EXPLORING NEURAL IMPLEMENTATION OF BAYESIAN INFERENCE BY A VARIABLE-RESISTANCE LEVER PUSH-PULL TASK

While the Bayesian brain hypothesis suggests that the cortex integrates prior knowledge and sensory evidence to estimate latent states, how neurons in different cortical layers represent underlying variables remains elusive.
We addressed this by imaging the primary somatosensory cortex forelimb area (S1fl) in mice performing a novel active perceptual decision-making task with a motorized lever with variable resistances for push and pull movements. Mice were rewarded for moving the lever in the direction of less resistance, the probability of which was manipulated in a block-wise manner. The mice's choices (12 mice, 124 sessions) depended on both the prior probability and the sensory evidence.
To investigate the neural basis, we performed single-cell resolution calcium imaging of S1fl across layer 2/3 (L2/3) and layer 5 (L5) using a miniscope and prism lens (16 sessions, 3 mice, 7096 total cells). Hierarchical clustering identified functional neural ensembles that modulated their activity during the action execution or motor preparation. L2/3 contained a higher percentage of neurons that showed either activation or inactivation during lever pushing.
To formalize this process, we constructed a Bayesian inference model that estimates the state of resistance based on the relationship between the force exerted by the mouse and the resulting lever movement. By combining model-based reinforcement learning with the Bayesian model, we reproduced behavioral characteristics, including psychometric curves and “change of mind” behavior.
Investigating layer differences in sensory inference using the behavioral task and computational model established in this study remains a subject for future research.