A Push-Pull Layer Improves Robustness of Convolutional Neural Networks

TL;DR

This paper introduces a biologically-inspired push-pull layer for CNNs that enhances robustness against noise perturbations without sacrificing accuracy on clean images.

Contribution

The paper presents a novel push-pull layer based on visual system neurons, improving CNN robustness to noise while maintaining high performance.

Findings

Significant robustness improvement on noisy images.

Maintains state-of-the-art accuracy on clean images.

Applicable to different CNN architectures.

Abstract

We propose a new layer in Convolutional Neural Networks (CNNs) to increase their robustness to several types of noise perturbations of the input images. We call this a push-pull layer and compute its response as the combination of two half-wave rectified convolutions, with kernels of opposite polarity. It is based on a biologically-motivated non-linear model of certain neurons in the visual system that exhibit a response suppression phenomenon, known as push-pull inhibition. We validate our method by substituting the first convolutional layer of the LeNet-5 and WideResNet architectures with our push-pull layer. We train the networks on nonperturbed training images from the MNIST, CIFAR-10 and CIFAR-100 data sets, and test on images perturbed by noise that is unseen by the training process. We demonstrate that our push-pull layers contribute to a considerable improvement in robustness of classification of images perturbed by noise, while maintaining state-of-the-art performance on the original image classification task.