Visual perception has all the hallmarks of an ongoing, cooperative-competitive process: probabilistic outcome, self-organization, order-disorder transitions, multi-stability, and hysteresis. Accordingly, it is tempting to speculate that the underlying collective neural activity performs an exploratory attractor dynamics (spontaneous transitions between distinct steady-states), perhaps at multiple spatial and temporal scales. Here I summarize our recent investigations of this dynamical hypothesis. In each case, a careful empirical study of perceptual dynamics fully constrains an idealized model of the stochastic dynamics of collective neural activity:
• The path-dependence of motion grouping (e.g., when motion coherence follows a random walk), reveals the effective energy landscape and relaxation time of grouping percepts, experimentally confirming the simultaneous presence of distinct attractor states.
• The scalar property of perceptual dominance times is readily explained by stochastic accumulation of activity across multiple independent nodes (idealized cortical columns), but not by other kinds of stochastic processes (e.g., diffusion-to-bound).
• The paradoxical input dependence of perceptual dominance in multi-stable phenomena (‘Levelt’s propositions’) constrains concurrent processing at two levels: stochastic accumulation of evidence by independent lower-level nodes, and cooperative-competitive interactions between tightly coupled upper-level nodes. This experimentally derived architecture maps naturally onto the well-known ‘saliency map’ and ‘predictive coding’ schemes of visual processing.
I conclude that the dynamical hypothesis outlined above permits a particularly close and direct back-and-forth between perceptual experiment and computational theory and thus has the potential to dramatically accelerate our progress in understanding visual function.