CONNECTIONIST APPROACHES TO VISUALLY-BASED FACIAL FEATURE EXTRACTION

We examine here some properties of a connectionist autoassociative matrix for storing, in a parallel and distributed fashion, face stimuli that are coded as simple patterns of spatially varying light intensities. First, we find that the opposition of positive and negative point contributions for nearly all the eigenvectors forms head/hair shapes, often containing the positions and shapes of eyes. Second, the opposition of positive and negative points that contribute strongly to the determination of the first eigenvector appear to separate male and female head/hair shapes. We find also that pixel positions that contribute strongly to the eigenvectors generally form spatially contiguous groups in the face pattern, often form face/head shapes, and occasionally consist of points that form a hairstyle. The results are discussed in terms of previous results indicating the salience of these ’features’ for discrimination and identification, and in terms of Bruce & Young’s (1986) visually-derived semantic