.
- Because the RNNs in ReNet look at the entire image, ReNet can see the global context. Convolution/pooling layers, by contrast, can only see local context.
- The experiments used two 256-dimensional ReNet layers, 2x2 patches, and a 4096-dimensional affine layer.
- Data was augmented by adding flipped and shifted images.