Bistable Perception in Conceptor Networks

Bistable perception describes the phenomenon of perception alternating between stable states when a subject is presented two incompatible stimuli. Besides intensive research in the last century many open questions remain. As a phenomenon occurring across different perceptual domains, understanding bistable perception can help to reveal properties of information processing in the human brain. It becomes apparent that bistable perception involves multiple distributed processes and several layers in the hierarchy of sensory processing. This observation directs research towards general models of perceptual inference and to the question whether these models can account for the spontaneous subjective changes in percepts that subjects experience when shown rivalling stimuli. We implemented a recurrent generative model based on hierarchical conceptors to investigate its behaviour when fed an ambiguous signal as input. With this model we can show that (1) it is possible to obtain precise predictions about the properties of bistable perception using a general model for perceptual inference, (2) hierarchical processes allow for reduction in prediction error, (3) random switches in the percept of the network are due to noise in the input and (4) dominance times exhibit a gamma distribution of stimulus dominance times compatible with experimental findings in psychophysics. Code for the experiments is available at https://​github.​com/​felixmzd/​Conceptors.