Deeply Felt Affect: The Emergence of Valence in Deep Active Inference

Abstract The positive-negative axis of emotional valence has long been recognized as fundamental to adaptive behavior, but its origin and underlying function have largely eluded formal theorizing computational modeling. Using deep active inference, a hierarchical inference scheme that rests on inverting model how sensory data are generated, we develop principled Bayesian valence. This formulation asserts agents infer their state based the expected precision action model—an internal estimate overall fitness (“subjective fitness”). index subjective can be estimated within any environment exploits domain generality second-order beliefs (beliefs about beliefs). We show maintaining representations allows ensuing affective agent optimize confidence in selection preemptively. Valence turn optimized by leveraging (Bayes-optimal) updating term for fitness, which label charge (AC). AC tracks changes estimates lends sign otherwise unsigned divergences between predictions outcomes. simulate resulting subjecting an silico T-maze paradigm requiring context learning, followed reversal. offers account link affect, (mental) action, implicit metacognition. It characterizes biological system reduce uncertainty such inferences through (i.e., top-down modulation priors underwrite confidence). Thus, demonstrate potential provide computationally tractable affect. Our demonstration face validity utility this represents first step larger research program. Next, leveraged test hypothesized role fitting behavioral neuronal responses.