Abstract
A wave of recent findings shows that it is possible to classify features, objects or scenes subjects are perceiving by “reading” their brain activity. Going beyond classification, (Kay et al., 2008) showed that a presented natural image can be identified among a large number of candidate images based on fMRI data. In that study a natural image was described by a comprehensive receptive field model consisting of many Gabor wavelets covering retinotopic locations of the visual field. We investigated whether a neural response model based on the two-parameter Weibull contrast distribution (Scholte et al, 2009) permits rapid natural image identification. We measured EEG activity of 32 subjects viewing brief flashes of 700 natural scenes. From these measurements, and from the contrast distribution of these scenes, we derived an across subject Weibull response model. We used this model to predict EEG responses to 100 new natural scenes and to estimate which scene from this set subjects viewed by finding the best match between predicted and measured EEG responses. In 90 percent of the cases our Weibull response model accurately identified the viewed scene. A different experiment with artificial occlusion images resulted in almost 95 percent correct identification. These image identification performances are comparable to the results of (Kay et al, 2009). We conclude that it is possible to identify natural images humans view on the basis of EEG data by establishing a relationship between neuronal responses to and Weibull contrast distribution of these images. The potency of this relationship lies in the ability of the Weibull distribution to structure the space of natural images in a highly meaningful and compact way, based only on two parameters. Our results support the idea that our brain may have evolved, among others, to estimate Weibull statistics of natural images for rapid scene gist extraction.