Speaker
Description
Supporters of the Bayesian Brain Hypothesis argue that the brain maintains a generative model of the environment and inverts this model to infer the states of the world. Message-passing is one class of algorithms that performs this computation tractably. While there is some evidence that the brain uses specific forms of message-passing, such as mean-field variational inference, other algorithms, such as Thouless-Anderson-Palmer message-passing, remain plausible candidates but have yet to be thoroughly considered. In this study, we investigate the measurement requirements for comparing message-passing algorithms as alternative explanations of data. We examine how the spatial precision of neural recordings would affect the ability to determine which message-passing algorithm the brain implements. Even with exact coarse-graining procedures using the renormalization group, we were not able to identify the ground-truth implemented by individual neurons. Thus, fine-grained measurements may be necessary to draw conclusions about how the brain could implement Bayesian inference through message-passing.
| Affiliation | University of Sussex |
|---|---|
| Career status | PhD student |