Natural scene statistics and contextual influences in vision
Prof Odelia Schwartz
Associate Professor, Computer Science
University of Miami
An appealing hypothesis suggests that neurons represent inputs in a coordinate system that is matched to the statistical structure of images in the natural environment. I discuss theoretical work on unsupervised learning of statistical regularities in natural images. In the model, Bayesian inference amounts to a generalized form of divisive normalization, a canonical computation that has been implicated in many neural areas. In our framework, divisive normalization is flexible: it is recruited only when the image is inferred to contain dependencies, and muted otherwise. I focus on recent work in which we have applied this approach to understanding spatial context effects in visual cortical processing of natural inputs. I also discuss our work in progress on understanding secondary visual cortex and integrating this approach with deep convolutional neural networks.
10 Crichton St
11:00 – 12:30
Lunch will be served in MF2 (4.04) after the seminar. If you plan to attend, and need any diet needs considered, then please email email@example.com.