Datensatz

Zusammenfassung

Multi-stable perception refers to the association of the same visual stimulus with multiple alternative percepts. So far multi-stability has been studied primarily in the context of low-level vision and shape recognition. Multi-stability has also been observed during the perception of body motion, especially if the associated depth information is ambiguous [1]. In this case the same action stimulus is associated, for example, with multiple alternative walking directions. In psychophysical experiments it has been demonstrated that body motion perception can show spontaneous perceptual switching between different interpretations and hysteresis, if a stimulus parameter is gradually varied that introduces a bias for one of the two perceptual interpretations. We present a physiologically-inspired neural model that provides a unifying account for this perceptual multi-stability and multiple psychophysical experiments that characterize the underlying perceptual dynamics. Our model includes the following parts: (1) a deep neural hierarchy that recognizes body shapes from silhouette features and shading gradients of the moving figure; (2) a fast dynamic neural layer that can be interpreted as 2D neural field whose dimensions encode the stimulus view and the temporal order of the body shapes within action sequences; (3) a slower bistable read-out network that pools neural responses over the body shapes belonging to the same action and view over time-points. Our model provides a unifying account for a number psychophysical results from the literature, and from our own experiments: (a) dependence of percept probabilities on shading cues, (b) illusory misperception of walking direction for body stimuli that are illuminated from below [2]; (c) perceptual hysteresis for the gradual variation of disambiguating shading cues of body motion stimuli [3]. Our results show that the multi-stability of body motion perception can be accounted for in a simple way by the interaction between deep example-based neural networks for the recognition of body shapes and an elementary physiologically plausible cortical activation dynamics (Fig. 1).