Automated Animal Training and Iterative Inference of Latent Learning Policy

Abstract

Progress in understanding how individual animals learn requires high-throughput standardized methods for behavioral training and ways of adapting training. During the course of training with hundreds or thousands of trials, an animal may change its underlying strategy abruptly, and capturing these changes requires real-time inference of the animal’s latent decision-making strategy. To address this challenge, we have developed an integrated platform for automated animal training, and an iterative decision-inference model that is able to infer the momentary decision-making policy, and predict the animal’s choice on each trial with an accuracy of ~80\%, even when the animal is performing poorly. We also combined decision predictions at single-trial resolution with automated pose estimation to assess movement trajectories. Analysis of these features revealed categories of movement trajectories that associate with decision confidence.

Tasks

Reproductions