Greedy layerwise training of convolutional neural networks

Abstract

Layerwise training presents an alternative approach to end-to-end back-propagation for training deep convolutional neural networks. Although previous work was unsuccessful in demonstrating the viability of layerwise training, especially on large-scale datasets such as ImageNet, recent work has shown that layerwise training on specific architectures can yield highly competitive performances. On ImageNet, the layerwise trained networks can perform comparably to many state-of-the-art end-to-end trained networks. In this thesis, we compare the performance gap between the two training procedures across a wide range of network architectures and further analyze the possible limitations of layerwise training. Our results show that layerwise training quickly saturates after a certain critical layer, due to the overfitting of early layers within the networks. We discuss several approaches we took to address this issue and help layerwise training improve across multiple architectures. From a fundamental standpoint, this study emphasizes the need to open the blackbox that is modern deep neural networks and investigate the layerwise interactions between intermediate hidden layers within deep networks, all through the lens of layerwise training.