Machine Learning and Friends Lunch
Learning Many Layers of Structure
David Stracuzzi
UMass
Abstract
Representational structure is an important aspect of machine learning. All learning algorithms are influenced by the structure of their underlying representation, yet very few algorithms make even the slightest attempt at structure learning. Structure is assumed to be provided. In this talk, I will motivate the importance of structure learning and discuss the "limited learning" assumption as a replacement for provided structure. This change is fundamental to our work and I will underscore the effects. Finally I will discuss two structure learning algorithms based on limited learning, along with several key issues, including victories and frustrations.