Humans, rats and mice show species-specific adaptations to sensory statistics in categorisation behaviour

One key aspect of animal intelligence is the ability to find, represent and update knowledge of statistical regularities in the sensory environment, in service of adaptive behaviour. This allows animals to build appropriate priors, in order to disambiguate noisy inputs, make predictions and act more efficiently. Here, we compare human, rat, and mouse performance on a 2-alternative-forced-choice sound categorisation task. Importantly, we manipulate the stimulus statistics without affecting the overall weight of each category, and we investigate decision-making in such distinct ‘statistical contexts’ associated with different stimulus prior probabilities. Humans, rats, and mice develop a sensory-dependent bias that increases reward to a near-optimal level. We also explore how flexible animals are in learning different types of statistics, without any explicit cue, when switching from one prior probability to another. All species can adapt to a new underlying distribution by gradually shifting their psychometric behaviour after the switch to a new statistical context, with mice showing the slowest rate of adaptation. Despite the overall sensory-dependent bias that is similar across species, humans versus rodents show distinct trial-to-trial learning updates. Rodents show an overall win-stay choice bias, with rats incorporating decision confidence more strongly than mice. Humans, on the other hand, show a more complex learning update, and more indicative of utilising detailed sensory statistics. We propose two classes of computational models, a boundary learning (discriminative) and a stimulus-category learning (generative) model. We show how they can account for the optimal exploitation of the sensory statistics, and predict the observed species-specific trial-to-trial updates.