Three-dimensional visual flow caused by self-movements provides important cues for the perception of space, and thus for the perception and control of own movement. At least 3 depth-cues built up by moving elements are contained in such a flow. One is purely dynamic signal independent of a visual pattern of a facing scene, i.e., ‘speed-gradient.’ The other two are rather static signal called ‘perspective,’ i.e., ‘size-gradient’ and ‘density-gradient’ in a texture pattern. We investigated relative contribution of the three cues for the perceptual impression of 3-dimensional space by presenting computer-generated virtual flow stimuli in which the three cues are controlled independently. As expected, the contribution of speed-gradient was most strong, a medium was texture size-gradient, and the weakest was texture density-gradient. We also investigated response-properties of cells in macaque MST-area using similar stimuli. MST cells are classified into three groups according to the different response-behaviors to 2-dimensional and 3-dimensional visual flow. Based on these results, a psychology-physiology correspondence is suggested.
We investigated the receptive fields of neurons in the visual cortex from a theoretical viewpoint. It was assumed that if two regions in the visual field are located far apart, local images that fall on these regions are mutually independent. An information-maximization solution was used to determine the optimal receptive field by first assuming a low signal-to-noise ratio limit. We found that the optimal receptive field is a spatially localized, Gabor-function-like receptive field, similar to that of a simple cell.
Recalling process of associative memory on the Little-Hopfield model is formulated based on an assumption that distortions of probability density of the inner potential from normal distribution are caused by self-correlations. The overlap (direction cosine) is formulated as an expected value of output function in terms of a pseudo Gaussian probability density. It is shown that the probability density as well as the overlap is in good agreement with simulations for various output functions including a non-monotone function.