Predicting individual learning trajectories in zebrafish via the free-energy principle

Abstract

The free-energy principle has been proposed as a unified theory of brain function, and recent evidence from in vitro experiments supports its validity. However, its empirical application to in vivo neuronal dynamics during active decision-making remains limited. This work reverse-engineered generative models--cast as canonical neural networks--from large-scale calcium imaging data of zebrafish performing a visually guided go/no-go task in a virtual-reality environment. Leveraging the formal equivalence between neural network dynamics and variational Bayesian inference, we constructed biologically plausible synthetic agents capable of active inference. These agents recapitulated individual variability in zebrafish neuronal dynamics and behaviour by identifying subject-specific prior and posterior beliefs. Additionally, they enabled quantitative predictions of long-term changes in neural activity, effective synaptic connectivity, and behavioural performance, including task accuracy after training. Our results demonstrate a powerful framework of active inference for modelling in vivo neuronal self-organisation and highlight the predictive validity of the free-energy principle in behaving animals.