The effect of spatial RNNs on neural network feature maps

Abstract

Study of LSTMs for feature correlation in order to improve generalization of networks when training data is not abundant.

Computer vision applications and specifically image classification tasks usually rely on convolutional layers in order to extract information form input images and process the feature maps. In this thesis we experiment and study the effects of applying sequence recurrent neural networks (RNN) to spatial feature maps. A new approach introduced by ReNet, Inside-Outside Network and PoseNet LSTM where sequence RNN are used to process 2D feature maps and improve the performance of the network. In this thesis we evaluate different toy models in the MNIST and CIFAR10 datasets to observe which are the best practices when applying the RNN and discover in which way can they improve the performance.