Manuel Baltieri
Honorary Senior Research Fellow

Research

Active Inference agents as a model of cognition

My interests span mainly artificial intelligence, computational cognitive neuroscience, information and control theory. In particular my current work is focused on the Bayesian brain hypothesis with an emphasis on predictive coding, a model of the cortex that originally addresses only perception [1] and active inference, an extension of predictive coding that includes accounts of behaviour, under the free energy principle [2].

(This set of ideas has also commonly been addressed as "Predictive Processing" in the cognitive sciences [3].)

Project Outline

Active inference has recently emerged as a possible unifying theory of perception and action in neuroscience and biology. On this view the brain is a probabilistic prediction machine whose role is to make sense of the world and to predict its own sensory input. Models of the visual cortex [1, 4, 5] provide examples of ways to implement such a mechanism.

Typical implementations inspired by work on the cortex propose that to do so, brains must build very detailed and accurate models that fully capture the complexity of the world. On the other hand, a more recent reading addressed as Radical Predictive Processing (RPP) [6] suggests that this need not be the case. In many situations simple representations of the environment with faster, action-oriented responses may provide valuable predictions without building overly accurate, but slow, solutions.

The main goal of my project is to investigate agents implementing action-oriented models in simulated environments inspired by Active Inference. The second one is to explore how such models could be learnt without external aid (e.g. ``teachers" that show how to perform a task). These models will provide results to better understand the implications of Predictive Processing/Active Inference for cognitive neuroscience and to challenge some of the assumptions being taken so far in its implementations. Possible future applications include engineering and control-theory uses, for instance implementations of robotic systems as part of the so-called ``probabilistic robotics" approach [7].

Supervisors: Dr. Christopher Buckley, Prof. Thomas Nowotny

References:

[1] Rao, Rajesh PN, and Dana H. Ballard. "Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects."Nature neuroscience 2.1 (1999): 79-87.

[2] Friston, Karl. "The free-energy principle: a unified brain theory?." Nature Reviews Neuroscience 11.2 (2010): 127-138.

[3] Clark, Andy. "Whatever next? Predictive brains, situated agents, and the future of cognitive science." Behavioral and Brain Sciences 36.03 (2013): 181-204.

[4] Lee, Tai Sing, and David Mumford. "Hierarchical Bayesian inference in the visual cortex." JOSA A 20.7 (2003): 1434-1448.

[5] Friston, Karl. "A theory of cortical responses." Philosophical Transactions of the Royal Society of London B: Biological Sciences 360.1456 (2005): 815-836.

[6] Clark, Andy. "Radical predictive processing." The Southern Journal of Philosophy 53.S1 (2015): 3-27.

[7] Thrun, Sebastian, Wolfram Burgard, and Dieter Fox. Probabilistic robotics. MIT press, 2005.