Re-examining Class Selectivity in Deep Convolutional Networks

Akshay Badola, Vineet Padmanabhan, Rajendra Prasad Lal

Abstract: While Deep Neural Networks have become ubiquitous, some of their properties have remained elusive. Among these is class selectivity, which measures the association of a unit of a model to its inputs and outputs. Current literature is split on the effect of class selectivity of CNNs as some have determined it to be harmful for generalization, while others have found it to be beneficial. The results of such analyses can vary widely with the definition of selectivity used as there is no consensus on it. In this work, we provide a new flexible definition of class selectivity towards rectifying this discrepancy, which can better describe the network’s association with a class at various layers. We compare with the standard class selectivity metric and associated regularizer. We show experimentally that our proposed metric quantifies selectivity in a more consistent manner. We also dispel the notion that selectivity is harmful for generalization by showing that evaluation results do not change by increasing selectivity. We also analyze the association between selectivity, feature disentanglement, and decomposability and show that selectivity and filter disentanglement are complementary. Finally, we also provide the source code for all our experiments.

External IDs:dblp:conf/miwai/BadolaPL23