Sensory priors, and choice and outcome history in service of optimal behaviour in noisy environments

While navigating the world, we constantly make decisions under uncertainty. In face of noisy perceptions, we build and continuously update our priors and expectations about future events in order to efficiently estimate and map sensory information to actions. In a non-stationary environment, exploratory behaviour may represent an optimal strategy to optimise reward at any given context. Even in stationary environments, animals may perform this innate exploratory behaviour and exhibit unexpected biases. In this work, we investigate how previous experience – sensory priors, as well as choice and outcome history ­– can inform subjects’ decision-making in a perceptual categorization task, where sensory statistics is manipulated. We compare human, rat, and mouse performance on a 2-alternative-forced-choice (2AFC) sound categorization task. Importantly, we expose subjects to different stimulus statistics, while keeping category probabilities equal. Humans, rats, and mice develop a stimulus-dependent bias that increases reward to a near-optimal level. Despite the overall stimulus-dependent bias that is similar across species, humans vs rodents show distinct trial-to-trial learning updates. Our proposed computational model accounts for this bias and predicts different trial-to-trial updates. Finally, we perform neuronal recordings from rat prefrontal cortex (PFC) and posterior thalamus. Stimulus identity and category are encoded on a single-neuron level in both areas. Additionally, animal choice can be decoded from neuronal population activity. Only in PFC, inter-trial activity reflects animal optimal biases and hence exhibits signatures of sensory prior representations. Overall, our study unravels an optimal behaviour that emerges from a trial-to-trial learning strategy that is different across species. Our results also suggest that the prefrontal structures in the rat brain might encode sensory prior representations.