Visual response inhibition for increased robustness of convolutional networks to distribution shifts

Convolutional neural networks have been shown to suffer from distribution shiftsin the test data, for instance caused by the so called common corruptions andperturbations. Test images can contain noise, digital transformations, and blurthat were not present in the training data, negatively impacting the performance oftrained models. Humans experience much stronger robustness to noise and visualdistortions than deep networks. In this work, we explore the effectiveness of aneuronal response inhibition mechanism, called push-pull, observed in the earlypart of the visual system, to increase the robustness of deep convolutional networks.We deploy a Push-Pull inhibition layer as a replacement of the initial convolutionallayers (input layer and in the first block of residual and dense architectures) ofstandard convolutional networks for image classification. We show that the PushPull inhibition component increases the robustness of standard networks for imageclassification to distribution shifts on the CIFAR10-C and CIFAR10-P test sets