Affective Active Inference and Precision Estimation

Abstract

As Neuroscience progresses, there is an increasing amount of research that endorses predictions and reducing of prediction errors as one of the main functions of the brain. active inference is a brain-inspired, mathematical framework that successfully implements this idea both in simulations as well as in robotics. The predictive nature of active inference might make current artificial intelligence agents more adaptive. However, the motives of these agents are often still hardwired as attractor dynamics or learnt using over-engineered rewards. Nature has come up with a different way of providing intelligent beings with drives for their actions: affect, more commonly known as emotions. Although various models integrate affect into active inference , none have yet applied Mark Solms' definition within a continuous active inference framework. According to Solms' interpretation, affect acts as an evaluative monitoring mechanism of an organism's homeostatic states and guides it through unpredictable environments. This active monitoring of homeostatic states is what according to Solms stands on the basis of consciousness. Key here is the prioritization of different homeostatic needs, where deviations in the most salient category of need come to the organism's affective(conscious) awareness. Mark Solms proposes that computationally, affect is constituted by the inference of changes in precision. Where increases in precision are positively- and decreases in precision are negatively valenced. This change in precision is obtained by performing a gradient descent on free energy with respect to precision, which results in an incremental precision updating scheme that determines the salience of prediction errors. This offers an adaptable mechanism that allows context, through precision modulation, to determine the relative influence of prediction errors.  This in turn allows an agent to prioritize homeostatic needs i.e. letting certain needs come to conscious awareness. "Context" in the light of Solms' research is defined as either: the relation of needs with respect to other needs or the relation of needs with respect to external opportunities. This research supports Solms' theory on affect and consciousness by successfully providing a computational implementation that can, through precision optimization, perform the prioritization of needs directed by "context" as just defined. By doing this successfully, this research shows that the principles used could potentially be useful in continuous active inference implementations, improving their adaptability.