State-dependent Filtering as a Mechanism towards Visual Robustness

Jing Yan¹Yunxuan Feng¹Wei P Dai^2*Yaoyu Zhang^1*

• ¹Shanghai Jiao Tong University, Shanghai, China
• ²Fudan University, Shanghai, China

Robustness, defined as a system's ability to maintain functional reliability in the face of perturbations, is achieved through the system's capacity to filter out external disturbances using internal priors encoded in its structure and states. While biophysical neural networks are widely recognized for their robustness, the precise mechanisms underlying this resilience remain poorly understood. In this study, we explore how orientation-selective neurons arranged in a one-dimensional ring network respond to perturbations with the aim of uncovering insights into the robustness of visual subsystems in the brain. Through analysis of the steady-state dynamics of a rate-based network, we demonstrate that the activation state of neurons, rather than their firing rates, along with lateral connectivity, governs the network's response to perturbations. We identify specific sinusoidal-like perturbation patterns that elicit maximal responses for different activation state configurations, while other patterns are significantly attenuated. These findings are corroborated in a spiking ring model. Furthermore, we map the perturbations in orientation back into two-dimensional image space using Gabor functions. We determine the optimal perturbation patterns for Gabor inputs and show that lateral connectivity profiles, motivated by biological observations and widely used in the theoretical models, prioritize input elongation over contrast enhancement while suppressing irrelevant perturbations. Through our detailed analysis of the ring model, this work elucidates how cortical circuits could amplify relevant perturbations to an input while suppressing irrelevant ones, highlighting the pivotal role of lateral connectivity in facilitating robust visual processing.