Seminar in Machine Learning (600.735)

Instructor: John W. Sheppard

Dr. John Sheppard is an Assistant Research Professor in the Department of Computer Science at Johns Hopkins University. Recently, he was elected as an IEEE Fellow "for contributions to system-level diagnosis and prognosis." Prior to joining Hopkins, he was a Fellow at ARINC Incorporated in Annapolis, MD where he worked for almost 20 years. Dr. Sheppard is the founding director of the Numerical Intelligent Systems Laboratory at Hopkins where he performs research in Bayesian classification, factorial hidden Markov models, recurrent neural networks, and reinforcement learning. In addition, Dr. Sheppard is active in IEEE Standards activities. Currently, he serves as a member of the IEEE Computer Society Standards Activities Board and is the Vice Chair of IEEE Standards Coordinating Committee 20 on Test and Diagnosis for Electronic Systems. He has served as co-chair of the Diagnostic and Maintenance Control Subcommittee of SCC20 and as an official US delegate to the International Electrotechnical Commission's Technical Committee 93 on Design Automation.

Course Description

This seminar course will look at currrent research in machine learning. Topics will be selected from those of mutual interest between students and the instructor. Sample topics include reinforcement learning, kernel methods, experimental methods in machine learning, computational learning theory, lazy learning, evolutionary computation, and neural networks. Students are expected to select papers and lead discussion.

Spring 2008 Topic

The machine learning topic of interest in the Spring 2008 semester will focus on learning in graphical models. The seminar will examine material including lectures, draft textbooks, and papers that have considered different representations for graphical models, inference techniques using graphical models, and parameter/structure learning in graphical models. While the most common form of graphical model is the Bayesian network, we plan to consider other types of networks such as dynamic Bayesian networks, factor graphs, Markov random fields, decision networks, etc.

Schedule

The current schedule has us meeting 9:00-9:50 on Thursdays in the NEB CS Conference Room. The following lists the schedule, based on that time, of when students will lead discussion. The schedule will be adapted based on the desires of those participating in the class. As papers are assigned and made available, they will be included here for download (password protected).

Graphical Models References

  • J. Pearl, Probabilistic Reasoning in Intelligent Systems, Morgan Kaufman, 1988. (This is the book that started it all. It provides the foundation of Bayesian networks and the inference algorithms applied to them.)
  • F. Jensen, Bayesian Networks and Decision Graphs, 2nd edition, Springer, 2007. (This has become a standard text in Bayesian network courses and introduces traditional directed graphical models, i.e., Bayesian nets, and decision networks. It was recently updated and greatly improved.)
  • R. Neopolitan, Learning Bayesian Networks, Prentice-Hall, 2003. (One of the first texts focusing specifically on machine learning and Bayesian networks. The book also provides detailed explanations and walkthroughs of various inference algorithms for Bayesian networks.)
  • M. Jordan, Learning in Graphical Models, The MIT Press, 1999. (A collection of papers and authored chapters by various experts in graphical models, this book focuses on machine learning issues but stretches beyond the traditional Bayesian networks formulation. It has become a fundamental reference for machine learning and graphical models.)
  • D. Koller and N. Friedman, Structured Probabilistic Models: Principles and Techniques, unpublished manuscript, 2007. (Sadly, this book is not yet available to the general public. It is a new textbook currently under development and available for preview only by faculty teaching courses in graphical models and Bayesian networks. Excerpts can be made available to students in qualified classes under strict, "no-distribution" rules. Once published, this book will most like become the standard reference for graphical models in the field.)

Machine Learning References

  • S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 2nd edition, Prentice-Hall, 2003. (This is an excellent reference for basic artificial intelligence and provides a lot of information introducing machine learning as well. For this class, I would recommend that chapters on decision/utility theory (16), Markov decision processes (17), and reinforcement learning (21). There is a little discussion of genetic algorithms in chapter 4.)
  • T. Mitchell, Machine Learning, McGraw-Hill, 1997. (This has become the "standard" textbook on machine learning and provides chapters on genetic algorithms (9) as well as reinforcement learning (13). While a bit dated, the book is still excellent.)
  • E. Alpaydin, Introduction to Machine Learning, The MIT Press, 2004. (This is the newest textbook on machine learning, but I am not particularly excited by it. I offer it up as a more recent resource if the date of Mitchell's text is a concern. This book tends to combine elements of machine learning from a traditional AI perspective with machine learning from the statistical pattern recognition perspective. It does have a chapter on reinforcement learning, but none on genetic algorithms.)
  • V. Cherkassky and F. Mulier, Learning from Data: Concepts, Theory, and Methods, Wiley Interscience, 1998. (While also a bit dated, I really like this book. Similar to the Alpaydin text, it approaches machine learning from a statistical point of view, but does so in both a rigorous and lucid manner.)
  • R. Duda, P. Hart, and D. Stork, Pattern Classification, Wiley Interscience, 2001. (This is an update to the classic "Duda and Hart" text titled Pattern Classification and Scene Analysis from 1973. Written from the perspective of statistical pattern recognition, this book became the standard for machine learning when it first formed into its own discipline. It provides good descriptions of neural networks, and even has a little bit on evolutionary algorithms.)
  • S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice-Hall, 1999. (This provides a solid, mathematical introduction to neural networks with some interesting chapters on radial basis function networks (5) and neurodynamic programming (12), as well as all the traditional ANN topics.)
  • K. DeJong, Evolutionary Computation: A Unified Approach, The MIT Press, 2006. (This is a brand new book that looks at the main topics of evolutionary computation from a consistent, unifying point of view. It is very readable and covers all of the main algorithms, including genetic algorithms, evolutionary programming, and evolution strategies. It treats genetic programming as a specialization of the genetic algorithm, so no special treatment of GP is provided.)
  • M. Mitchell, An Introduction to Genetic Algorithms, The MIT Press, 1996. (This provides a nice introduction and overview to standard genetic algorithms. Similar to other MIT Press books, it is small but well written. It is a bit dated, but given the overview nature of the book, the material is still relevant.)
  • R. Sutton and A. Barto, Reinforcement Learning: An Introduction, The MIT Press, 1998. (As far as I know, this is the only book dedicated to reinforcement learning. There is nothing specific to evolutionary or connectionist techniques in the book, except or a chapter on function approximation, but it still provides a good overview. A version of the book is available online at http://www.cs.ualberta.ca/%7Esutton/book/ebook/the-book.html.)