Active Inference in Cortical Circuits

This aim of this project is to produce new machine learning constructs based on theories of
brain dynamics, particularly those of functional connectivity and self-organisation in the
cortex. We will develop biologically interpretable agents that embody theoretical constructs
of Active Inference and the Free Energy Principle. Such an agent will aim to minimize its
own free energy, or the divergence between perception and its internal representation of the
environment, by either interacting with its environment in a way that fulfils its own
predictions or by learning updating its own internal model of the environment. Such
adaptive model-driven agents may be able to benefit from reduced learning times through
hierarchical learning or be more robust when faced with dynamic or noisy environments.
Our research pursues two agendas - healthcare and robotics.
In Robotics:
These agents will be capable of making inferences about their environment and about future
events and as such will make interesting candidates for solving general problems in artificial
intelligence. Using the AI Gym & Universe platforms we will investigate habitual and goaldirected
behaviours in a wide range of environments. An optimal agent will be able to
generalise and perform well regardless of the determinability, stasis, observability or agency
of a given environment.
Multiscale computational models will be used as tools to investigate functional and structural
cortical circuits and to identify network properties and constraints that can be modelled in
silico and validated in machina. Novel cognitive-affective behaviours or topographical
properties observed during the modelling phase will be embedded into robotic agents.
These algorithms will be assessed by the agent's ability to perform meaningful benchmark
associative learning & sensory motor integration tasks. A key example of where the Free
energy principle offers new computational avenues is in the explore-exploit trade-off. This
classical dilemma of biological & artificial agents is often hand-crafted. Using this
theoretical framework this problem is solved by considering action policies with epistemic
& pragmatic outcomes.
In Healthcare & Computational Psychiatry:
We will pursue computational psychiatric applications of our models. The main
advantage of applying a Free Energy model for phenotyping behaviour in neuropsychiatric
diseases like dementia or schizophrenia is that is subsumes perception, learning, memory
and action. Hence parameterizations for a given individual in one task should be able to
predict behaviour in many (most) cognitive protocols - providing for consistent comparisons
across different tasks & laboratories. The formulation has recently been applied to
neuroeconomic games such as probabilistic lotteries involving food rewards and
demonstrated that individual choices in these games can be best explained by a
combination of individual preferences - mathematically represented by prior beliefs - and
individual confidence in the impact of actions selected - mathematically represented by a
precision or inverse temperature parameter. Importantly elements of the update procedure
within this modelling framework have been mapped to putative neurobiological components.
This attendant process model hypothesises prior preferences may be encoded in specific
cortical regions e.g. ventromedial prefrontal cortex, while the inverse temperature or
precision parameter may be encoded by neuromodulatory systems, such as dopamine or
acetylcholine. Thus given a sufficiently simple design for patient populations, where the
task structure can be understood, distinctions between cortical & subcortical effects on
behaviour can be distinguished. We will use the games developed above to test for
correlations of model parameters & behavioural phenotypes in control subjects who
submit personality questionnaires - employing the amazon trunk & local data acquisition.