Master Thesis

An Information Theoretic Feature Learning Approach to Visual Recognition

For my Master’s thesis dissertation, I worked on an unsupervised feature learning approach to Visual Recognition based on information theoretic principles.

Rather than the mere ability to identify and localize categories, places, and objects in a complex visual scene, I’ve grown to consider Visual Recognition as a journey from features to meaning, with the goal of modeling the process by which visual inputs become understanding. Indeed, at the highest peak of its achievement, Visual Recognition tries to capture the essence of the formation of human knowledge, which I found to be a very ambitious task but also an incredibly fascinating objective.

I have been fully engaged for several months in the understanding and numerical validation of a new approach to Visual Recognition (the Developmental Vision Agent, or DVA) developed in the Artificial Intelligence Research Lab of the University of Siena. Specifically, my Thesis centered on the DVA’s key idea of using deep architectures for extracting pixel-based features that can be extended to the higher layers of the learning hierarchy. The learning principles are derived using an Information Theoretic Learning approach, which allows identifying an unsupervised learning algorithm by minimizing the conditional entropy under the soft-constraint of optimizing the development of different visual features.