Adaptive coding efficiency through joint gain control in neural populations

Efficiently transmitting information from dynamic environments necessitates sensory systems that rapidly and reversibly adapt to changes in input statistics. Single neurons adjust their input-output gains to adaptively normalize their stimulus-driven response variance (Fairhall et al. 2001). In neural populations, statistical whitening and related adaptations have been observed (Dan et al. 1996; Benucci et al. 2013; Wanner and Friedrich 2020), whereby joint statistics of neurons are decorrelated in addition to being re-scaled to have equal variance. Existing models of neural population adaptation rely on synaptic plasticity mechanisms, which, while more flexible, are unlikely to operate as transiently or reversibly as gain control. In this study, we develop a novel circuit which generalizes single-neuron adaptive gain control to the level of a population. We derive a normative adaptive whitening algorithm which regulates joint second-order statistics of a neural population by adjusting the marginal statistics of an overcomplete auxiliary population. The algorithm operates online, and can be mapped onto a recurrent neural network comprising principal cells and gain-modulating interneurons. Remarkably, the interneurons adjust gains using only local signals, and feed back onto principal cells to achieve a globally statistically white output. Our framework can be generalized to handle biophysical constraints, and we demonstrate its use in statistically whitening local image patches using convolutional weights. Finally, we compare our model to experimental observations of adaptation in early sensory systems.