Learning in nonstationary environments with minimal neuralcircuitry
Animals must learn and make predictions in a world that is not only uncertain, but shaped by unmodelled contingencies: unknown unknowns. Despite this, current theories of how this is achieved are based on optimal (Bayesian) strategies in simple settings, where it is possible to fully represent a model of the world and infer its state based on sensory data. This could be impractical in situations where an animal has limited prior experience and cognitive capacity, particularly in short-lived animals with small brains. Using the neural circuitry of Drosophila as inspiration, we derive a learning framework, which we call Hetlearn, for coping in such situations. Hetlearn sacrifices asymptotic optimality on idealised inference tasks, making it less fragile to abrupt changes in environmental statistics and providing more accurate state information than standard Bayesian procedures when data are limited. Hetlearn explains key physiological and anatomical features of the Drosophila Mushroom body (MB), and predicts behavioural features of learning that are conserved across species, such as second-order conditioning and reversal learning.