COMPETITIVE LEARNING, NATURAL IMAGES AND CORTICAL-CELLS

It has been suggested that the paradigm of competitive learning might constitute a viable mechanism by which the response properties of the cells of the visual cortex could develop to form coding units suitable for representing the visual stimuli encountered in natural life. This suggestion was founded on the observation that, in computational simulation, competitive adaptation to stimuli which have the character of natural images can give rise to receptive fields that are primarily sensitive to simple oriented features. In this paper it is argued that the winner-take-all paradigm of competitive learning is not adequate to account for the distribution of responses in the domain of spatial frequency that is necessary to form a complete representation of the information contained in images. The competitive mechanism gives rise to a gross imbalance in the distribution of spatial frequency responses in favour of the lowest spatial frequencies, as a consequence of the form of the spatial frequency spectra of natural images. It is widely believed that the function of the primary visual cortex is to provide an efficient, non-redundant representation of the visual stimulus and it has been suggested that the total activity of the simple cells is at any one time concentrated among relatively few cells. Here it is argued that the competitive mechanism is inappropriate for the establishment of such a scheme of representation. Characterization of the inadequacies of the algorithm provides some insight into the requirements which must be met by a mechanism responsible for determining the response properties of cortical cells from experience of the visual world.