Item

Abstract

Cortical dynamics unfold across multiple timescales reflecting networks’ specialization for task-relevant computations [1,2]. However, it is unknown how these timescales emerge from the network connectivity and whether they can be flexibly modulated by cognitive demands, e.g., during attention. We analyzed the timescales in autocorrelations of population spiking activity recorded from single cortical columns in area V4 from monkeys performing a spatial attention task (AT) and a fixation task (FT) (fig 1A). We observed that both spontaneous (FT) and stimulus-driven (AT) activity exhibit two distinct timescales (one slow and one fast). To validate the presence of two timescales and estimate their values, we developed a method based on Approximate Bayesian Computations (ABC) [3]. Our method estimates the timescales from spiking activity by fitting autocorrelations using a generative model with multiple timescales to overcome statistical biases due to finite sample size [4] (fig 1B,C). We found that most recordings (31 out of 37, 84%) were better fitted with two timescales than one timescale. Moreover, the slow timescale was significantly longer on trials when monkeys attended to the receptive fields (RFs) location of the recorded neurons than on control trials when monkeys attended to a different location (fig 1C,D).
We hypothesized that the observed timescales emerge from the recurrent network dynamics shaped by the spatial connectivity structure. We developed a network model consisting of binary units representing cortical minicolumns [5] with local spatial connectivity among them (fig 1E). We found that the activity of model minicolumns exhibits two distinct timescales: A fast timescale induced by vertical recurrent excitation within a minicolumn, and a slow timescale induced by horizontal interactions among minicolumns. Both timescales depend on the network topology, and the slow timescale disappears in networks with random connectivity (fig 1F). We derived an analytical relationship between the timescales and connectivity parameters, enabling us to identify model parameters best matching the timescales in the data. The model indicates that modulation of timescales during attention arises from a slight increase in the efficacy of horizontal interactions (fig 1G). Our results suggest that timescales of local neural dynamics emerge from the spatial network structure and can flexibly change due to top-down influences according to task demands.