Modelling Bayesian computation in the brain: unification, explanation, and constraints

Colombo and Hartmann (Br J Philos Sci 68(2):451–484. https://doi.org/10.1093/bjps/axv036, 2017) recently argued that Bayesian modelling in neuroscience can not only unify a diverse range of behavioral phenomena under a common mathematical framework, but can also place useful constraints on both mechanism discovery and confirmation among competing mechanistic models. After reviewing some reasons for decoupling unification and explanation, we raise two challenges for their view. First, although they attempt to distance themselves from the view that Bayesian models provide mechanistic explanations, to the extent that a given model successfully constrains the search space for possible mechanisms, it will convey at least some mechanistic information and therefore automatically qualify as a partial or incomplete mechanistic explanation. Second, according to their view, one widely used strategy to guide and constrain mechanism discovery involves assuming a mapping between features of a behaviorally confirmed Bayesian model and features of the neural mechanisms underlying the behavior. Using their own example of multisensory integration, we discuss how competing mechanistic models can be consistent with all available behavioral data and yet be inconsistent with each other. This tension reveals that there are exploitable degrees of freedom in the mapping relationship between models of behavioral phenomena and neural mechanisms, and points to the role that other background assumptions play including level-assumptions about the appropriate level at which the neural model should be specified (e.g., individual neuron or population level) and localization-assumptions about where in the system the underlying mechanism might occur. These considerations highlight the need for a more refined account of modelling constraints in neuroscience.