One of the least understood aspects of mammalian vision is the ability to recognize scenes through significant degradations in image quality. Neural receptive fields have traditionally been described with coherent structures - for example, oriented gratings in V1. However, this does not address how neurons respond to noisy, less coherent visual input, which is arguably more prevalent in the natural world. Previous studies with natural images show that recognition is highly non-linear with respect to noise, and more importantly, that recognition in noise is facilitated by prior experience with the stimuli (Sadr and Sinha, 2004). These studies entail using RISE sequences (Random Image Structure Evolution) to present subjects with structured images evolving from noise. Specifically, the direction of RISE evolution – ascending or descending in information content – allows us to control for low-level image features, such as luminance, while trending towards or away from a neuron’s experimentally defined preferred stimulus. Any difference in response to ascending and descending stimuli thus reflects prior knowledge facilitating visual recognition in noise. Expanding on our previous work, subjects were presented RISE stimuli in an MRI scanner. Recent machine-learning techniques allowed us to classify the BOLD images (predicting the current visual stimulus) while subjects performed the same recognition task in the scanner. Classification was better for degraded images when subjects had prior knowledge of the coherent stimuli. We show that this hysteresis in machine classification matches the subject’s behavior, indicating more information in the brain with prior knowledge of the stimulus. Furthermore, we can localize this information by showing a graded hysteresis response from V1 through prefrontal cortex, suggesting that prior knowledge affects lower and higher visual areas in different ways. We also compare the location of hysteresis with respect to the kinds of stimuli used, to elucidate what types of image features are facilitated with prior knowledge and which are strictly processed in a feed-forward manner.

Supported by EY07023 to MS.
Jim and Marilyn Simons Fund to PS.
NEI R21EY015521-01 to PS.