Transitions in dynamical regime and neural mode underlie perceptual decision-making

Abstract

Perceptual decision-making is the process by which an animal uses sensory stimuli to choose an action or mental proposition. This process is thought to be mediated by neurons organized as attractor networks. However, whether attractor dynamics underlie decision behavior and the complex neuronal responses remains unclear. Here we use an unsupervised, deep learning-based method to discover decision-related dynamics from the simultaneous activity of neurons in frontal cortex and striatum of rats while they accumulate pulsatile auditory evidence. We show that contrary to prevailing hypotheses, perceptual choices emerge from the dynamics driven by sensory inputs that are not aligned to discrete attractors in the input-independent dynamics. Input-driven and -independent dynamics differ in strength across the decision state space, resulting in the input-driven dynamics playing a dominant role in evidence integration, while input-independent dynamics playing a principal role in decision commitment. An extension of the classic drift-diffusion hypothesis to approximate the non-canonical attractor dynamics precisely predicts the internal decision commitment time and captures diverse and complex single-neuron temporal profiles, such as ramping and stepping. It also captures choice behavior and trial-averaged curved trajectories and reveals distinctions between brain regions. Thus, non-canonical attractor dynamics inferred from unsupervised discovery conceptually extend a classic hypothesis and parsimoniously account for multiple neural and behavioral phenomena.

Dynamic reinforcement learning reveals time-dependent shifts in strategy during reward learning

“…However, it does suggest that the inclusion of these additional metrics may provide a richer description of changing strategy that modeling choice alone cannot capture. Indeed, models combining both behavioral and neural data have found success in capturing behavioral and neural patterns that could not be captures by one modality alone ( Shahar et al, 2019; DePasquale et al, 2022; Luo et al, 2023 ). Furthermore, while the MoA-HMM has been applied here in the context of RL, the framework can be applied more generally to arbitrary learning rules and tasks, and would apply particular naturally to rule-switching or task-switching settings.…”

Neural circuit models for evidence accumulation through choice-selective sequences

“…This transition could reflect a change in dynamics as a result of decision commitment 109,110 or increasing urgency 111,112 , and be produced in our model by increasing the strength of inhibition between the chains at later positions. The observation of a transition across the population from graded evidence coding to binary choice coding is reminiscent of the conclusions from a subpopulation of cells in another accumulation of evidence task in rats 110 .…”

“…While the graded representations of evidence during the early cue period in ACC and RSC are consistent with the competing chains models of evidence accumulation, our models do not explain the transition towards choice tuning later in the trial. This transition could reflect a change in dynamics as a result of decision commitment 113,114 or increasing urgency 115,116 , and could be produced in our model through many different mechanisms, including by increasing the strength of inhibition between the chains at later positions or introducing a saturating nonlinearity 33 . The observation of a transition across the population from graded evidence coding to binary choice coding is reminiscent of the conclusions from a subpopulation of neurons in another accumulation of evidence task in rats 114 and from the evolution across a trial of the power to predict choice from neural responses in monkeys 40 .…”