Machine Learning and Friends Lunch

home
past talks
resources
conferences

Learning From Textbooks

Abstract

I will present some necessary limitations of statistical learning and then discuss the problem of how to acquire knowledge from the kind of presentation that one finds in a textbook.

Clark and Thornton make a critical distinction between Type-1 and Type-2 learning. In Type-1 learning, a statistical relationship, however weak, exists between the dependent variable (e.g. the class label) and the independent variables (e.g. the instance descriptors). (The great majority of the research in Machine Learning is concerned with Type-1 learning.) For Type-2 learning, there is no such statistical relationship to be found. Instead, the independent variables must be mapped first to one or more intermediate representations. A textbook embraces the apparent reality that much of human learning requires bottom-up learning of knowledge, representations, and skills that enable further learning. Such bottom-up learning produces mapped representations that make Type-2 learning possible. To achieve the next fundamental breakthroughs in the mechanization of learning, we need to explore and develop approaches to Type-2 learning.