All You Need is Beyond a Good Init: Exploring Better Solution for Training Extremely Deep Convolutional Neural Networks with Orthonormality and Modulation

TL;DR

This paper introduces orthonormality regularization and error modulation techniques to effectively train extremely deep plain CNNs without shortcuts, achieving performance comparable to residual networks.

Contribution

It proposes novel optimization methods leveraging orthonormality and error modulation to train very deep plain CNNs from scratch, eliminating the need for residual connections.

Findings

Improved training of 44-layer and 110-layer plain CNNs on CIFAR-10 and ImageNet.

Plain CNNs trained with these methods match residual network performance.

New principles for network design based on orthonormality insights.

Abstract

Deep neural network is difficult to train and this predicament becomes worse as the depth increases. The essence of this problem exists in the magnitude of backpropagated errors that will result in gradient vanishing or exploding phenomenon. We show that a variant of regularizer which utilizes orthonormality among different filter banks can alleviate this problem. Moreover, we design a backward error modulation mechanism based on the quasi-isometry assumption between two consecutive parametric layers. Equipped with these two ingredients, we propose several novel optimization solutions that can be utilized for training a specific-structured (repetitively triple modules of Conv-BNReLU) extremely deep convolutional neural network (CNN) WITHOUT any shortcuts/ identity mappings from scratch. Experiments show that our proposed solutions can achieve distinct improvements for a 44-layer and a 110-layer plain networks on both the CIFAR-10 and ImageNet datasets. Moreover, we can successfully train plain CNNs to match the performance of the residual counterparts. Besides, we propose new principles for designing network structure from the insights evoked by orthonormality. Combined with residual structure, we achieve comparative performance on the ImageNet dataset.