Rapid Categorization of Extrafoveal Natural Images: Implications for Biological Models

Abstract

Despite intensive research over the past 30 years, the performance of artificial visual systems in object recognition is still poor when compared with humans. For humans, the identification of objects in natural scenes appears both effortless and fast. Just how fast was a question we addressed recently in a study that associated behavioral measurements and event related potential (ERP) recordings¹. The task used was a go/no-go visual categorization task in which human subjects had to respond when a photograph of a natural scene contained an animal. The photographs had never been seen before and were flashed centrally on a screen for only 20 ms. Humans scored 94% correct, moreover, their ERPs recorded on animal and non-animal trials showed a clear difference on all frontal electrodes that started around 150 ms after stimuli onset. Thus, it appears that the human visual system can process such previously unseen complex natural scenes in less than 150 ms, a level of performance well above that of any currently available artificial system. This is despite the fact that the neurons that constitute the human visual system are relatively slow — firing rates rarely exceed 200 Hz, a value far slower than the transistors of a modern microprocessor, which can change state over a million times faster.