Parametric Surround modulation improves the robustness of the deep neural networks

Abstract

In this study, we propose a bio-inspired deep neural network that is robust to classification tasks in an unstructured environment. Deep neural networks show high classification accuracy when training and testing data are independent and identically distributed. However, deep neural networks show significantly low classification accuracy on the corrupted data. In this study, we propose Parametric Surround Modulation for robust image classification. Surround modulation is an important biological mechanism of mammalian visual systems. It detects object boundaries, controls contrast gain, and perceives semantic features invariant to the environmental changes. Parametric Surround Modulation applies the surround modulation function to deep neural networks with trainable parameters. Parametric Surround Modulation is trained on ImageNet and the performance of the module is evaluated on the corrupted dataset, ImageNet-C. Experimental results show that the proposed Para- metric Surround Modulation accurately performs image classification for the corrupted data.