A neuro-computational model of multi-stable perception: Switching border-ownership
Naoki Kogo, Alessandra Galli, Luc Van Gool, Johan P. Wagemans · Lirias · 2009
Perceiving the famous “face or vase” image, we either see the face or the vase at a particular time, and this perception alternates from one to the other. In other words, our perception of one area as figure and the other as background (“figure-ground” segregation) changes in time. This multi-stable perception can also be described in terms of the alternation of border-ownership. The borderlines dividing the face area and the vase area are owned by one of the two areas at a time, and the ownership alternates in time. Multi-stable perception indicates a change in the perception (interpretation) of an input image which is kept constant. It has, therefore, been an important tool to investigate the subjective nature of visual processing. A neuro-computational model with stochasticity and a feedback system is developed to explain the underlying mechanisms of the phenomenon. The core of the model is a feedback system. Once the model starts to detect one of the two objects (face or vase) as a figure, the top-down influence changes the response properties in favour of the detected figure. This, in turn, enhances the figure further. In a face or vase image, the borderlines and junctions are detected first. The border-ownership (BOWN) is computed by the global interaction between all elements of the borderlines and the junctions. The BOWN map indicates the existence of a difference in depth and hence it is considered as a 2-D differentiation of the depth map. Its 2-D integration, therefore, creates a depth map of the image and determines the figure-ground relationships. To introduce stochasticity, random numbers are multiplied to the BOWN signals. The top-down feedback influences the random numbers by skewing their probability density function so that the BOWN signals that are in agreement with the detected figure-ground relationship are enhanced while those in disagreement are inhibited. This enhances the figure-ground segregation further. Adaptation results in a decay of the response, which leads to alternation. This triggers recovery from adaptation. The model is tested with parametrically modified images. By changing the size of the two areas, it is possible to “disambiguate” the image, e.g., perceiving the face more often than the vase. When the image is disambiguated in such a way, the total duration during which the model responds to one figure becomes longer than the other. The model is also tested with “intermittent presentation”: the image is given to the model intermittently, i.e., mixed with blank periods. Experimentally, it has been shown that the alternation is prolonged in other cases of bi-stable perception. Due to the recovery of the adaptation not only during the blank periods but also during the periods when the model is giving the opposite figure-ground relationship to the current response, the model shows a prolongation of the alternation. Therefore, the adaptation and the recovery processes need to be perception-dependent (not physical-input-dependent) to reproduce human perception. This model provides a framework for a dynamic feedback system, which may be playing an important role in visual system.