Séminaire de Sophie Deneve
Friday February 16, 2024, at 14:30, at Henri Gastaut meeting room
Sophie Deneve (Institut de Neurosciences de la Timone) invited by Fredéric Chavane
Circular inference in brain circuits and social networks
Abstract:
To perceive and act in an uncertain world, we must interpret ambiguous sensory data and predict action consequences. Neural processing could be seen as a form of Bayesian inference where feedforward and feedback processes propagate beliefs about an underlying hierarchy of causes. But is it really so simple? First, deviations from rationality (e.g. confirmatory biases) are common. Second, while average behavior may appear Bayesian, individuals can differ very significantly. Finally, naive belief propagation is a poor method for inference, since it leads to circular inference (CI), a reverberation and amplification of sensory evidence and prior knowledge through feedforward/feedback loops. On the theory side, we showed that CI could lead to aberrant beliefs such as hallucinations or delusions. Moreover, this could be prevented by balancing excitation and inhibition in cortical microcircuits. Interestingly, CI could also account for polarization and radicalization in social echo chambers, suggesting brain-like solutions to address this general societal issue. On the experimental side, we measured significant levels of CI in human subjects performing probabilistic reasoning and bistable perception tasks. Moreover, CI severity was linked to individual propensity to form conspiracy theories. In schizophrenic patients, CI was stronger than in controls and linked to the severity of positive symptoms. Far more work needs to be done, in particular in identifying neural signatures, exploring candidate microcircuit mechanisms/neuromodulation, and extending CI to sensorimotor control or reward-guided behavior.