Gain-mediated statistical adaptation in recurrent neural networks

L Duong, C Bredenberg, D J Heeger and E P Simoncelli

Published in Computational and Systems Neuroscience (CoSyNe), Mar 2022.
Sensory neurons in different species, brain areas, and modalities adjust their sensitivity (gain) in response to recent stimulus history. This gain control offers a mechanism for single neurons to rapidly and reversibly adapt to different stimulus contexts while preserving synaptic weights that serve to represent features that remain consistent across contexts. From a normative perspective, this allows an individual neuron to adjust the dynamic range of its responses to accommodate changes in input statistics -- a core tenet of Barlow's efficient coding theory. However, experimental measurements reveal that adaptation induces more complex changes in neural responses: tuning-dependent reductions in both response maxima and minima; tuning curve repulsion; and stimulus-driven decorrelation. Although coding efficiency and gain-mediated adaptation is well-studied in single neurons, this combination of normative principle and simple mechanism appears insufficient in accounting for these nuanced effects. Indeed, to explain these phenomena, previous studies have relied on adaptive changes in feedforward or recurrent synaptic efficacy [1, 2], mechanisms which, while more flexible, are unlikely to operate as transiently as simple gain modulation.

We propose a normative framework for adaptive gain control in which recurrently-connected neurons dynamically adjust their gains in response to novel stimulus statistics. Specifically, gains are modulated to optimize an objective enforcing an accurate representation of their inputs while minimizing total population activity. We compare model predictions to experimental measurements of V1 neurons responding to a sequence of gratings drawn from an ensemble with either uniform or biased orientation probability [3]. The model captures the full set of adaptation phenomena in the data simply by propagating the effects of single-neuron gain changes through the network. Thus, single-neuron gain control within a recurrent circuit, coupled with a coding efficiency objective, is sufficient to capture the observed diversity of neural adaptation responses.

[1] M. J. Wainwright, O. Schwartz, and E. P. Simoncelli. ``Natural Image Statistics and Divisive Normalization: Modeling Nonlinearities and Adaptation in Cortical Neurons''. In: Probabilistic Models of the Brain: Perception and Neural Function. Ed. by R Rao, B Olshausen, and M Lewicki. MIT Press, Feb. 2002. Chap. 10, pp. 203-222.
[2] Z. M. Westrick, D. J. Heeger, and M. S. Landy. ``Pattern Adaptation and Normalization Reweighting''. In: Journal of Neuroscience 36.38 (), pp. 9805-9816.
[3] A. Benucci, A.B. Saleem, and M. Carandini. ``Adaptation maintains population homeostasis in primary visual cortex''. In: Nature Neuroscience 16.6 (2013), pp. 724-729.

  • Related Publications: Wainwright00c
  • Listing of all publications