A Method for Testing Bayesian Models Using Neural Data

Bayesian models have been successful at accounting for human and animal behavior, yet to what degree they can also explain neural activity is still an open question. While decoding approaches that link neural variability to behavioral uncertainty provide some evidence, stronger tests have tried to link posterior beliefs about specific latent variables in a generative model to neural responses. On one hand, the specificity of the resulting predictions is desirable since it allows us to decide which of the infinitely many parameterizations of the task model (ideal observer) is more closely aligned with the brain's internal model. On the other hand, it is unclear under what conditions we can even expect a match of predictions and data given that current models are drastic simplifications of the rich internal model the brain uses. Furthermore, this approach so far has required strong assumptions about how probabilities are represented in neural responses. Here, we formalize and address both of these problems and derive predictions for when they can be overcome. In particular, we show how to meaningfully differentiate between Bayesian models using neural data with a weak assumption about the neural representation of probabilities, i.e. a kind of linearity that holds for a wide class of probabilistic representations including distributed distributional codes (DDCs) and neural sampling schemes. We demonstrate our method by using simulated V1 neural data to differentiate between two Bayesian models for an orientation discrimination task that are practically indistinguishable based on behavior. The first model contains orientation as an explicit variable to be inferred, while the second model assumes inference over a set of oriented gratings. Our results pave the way for strong and rigorous neural tests of Bayesian models of behavior using neural data, and give us deeper insights into how to correctly interpret neural data.