Deeply Felt Affect: The Emergence of Valence in Deep Active Inference Casper Hesp†*1,2,3,4, Ryan Smith†5, Thomas Parr4, Micah Allen6,7,8, Karl J. Friston4, Maxwell J. D. Ramstead4,9,10 *Corresponding author (email:
[email protected], twitter: @casper_hesp) †Shared first-authorship. These authors made equal contributions and are designated co-first authors. 1. Department of Psychology, University of Amsterdam, Science Park 904, 1098 XH Amsterdam, Netherlands. 2. Amsterdam Brain and Cognition Centre, University of Amsterdam, Science Park 904, 1098 XH Amsterdam, Netherlands. 3. Institute for Advanced Study, University of Amsterdam, Oude Turfmarkt 147, 1012 GC Amsterdam, Netherlands. 4. Wellcome Centre for Human Neuroimaging, University College London, London, UK, WC1N 3BG. 5. Laureate Institute for Brain Research. 6655 South Yale Avenue, Tulsa, OK 74136. 6. Aarhus Institute of Advanced Studies, Aarhus University, Aarhus, Denmark 7. Centre of Functionally Integrative Neuroscience, Aarhus University Hospital, Aarhus, Denmark 8. Cambridge Psychiatry, Cambridge University, 18b Trumpington Road, Cambridge, CB2 8AH 1 9. Division of Social and Transcultural Psychiatry, Department of Psychiatry, McGill University, 1033 Pine Avenue, Montreal, QC, Canada. 10. Culture, Mind, and Brain Program, McGill University, 1033 Pine Avenue, Montreal, QC, Canada. 2 Abstract The positive-negative axis of emotional valence has long been recognised as fundamental to adaptive behaviour, but its origin and underlying function has largely eluded formal theorising and computational modelling. Using deep active inference – a hierarchical inference scheme that rests upon inverting a model of how sensory data are generated – we develop a principled Bayesian model of emotional valence. This formulation asserts that agents infer their valence state based on the expected precision of their action model – an internal estimate of overall model fitness (“subjective fitness”).