Bradley, S. D. (2004). Visual expectancy and orienting behavior in a recurrent artificial neural network [Abstract]. Psychophysiology, 41(Suppl. 1), S62.

 

Abstract
The evolutionary importance of the orienting reflex (OR) has been known for decades. To paraphrase Pavlov, if an animal does not orient to novelty in its environment, it will not live long. Although much is known about the OR, the computation of novelty at the neuronal level is not fully understood. A variety of stimuli trigger an OR, and at some level they must elicit the same pattern of low-level neural activation. Sokolov and colleagues posit that “expectancy neurons” compare incoming stimuli with what is expected in short-term memory. This study outlines the development of a connectionist simple recurrent network (SRN) that displays orienting behavior without having been taught to do so. Participants in laboratory experiments reliably orient to scene changes (cuts) in television programs. In order to capture the dynamic nature of visual processing, a series of television clips was input to the SRN frame by frame. The task of the network was to pre! dict the pixel values for the subsequent frame. The model learns to “expect” consistency in the visual world, and predictions falter noticeably at scene boundaries. When model expectancy is plotted over time as perfect performance minus error, the resulting curve closely resembles a cardiac response curve characteristic of orienting to scene changes. Model data are compared to human EKG data recorded while participants watched television clips. Thus, this simulation suggests that neurons that learn to expect consistency from incoming stimuli are sufficient to trigger an OR without higher-order processing.