Dr. Vishnu Boddeti, Assistant Professor, Computer Science Department, Michigan State University gave a talk on The Capacity and Intrinsic Dimensionality of Image Representations on 10 August. His research areas are Computer Vision, Pattern Recognition and Machine Learning.

Image representations are functions of images that enable us to distill the information present within them and are useful for prediction tasks. Significant amount of effort has been devoted to designing effective image representations over the past few decades, culminating in the current day paradigm of data-driven representation learning. Indeed, automated image recognition systems are now believed to surpass human performance in some scenarios. Despite this tremendous empirical progress, many crucial questions about image representations still remain unanswered, (1) What is the minimal number of degrees of freedom or intrinsic dimensionality of a given image representation?, (2) What is the capacity or maximal number of categories that can be fully resolved by a given image representation at a given tolerable error metric? A scientific basis for answering these questions will not only benefit the evaluation and comparison of different image representation methods, but will also establish bounds on the compactness and scalability of a given automatic image recognition system.

In his  talk Dr. Boddeti presented his attempts to quantitatively answer these questions using concepts from information theory and then discuss their implications.