Xin Huang, Ph.D.
Associate Professor, Department of Neuroscience
University of Wisconsin-Madison
Natural Scene Statistics and Neural Coding of Multiple Visual Stimuli
Natural scenes are often complex and contain multiple entities. Visual segmentation refers to the processes of partitioning visual scenes into distinct objects and surfaces, as well as segregating figural objects from their background. Segmentation is crucial for scene interpretation, object recognition, and visually guided action. However, it is still unclear how the brain represents and segregates multiple stimuli.
We hypothesize that the visual system exploits statistical regularities in natural scenes to represent multiple stimuli and facilitate segmentation. To test this hypothesis, we characterized natural scene statistics of motion and depth pertinent to visual segmentation, as these cues are potent for segmentation. We found that the figural region tended to move faster and more coherently and tended to be nearer in depth than the background region. In neurophysiological experiments, we recorded the activities of neurons in cortical area MT, a crucial hub for processing visual motion and depth information. We found that the responses of MT neurons to multiple stimuli within their receptive fields tended to be biased toward the stimulus component that moved at a faster speed, more coherently, and at a nearer depth.
Previous theoretical studies suggest that mixing multiple stimuli with different weights (rather than equal weights) can enhance the ability to encode multiple stimuli in neuronal populations. Our neural results revealed that MT neurons indeed incorporate this strategy, but in an interesting way: the response biases and hence the response weights for different stimuli reflect optimization for performing behavioral tasks such as figure-ground segregation, given our measured natural scene statistics related to the figure and ground regions. Together, these results enrich our understanding of neural representation and segregation of multiple visual stimuli, as well as demonstrate that neural coding can be optimized to perform essential behavioral tasks, in contrast to preserving information and efficiently using resources.