FP-nets as novel deep networks inspired by vision

Feature-product networks (FP-nets) are inspired by end-stopped cortical cells with FP-units that multiply theoutputs of two filters.We enhance state-of-the-art deepnetworks, such as the ResNet and MobileNet, withFP-units and show that the resulting FP-nets performbetter on the Cifar-10 and ImageNet benchmarks.Moreover, we analyze the hyperselectivity of the FP-netmodel neurons and show that this property makesFP-nets less sensitive to adversarial attacks and JPEGartifacts.We then show that the learned model neuronsare end-stopped to different degrees and that theyprovide sparse representations with an entropy thatdecreases with hyperselectivity