engleski

A Neurodynamical Account of How Expectations Affect Color Perception in Afterimages

A recent study by Lupyan (2015) showed that adapting to objects with intrinsic colors (tomato) creates stronger afterimages (more vivid colors) than adapting to arbitrarily colored objects (car). Also, stronger afterimages were created by scenes containing intrinsically colored elements (sky) than scenes with arbitrarily colored objects (book) similar to the “Spanish castle illusion”. He interpreted this finding as an evidence for predictive coding models of visual perception and for the cognitive penetrability of vision. Here, we offered an alternative explanation based on the theoretical framework of adaptive resonance theory (ART) that led to opposite conclusion (Grossberg, 2013). In the ART cortical circuit, categorization of sensory input is achieved by matching bottom-up sensory signals with top-down expectations. When the bottom-up and top-down signals are closely aligned, resonant state develops that indicate successful recognition of the input pattern. Resonance corresponds to conscious perception of the presented input. On the other hand, mismatch between bottom-up and top-down signals produces a global reset wave that clears the traces of erroneous top-down expectation. This is a crucial difference with respect to predictive coding models where perception is a compromise between bottom-up and top-down signals. To explain the perception of afterimages, we introduced a model of opponent interactions with slowly adapting neurotransmitter release as a frontend to the ART circuit. It is closely related to the model of gated dipole proposed by Grossberg, Hwang and Mingolla (2002) where adapting to one color (red) produces subsequent reaction in the complement color channel (green) when the network is exposed to the homogenous achromatic field. Also, we developed a real-time implementation of the fuzzy ART algorithm (Carpenter, Grossberg & Rosen, 1991) in order to categorize analog patterns corresponding to different hues. Results of computer simulations showed that expected colors indeed produced stronger and more lasting resonant states comparing to non-expected colors. In the latter condition, several reset waves make network activity less stable and consequently color perception less vivid. Furthermore, it should be noted that the effect of expected colors is short lived and eventually the network dynamics settle to the same neutral gray level in both conditions resulting in veridical perception of achromatic input. In conclusion, we showed that vivid perception of expected colors in afterimages is not necessarily a product of the predictive coding. It also arises in a neural network where cognitive penetration or top-down influence is restricted and tightly controlled in order to achieve stable and unbiased categorization and perception.

cognitive penetrability of vision ; color perception ; afterimages ; neural networks

nije evidentirano