Learning To Learn
Sebastian Thrun and Lorien Y. Pratt
Kluwer Academic Publishers
Over the past three decades, research on machine learning and data mining has led to a wide variety of algorithms that induce general functions from examples. As machine learning is maturing, it has begun to make the successful transition from academic research to various practical applications. Generic techniques such as decision trees and artificial neural networks, for example, are now being used in various commercial and industrial applications.
Learning to learn is an exciting new research direction within machine learning. Similar to traditional machine learning algorithms, the methods described in LEARNING TO LEARN induce general functions from experience. However, the book investigates algorithms that can change the way they generalize, i.e., practice the last of learning itself, and improve on it.
To illustrate the utility of learning to learn, it is worthwhile to compare machine learning to human learning. Humans encounter a continual stream of learning tasks. They do not just learn concepts of motor skills, they also learn bias, i.e., they learn how to generalize. As a result, humans are often able to generalize correctly from extremely few examples---often just a single example suffices to teach us a new thing.
A deeper understanding of computer programs that improve their ability to learn can have large practical impact on the field of machine learning and beyond. In recent years, the field has made significant progress towards a theory of learning to learn along with practical new algorithms, some of which led to impressive results in real-world applications.
LEARNING TO LEARN provides a survey of some of the most exciting new research approaches, written by leading researchers in the field. Its objective is to investigate the utility and feasibility of computer programs that can learn how to learn, both from a practical and a theoretical point of view.
This book is organized into four parts:
- Part I: Overview articles (chapter 1-3), in which basic taxonomies and the cognitive foundations for algorithms that ``learn to learn'' are introduced and discussed,
- Part II: Prediction/Supervised Learning (chapter 4-8), in which specific algorithms are presented that exploit information in multiple learning tasks in the context of supervised learning,
- Part III: Relatedness (chapter 9-10), in which the issue of ``task relatedness'' is investigated and algorithms are described that selectively transfer knowledge across learning tasks, and
- Part IV: Control (chapter 11-13), in which algorithms specifically designed for learning mappings from percepts to actions are presented.
LEARNING TO LEARN features contributions by the following authors:
- Learning To Learn: Introduction and Overview, by Sebastian Thrun and Lorien Pratt
- A Survey of Connectionist Network Reuse Through Transfer, by Lorien Pratt and Barbara Jennings
- Transfer in Cognition, by Anthony Robins
- Theoretical Models of Learning to Learn, by Jonathan Baxter
- Multitask Learning, by Rich Caruana
- Making a Low-Dimensional Representation Suitable for Diverse Tasks, by Nathan Intrator and Shimon Edelman
- The Canonical Distortion Measure for Vector Quantization and Function Approximation, by Jonathan Baxter
- Lifelong Learning Algorithms, by Sebastian Thrun
- The Parallel Transfer of Task Knowledge Using Dynamic Learning Rates , by Daniel L. Silver and Robert E. Mercer
- Clustering Learning Tasks and the Selective Cross-Task Transfer of Knowledge, by Sebastian Thrun and Joseph O'Sullivan
- CHILD: A First Step Towards Continual Learning, by Mark B. Ring
- Reinforcement Learning With Self-Modifying Policies, by Juergen Schmidhuber, Jieyu Zhao, Nicol N. Schraudolph
- Creating Advice-Taking Reinforcement Learners, by Richard Maclin and Jude W. Shavlik