Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

How to test Bayesian models using neural data

Gabor Lengyel¹, Sabyasachi Shivkumar², Ralf M Haefner¹; ¹University of Rochester, ²Columbia University

Presenter: Ralf M Haefner

The Bayesian Brain hypothesis suggests that the brain can be understood in terms of Bayesian computations. While many studies have provided perceptual and sensorimotor evidence for this hypothesis, the question of whether neural responses can also be understood in Bayesian terms remains open. Answering this question requires the specification of two main unknowns: (1) what is the generative model that relates the variables inferred by some population of neurons to the sensory observations, and (2) what is the 'neural code', i.e. what is the relationship between posterior beliefs and neural responses? Much attention has been directed at answering the second of these questions while ignoring the first question, however without reaching consensus. At least in part this is because a given set of observed neural responses can imply different codes under different assumptions about the generative model. Here, we propose answering both questions in the opposite order. First, we present a method to test a given generative model using metamers -- stimuli that give rise to the same posterior under this model -- and confirming that they elicit the same neural responses. This approach can be interpreted as a special case of representational similarity analysis, and generalized accordingly. Second, we propose a `mixture method' that tests whether the relationship between posteriors for different stimuli matches the relationship for the measured neural responses to the same stimuli. If applied to the full response distribution, model and data are only expected to match for neural sampling codes. If applied to average neural responses, they are expected to match for any linear distributional code, including neural sampling and distributed distributional codes, but not probabilistic population codes. We illustrate our approach using simulations where the ground truth is known -- both for a sparse coding model of V1, and a hierarchical motion model for area MT.

Topic Area: Methods & Computational Tools

Extended Abstract: Full Text PDF