Visual Computing Seminar: Learning a distance measure from the information-estimation geometry of data

Abstract:

The perceptual distance between images is widely believed to be related to the distribution of natural images. But how can a probability distribution give rise to a distance measure—let alone one that aligns with human perception? What properties should such a distance satisfy, and how can it be learned from an image database in an unsupervised manner? In this talk, I will address these questions by presenting the Information–Estimation Metric (IEM), a novel form of distance function derived from a given probability density over a domain of signals. The IEM is rooted in a fundamental relationship between information theory and estimation theory, which links the log-probability of a signal with the errors of an optimal denoiser, applied to noisy observations of the signal. For Gaussian-distributed signals, the IEM coincides with the Mahalanobis distance. But for more complex distributions, it adapts, both locally and globally, to the geometry of the distribution. I will discuss and illustrate the theoretical properties of the IEM—including its global and local behavior. Finally, I will demonstrate that the IEM effectively predicts human perceptual judgments when trained (unsupervised) on natural images.

Bio:

Guy is a postdoctoral researcher working with Eero Simoncelli at the Flatiron Institute. His research focuses on developing computational models of human perception that are grounded in principles from information theory. He received his PhD in Computer Science from the Technion—Israel Institute of Technology, where he worked with Michael Elad and Tomer Michaeli on the design and theoretical analysis of image restoration and compression methods that rely on generative models.