Machine Learning and Friends Lunch

home
past talks
resources
conferences

... But What About The Structure?

Abstract

For 30 years, structure was at the heart of AI, but in the '80s the
emphasis shifted to parameter estimation. Structure is becoming
popular again as efforts focus on model selection and the minimum
description length principle. In this talk I argue that ignoring
structure is the main cause of the curse of dimensionality, which
plagues popular machine learning approaches such as Support Vector
Machines (SVMs) and Reinforcement Learning (RL). This problem can be
dealt a mortal blow by embracing the optimization of structure in
parallel with the optimization of parameters.

An optimization criterion is described, along with a straight forward
search methodology. Encouraging results in computer vision for a
mobile robot are presented. The methodology bears resemblance to
certain psychological theories of learning, and exhibits some well
known psychological phenomena.