Structured Neural Noise Boosts Artificial Network Robustness

This study explores the role of neural variability, specifically correlated noise, in enhancing the robustness of artificial neural networks (ANNs). Drawing inspiration from biological neural systems where trial-to-trial variability is common, the research investigates whether introducing structured noise into ANN activations can improve their resilience to various perturbations. The findings indicate that structured noise indeed significantly boosts network robustness, particularly against adversarial attacks and naturalistic image modifications. Interestingly, the benefits of noise structure for naturalistic modifications are highly specific to the type of modification. In contrast, noise structure learned from adversarial attacks demonstrates a valuable ability to generalize across different kinds of attacks. These results establish a biologically plausible strategy for developing more robust ANNs. By leveraging local information to induce structured noise in activations, it's possible to create networks that are more resilient to unforeseen inputs and malicious manipulations, without requiring global architectural changes.