Adaptive Coding Emerges in Stabilized Supralinear Networks Trained with Local Plasticity

Lateral connections (LCs) are ubiquitous in the cortical circuits. While DL architectures have rich intralayer interactions to support feature selectivity and contextual modulation, explicit excitatory and inhibitory (E-I) LCs remain underexplored and less-justified for encoding models in both DL and visual neuroscience. In this work, we analyze and train stabilized supralinear networks (SSNs) with strong E-I LCs, using local plasticity rules and natural images. We demonstrate that these LCs support a transition between dynamical regimes under different input conditions. During the transition, the network shifts from population coding that extracts features from low-contrast or noisy inputs by recruiting more neurons, to sparse coding at high contrast, utilizing considerably fewer neurons. This reduction in the number of active neurons has been generally associated with lower metabolic demand in previous experiments and models. We find the model showing better robustness and adaptiveness against sparse coding, ICA and other unsupervised models under degraded inputs, but not when LCs are ablated. These results support the role of E-I recurrence in dynamic coding strategies and the design of more adaptive and robust systems with a concrete example in vision.

AI vision often fail when images are faint, noisy, or otherwise degraded, while our vision are more robust. We asked whether a specific circuit motif in the brain—local excitatory and inhibitory connections between neurons—could help explain this difference. We studied a brain-inspired recurrent model called a stabilized supralinear network and trained it on natural images using only local learning rules, rather than hand-programming a coding strategy. After training, the model learned visual features and automatically changes how it represents images as input quality changes. For weak or noisy inputs, it used a “many-neuron” population code, pooling evidence across more neurons to see the feature better. For clear, high-contrast inputs, it shifted to a sparse code, relying on fewer highly selective neurons, a pattern associated with lower metabolic demand in biological models. With this adaptive coding, the model show improved robustness under low contrast and noise compared with various other unsupervised learning models in larger experiments, but could not do so without the excitatory-inhibitory recurrence. The results suggest that this recurrence is not just a biological detail, but a mechanism for adaptively balancing robust perception with efficient neural activity.