Predicting neuronal dynamics with a delayed gain control model

Abstract

Visual neurons respond to static images with specific dynamics: neuronal responses sum sub-additively over time, reduce in amplitude with repeated or sustained stimuli (neuronal adaptation), and are slower at low stimulus contrast. Here, we propose a simple model that predicts these seemingly disparate response patterns observed in a diverse set of measurements–intracranial electrodes in patients, fMRI, and macaque single unit spiking. The model takes a time-varying contrast time course of a stimulus as input, and produces predicted neuronal dynamics as output. Model computation consists of linear filtering, expansive exponentiation, and a divisive gain control. The gain control signal relates to but is slower than the linear signal, and this delay is critical in giving rise to predictions matched to the observed dynamics. Our model is simpler than previously proposed related models, and fitting the model to intracranial EEG data uncovers two regularities across human visual field maps: estimated linear filters (temporal receptive fields) systematically differ across and within visual field maps, and later areas exhibit more rapid and substantial gain control. The model is further generalizable to account for dynamics of contrast-dependent spike rates in macaque V1, and amplitudes of fMRI BOLD in human V1.

Interaction between steady-state visually evoked potentials at nearby flicker frequencies

“…Temporal domain models have been used previously in the masking literature, but most of these studies relied on multiple contrast levels for the target or the mask for evaluating between candidate models [15][16][17] , something our study lacks. A recent study 29 has used a delayed gain control model to explain the temporal structure of neural responses in various modalities of recording. Although they do not focus on temporal frequency steady-state responses, it seems conceivable that the impulse response function in their model formulation would relate to the observed temporal frequency tuning (Fig.…”

“…The free parameters are the 7 amplitudes L amp for each TF, time constants τ 0 and τ 1 , and exponent n. The amplitude parameters L amp characterize the TF tuning of the electrode at maximum contrast, and time constants τ 0 and τ 1 encapsulate the responsiveness of the neuronal population. We cannot explicitly attach any biophysical interpretation to these time constants or directly compare them to time constants obtained in other models of temporal response and gain control 16,29,73,74 , except noting that they may be thought of as "aggregate" measures of conductance at rest and at maximum contrast 23 . Finally, the exponent n captures any nonlinearities in the TF response generation process.…”

“…Unsurprisingly, masking studies have also used variants of the normalization model to explain SSVEP suppression [15][16][17] . Normalization models in the neurophysiology literature commonly explain the aggregate response of a cell or neural population in a given time window 23,25,27 , but a few recent studies have also explored more dynamic formulations 28,29 . Masking studies have usually employed temporal domain models, which can also account for IM responses across different stimulus conditions [15][16][17] .…”

Diverse Temporal Dynamics of Repetition Suppression Revealed by Intracranial Recordings in the Human Ventral Temporal Cortex

“…shown that modeling the temporal responses of neurons in millisecond resolution better predicts BOLD responses (with a resolution of seconds) than the general linear model (Stigliani A et al 2017;Zhou J et al 2018;Stigliani A et al 2019;Zhou J et al 2019), future research can measure both ECoG and fMRI responses in an event-related design for long-lagged repetitions to directly relate the impact of the repetition on the combined changes to PT, PM, and AUC, measured with ECoG, on BOLD responses.…”

Effect of Cross-Orientation Normalization on Different Neural Measures in Macaque Primary Visual Cortex

“…Although our model was able to explain the populationlevel neural measures, it is descriptive in nature and does not provide a comprehensive biophysical explanation of why the normalization strength is different for different measures. It also lacks dynamics (fluctuations in normalization strength), which has been introduced in recent normalization models ( Tsai et al 2012 ; Zhou et al 2019 ) and cannot explain differences in crossorientation normalization produced by overlapping gratings of arbitrary relative orientations, as discussed and implemented in other normalization models ( Candy et al 2001 ; Hermes et al 2019 ). Our analysis was limited to variants of the standard normalization model, because these were directly comparable to a large body of literature that have used similar simplistic models to explain spike responses ( Ni et al 2012 ; Ruff et al 2016 ; Ni and Maunsell 2017 ).…”