Modeling the Diverse Effects of Divisive Normalization on Noise Correlations

Abstract

Divisive normalization, a prominent model of interactions between neurons, plays a key role in influential theories of neural coding. Testing the mechanistic basis of those theories requires understanding the relationship between normalization and the statistics of neural activity, particularly noise correlations which provide strong constrains on circuit models. Yet the relation between normalization and correlations remains unclear, because models of covariability typically ignore normalization despite empirical evidence suggesting it contributes to stimulus and state modulations of covariability. We propose a pairwise stochastic divisive normalization model that accounts for the effects of normalization and other factors on covariability, and accurately fits Ca2+ imaging data from mouse primary visual cortex (V1). We show that normalization modulates noise correlations in qualitatively different ways depending on whether normalization is shared between neurons, and we outline the data regime in which the corresponding model parameter is identifiable. Our analysis provides the first quantitative evidence that similarly tuned neurons in mouse V1 share their normalization signals, and explains our observation that stronger normalization decorrelates those pairs, particularly at low stimulus contrast. Our model could enable quantifying the role of normalization on information transmission and representation, and mapping the parameters of normalization onto circuit and cellular mechanisms.

Coordinated multiplexing of information about separate objects in visual cortex

“…For example, several important recent studies have begun to explore how recurrent circuit mechanisms might implement dynamic fluctuations in neural activity ( Heeger and Mackey, 2019 ; Heeger and Zemlianova, 2020 ) and have postulated that shifts in the balanced excitation and inhibition that is thought to underlie normalized average responses when two stimuli are presented might contribute to sizeable positive or negative spike count correlations ( Verhoef and Maunsell, 2017 ). Finally, recent work by Coen-Cagli and colleagues proposes a method of assessing normalization strength on individual trials and demonstrated a connection between the neural responses that are well-described under a normalization model and the level of variability of firing that they show ( Coen-Cagli and Solomon, 2019 ; Weiss et al, 2022 ).…”

Response sub-additivity and variability quenching in visual cortex

“…A natural and important extension of this work is to additionally consider pairwise and population-level response statistics (such as pairwise noise correlations and the geometry of population activity). These statistics have been studied extensively in cortical areas [177][178][179] , are influenced by similar factors as those that elicit response sub-additivity and variability quenching 38,51,[180][181][182][183][184] and further constrain models of neural activity. Response sub-additivity and variability quenching co-occur under various experimental manipulations.…”

“…Note that we primarily discuss data collected in cat and monkey. Recent work in rodent V1 has revealed similar sub-additive phenomena (and is rapidly advancing our understanding of the underlying circuit mechanisms [42][43][44][45][46][47][48] ), but studies of response variability in rodent V1 have thus far been less extensive than in these other species (but see [49][50][51] ).…”