Adaptive coding efficiency with fast gain modulation and slow synaptic plasticity
D Lipshutz, L R. Duong, D B Chklovskii and E P Simoncelli
Published in Computational and Systems Neuroscience (CoSyNe), Mar 2024.
Efficient transmission of information from dynamic environments necessitates sensory systems that rapidly adapt to changes in sensory statistics. Neurons in early sensory areas rapidly adapt to changes in sensory statistics to both normalize their response variances (Fairhall, Lewen & Bialek, 2001) and reduce between-neuron response correlations (Benucci, Saleem & Carandini, 2013; Wanner & Friedrich, 2020), which together can be viewed as a form of adaptive whitening. While it is well-established that single neurons can modulate their gains to normalize their response variance, the mechanism underlying adaptive decorrelation of neural populations remains unclear. Existing mechanistic models of adaptive whitening exclusively use either synaptic plasticity or gain modulation as the biological substrate for decorrelation; however, on their own, each of these models has significant limitations.
We unify these approaches in a multi-timescale mechanistic model that adaptively whitens its responses using a combination of fast gain modulation and slow synaptic plasticity. Our model is derived from a novel whitening objective that factorizes the whitening transformation into context-independent basis vectors, which correspond to synaptic weights, and a context-dependent diagonal matrix, which corresponds to neuronal gains. We test our model on synthetic and natural datasets and find that the synapses learn to match shared structure of the stimulus inputs over long timescales that enable adaptive whitening on short timescales using gain modulation.