ausblenden:
Schlagwörter:
-
Zusammenfassung:
Perceptual multistability (when two or more percepts alternate in response to a single ambiguous sensory input) has been studied for centuries using myriad approaches, and has illuminated diverse cognitive functions (e.g., perceptual inference, attention, visual awareness) [1-3]. Traditionally, multistability has been viewed in Helmholtzian terms, i.e., treating perception as a passive [Bayesian] inference about the contents of the world [2,4]. However, this view neglects the crucial role played by value [3,5-8]: e.g., percepts paired with reward tend to dominate for longer periods than unpaired ones [5-6]. We reformulate visual multistability in terms of a decision process, employing the formalism of a partially observable Markov decision process (POMDP) [9]. Each percept is potentially associated with different sources of rewards or punishments (including aesthetic value [10]), and switching between percepts is a form of (costly) internal action - the attentional equivalent of the external action of moving eye gaze between objects. Selecting one percept is accompanied by reduced observation noise, and ultimately stronger beliefs about the perceived state (dominant percept). The solution of the POMDP is the (approximately) optimal perceptual policy; this replicates and explains several classic and elusive aspects of rivalry. It reproduces apparently spontaneous random switches, with roughly gamma-distributed dominance periods (which are two key hallmarks of multistability [1], see Figure 1A-B). It captures the modulation by reward (Figure 1C-E) [5-8]. It explains the rich temporal dynamics of perceptual switching rates [11], i.e. the increase in switching rate initially observed in naive participants, then decreases within single observation periods in subsequent sessions, and finally slowly increasing switching rates across days. To our knowledge, this model is unique in explaining the last two observations. Overall, our value-based decision-making account of perceptual multistability synergizes with previous accounts and also offers a more comprehensive treatment of computational and algorithmic facets of multistability. Furthermore, the dynamic nature of value in our framework might help explain the differences reported between psychiatric and healthy populations that concern the temporal dynamics of perception (e.g., the rate of perceptual switching, see [3]).
