Exploring Robust Architectures for Deep Artificial Neural Networks

Abstract The architectures of deep artificial neural networks (DANNs) are routinelystudied to improve their predictive performance. However, the relationshipbetween the architecture of a DANN and its robustness to noiseand adversarial attacks is less explored. We investigate how the robustnessof DANNs relates to their underlying graph architectures or structures.This study: (1) starts by exploring the design space of architecturesof DANNs using graph-theoretic robustness measures; (2) transforms thegraphs to DANN architectures to train/validate/test on various imageclassification tasks; (3) explores the relationship between the robustnessof trained DANNs against noise and adversarial attacks and the robustnessof their underlying architectures estimated via graph-theoreticmeasures. We show that the topological entropy and Olivier-Ricci curvatureof the underlying graphs can quantify the robustness performanceof DANNs. The said relationship is stronger for complex tasks and largeDANNs. Our work will allow autoML and neural architecture searchcommunity to explore design spaces of robust and accurate DANNs.