Toward a decision-theoretic model of perceptual multistability

Perceptual multistability (PM) has been studied for centuries using a diverse range of approaches. Insights derived range from core principles of information processing such as perceptual inference, to high-level concerns such as visual awareness. The dominant computational explanations of PM are based on analysis by synthesis [1]. However, they fail to account for the crucial role played by value, e.g. with percepts paired with reward dominating for longer periods than unpaired ones [2]. We reformulate PM in terms of dynamic, value-based, choice, employing the formalism of a partially observable Markov decision process (POMDP) and using binocular rivalry as an example. As in conventional Bayesian accounts of perception, we consider there to be a single actual state of the world at any one time to be inferred; but (in the simplest case) with two plausible, but incompatible candidates for what that state might be. Given inevitable environmental volatility, the generative model allows for stochastic dynamics in the actual state. Perceiving one of the states is assumed to be accompanied by reduced observation noise and thus stronger belief about the perceived state. Switching between percepts is a form of (costly) internal action - the attentional equivalent of the external action of moving eye gaze from one object to another. Given implicit reward or punishment associated with the aesthetic value (AV) of the percepts and/or explicit reward or punishment (e.g through conditioning to an image as [2-6]), the optimal policy of the agent is the solution of the described POMDP; and specifies apparently spontaneous random switches, with approximately gamma-distributed dominance periods - two hallmarks of PM (Fig 1A-B). Also, the model captures the modulation by the reward that eluded previous models of PM (Fig 1C-D). We appeal to a recent theory of AV [7] which derives it from the task of optimizing a sensory system to process present and future stimuli well. The resulting values are time-dependent and influenced by novelty as a form of exploration. Taking this into account, we can explain the rich temporal dynamics of perceptual switching rates, i.e. an initial increase for the naive participants, a decrease within a single trial, and a longer-timescale speeding across several sessions [8]. In sum, our value-based decision-making account synergizes with previous models and also offers a more comprehensive treatment of computational and algorithmic facets of PM.