Gain signal manifest in prestimulus neural population dynamics underlies decision-making
Embargo Date
2022-07-26
OA Version
Citation
Abstract
Decision-making is thought to be shaped by factors such as “urgency” and past outcomes. However, the effect of such factors on neural population dynamics and its link to decision-making behavior is largely unclear. Here, we addressed this gap by investigating the neuronal population dynamics in a heterogeneous population of dorsal premotor cortex (PMd) neurons recorded from monkeys performing a red-green checkerboard discrimination task.
We investigated effects of urgency by analyzing firing rates of neurons organized by reaction time (RT) and choice. Dimensionality reduction, regression, and decoding analyses suggested that prestimulus neural population state covaried with RT but not choice. Critically, effects were observed within a stimulus difficulty. Subsequent analysis suggested that faster RTs involved faster pre- and post-stimulus dynamics whereas slower RTs involved slower dynamics. This relationship between prestimulus state and RT but not choice suggests a gain signal that amplifies the sensory evidence and modulates prestimulus activity rather than a bias to choose a particular side. Furthermore, errors on the previous trial led to shifts in prestimulus state and slower RTs, suggesting that this gain signal is modulated by trial history and linked to an internal speed-accuracy tradeoff.
We used drift diffusion and recurrent neural network (RNN) modeling to test if variability in behavior and neural activity represents fluctuations in a gain signal. Drift diffusion models with a trial-by-trial varying multiplicative gain signal on the sensory evidence provided the best description of the RT and choice behavior. Similarly, in optimized recurrent neural networks, trial-by-trial variation in multiplicative gains on the rectified linear unit (ReLU) nonlinearity were necessary to recapitulate network dynamics consistent with our PMd data.
Collectively, these results suggest that a gain signal dependent upon previous trial outcome alters the prestimulus state and is an important component of decision-related neural population dynamics and behavior.