Wellcome Centre for Human Neuroimaging, Institute of Neurology, UCL, UK
Title: The anatomy of (active) inference
Modern theories of brain function often appeal to the metaphor of the brain as a scientist, engaged in a process of inference about incoming sensory data, and selecting actions to obtain new data. This can be formalised through an appeal to Bayesian inference – where perception and action can be thought of as optimising the evidence (a.k.a. marginal likelihood) for a statistical model of the world. This view has important consequences for brain circuitry, in virtue of the Markov blankets (statistical independencies) implied by a given model of the world. In brief, if we knew everything about the Markov blanket of a variable, we would need no additional information to draw inferences about it. This fact is exploited to derive efficient inferential message passing schemes in machine learning and provides a useful perspective on the local (synaptic) computations that underwrite inference in biological systems. Importantly, this means that the generative model employed by the brain is an important constraint upon (and is constrained by) brain architectures. Drawing from examples in active vision and oculomotion, this talk will unpack the notion of a generative model and the neuronal message passing it implies. In doing so, we will illustrate through simulation how this approach may be used to reproduce features of behaviour, electrophysiology, and pathology.
Thomas began his academic career at University College London (UCL) medical school. After completing an undergraduate degree in Medical Sciences with Neuroscience, he enrolled on the UCL MBPhD programme, combining his PhD with clinical studies. He works in the theoretical neurobiology group, led by Professor Karl Friston, at the Wellcome Centre for Human Neuroimaging at the UCL Institute of Neurology. His research interests include active inference, computational neuropsychology, and the oculomotor system.