A Neural Network Model of Dynamically Fluctuating Perception of Necker Cube as well as Dot Patterns

The mechanism underlying perceptual grouping of visual stimuli is not static, but dynamic. In this paper, the dynamical grouping process is implemented with a neural network model consisting of an array of (hyper)columns suggested by Hubel ̄ "vViesel, where intracolumnax inhibition and intercolumnax facilitation axe incorporated ̄ The model was applied successfully to figures consisting of a set of dots yielding either of two ways of groupings from time to time due to neural fluctuations and fatigue. Then the model was extended to introduce dependency on fixation points as well as neural fluctuations and fatigue. Then, it was applied to the Necker Cube. The model output from time to time either of two ways of 3D interpretations depending on the fixation points. Introduction Perceptual grouping plays an essential role in segmenting objects in the scene and recognizing each of them. Gestalt psychologists have proposed that there are several factors underlying the grouping: they are factor of proximity, factor of similarity, factor of smooth continuation, and so on. Recently, computer implementations of the grouping processes have been reported (Stevens 1978; Hiratsuka, Ohnishi, and Sugie 1992). However, the mechanism underlying perceptual grouping of visuM stimuli is not static; but dynamic as in Marroquin pattern (Fig.l) (Marr 1982). The namical aspect of grouping seems to reflect the flexible nature of human visual information processing to deal with ambiguous patterns. However, it has not been studied seriously. In this paper, the dynamical grouping process is implemented with a neural network model consisting of a 2D array of hypercolumns suggested by Hubel & Wiesel (1977), where intracolumnar inhibition and intercolumnar facilitation are incorporated. The model was applied successfully to figures which consist of a set of dots yielding either of two ways Copyright Q1999, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. ° Figure 1: Marroquin pattern. Figure 2: Necker Cube. of groupings from time to time due to neural fluctuations and fatigue ̄ Moreover, the model was able to interpret line drawings, a grouping at a higher level; it was applied to the Necker cube (Fig.2). It also exhibited dependence on fixation points about the Necker cube. The model output either of two 3D interpretations reflecting fixation point dependence as reported by Kawabata et al. (1978). Neural Network Model This model is based on the neural network model consisting of hypercolumnar structure. It has been used to explain the early visual process such as retinal rivalry (Sugie 1982). We extend it to deal with highly ambiguous figures of more than two interpretations. From: AAAI-99 Proceedings. Copyright © 1999, AAAI (www.aaai.org). All rights reserved.