Robustness of Biologically-Inspired Filter-Based ConvNet to Signal Perturbation

We have studied the effectiveness of a biologically-inspired filter in improving the performance of the VGG-16 ConvNet when presented with images perturbed by noise and distortion. Our work builds on the findings of Evans et al. (2021), who reported that biologically-inspired Gabor filter-based VGG-16 improved tolerance to perturbations. Previously, we have demonstrated that foveal filters perform better than Difference-of-Gaussian (DoG) filters, both of which are inspired by the primate retina, in terms of perceptual content retrieval from input images. Subsequently, we have observed that using foveal kernels improved the accuracy of a spiking ConvNet. In this work, we introduce the foveal filter-based VGG-16; our goal is to compare its robustness with other biologically-inspired filters, viz. DoG and Gabor, when presented with perturbed images. Our results showed that when tested with perturbed images, the foveal filter-based VGG-16 outperformed the standard, Gabor filter-based, and the DoG filter-based VGG-16. However, when tested with unperturbed images, the performances of all the biologically-inspired filter-based VGG-16 models were similar to, but not better than, those of the standard VGG-16. This implies that the biologically-inspired filters are particularly robust to noisy, distorted inputs; of these, the foveal filters are seen to be the most robust. To further test the foveal filters, we added perturbation to the training dataset and compared the standard and foveal filter-based VGG-16. The foveal filter-based VGG-16 once again outperformed the standard VGG16, thus affirming its robustness to perturbation.