Learning in nonstationary environments with minimal neural circuitry

Animals inhabit continually changing environments where it is not always possible to infer causes of relevant changes, such as the appearance of a new threat. In such nonstationary settings, learning a predictive model is challenging because a surprising observation could be due to chance, or due to systematic but uninferrable causes. This issue magnifies for short-lived animals with small nervous systems and limited opportunity to explore. To understand how the nervous system negotiates these challenges, we derived a new learning principle ‘Hetlearn’ (Heterologous Learners) based on the physiology and multi-compartmental architecture of the Drosophila mushroom body, a key associative learning circuit. Hetlearn circumvents learning full causal models of the environment, making it less fragile to nonstationarity. In stationary settings, Hetlearn is efficient and outperforms Bayesian inference when data are limited. Hetlearn predicts observed learning behaviour in Drosophila across multiple tasks and provides new, testable predictions about circuit interactions during learning.