[edit]

AISTATS-97 Tutorial Program

NOTE: we have changed the schedule for the Tutorial Program as of November 7th: due to demand we have rescheduled so that attendees can attend both the Dawid and Jordan tutorials. The Mitchell and West tutorials will still be in parallel. Tutorial fees remain the same: attendees now get to attend 3 tutorials instead of just 2 with no change in cost.

General Information

• Each tutorial will last about 3 hours.
• Your tutorial attendance fee (see the registration form and instructions for registering ) covers attendance at each of the Dawid and Jordan tutorials and one of the Mitchell and West tutorials (which will be held in parallel), plus a copy of notes from the tutorial speakers for all 4 tutorials.
• Date: January 4th 1996
• Location: Radisson Bahia Mar Beach Resort

Tutorial A: 8:30 to 11:30am

Conditional Independence for Statistics and AI
A. P. Dawid, University College London

The axiomatic theory of Conditional Independence provides a general language for formulating and determining questions relating to the intuitive idea of "relevance", in a wide variety of different contexts. This tutorial will describe the basic theory and various interesting models of it, with special emphasis on its use in conjunction with modular graphical representations of problems in Probability, Statistics and Expert Systems.

Tutorial B: 12:30 to 3:30pm

Bayesian Time Series Analysis and Forecasting
Mike West, Duke University

This tutorial will overview the historical development and current status of Bayesian approaches to time series modelling and analysis. Particular emphasis will be given to the conceptual bases and foundations in: time-varying parameter models, sequential modelling and adaptive learning, interventionist ideas and tools, and component model structuring. Developments of specific model classes will be given with illustrative applications, leading through standard dynamic linear models, non-linear and mixture models, multivariate models, and others. This will be complemented with discussion of recent developments, especially in computation and simulation, and current research frontiers.

Tutorial C: 12:30 to 3:30pm

Learning in Information Agents
Tom M. Mitchell, Carnegie Mellon University

What kind of software agents should we create for our workstations and the Internet over the next few years, and what role should be played by AI and statistical methods in these systems? This tutorial will examine some recent examples of software agents that learn from and about users. For example, we will cover a newsreader that automatically learns users' reading interests, and a web browser that learns which hyperlinks to suggest based on user interests. Underlying these applications are interesting new research problems for machine learning and statistics, such as how best to learn when the data consists of text, images, and other unstructured data, how to learn from multiple sources of information across multiple users, and how best to adapt to changing environments and users interests. This tutorial will be part overview of known approaches, and part discussion of open research issues.

Tutorial D: 4:00 to 7:00pm

Graphical models, neural networks and machine learning algorithms
Michael Jordan, MIT

What are the commonalities between graphical models, neural networks and other network-based statistical and machine learning methods? More importantly, what are the strengths of the ideas developed thus far by the various research communities that should be retained as we consider a more unified methodology? I will present a tutorial on probabilistic learning systems that aims at a unified perspective on network-based modeling. I emphasize graphical models as providing the basic formalism, but I will also emphasize the nonparametric and (particularly) semiparametric methods characteristic of the neural network literature. Examples discussed will include Bayesian belief networks with logistic or noisy-OR nodes, Hidden Markov models (including several ``intractable'' variations of HMM's), mixture models, probabilistic decision trees, Helmholtz machines and variations on Kalman filters and Markov random fields. I will provide a detailed discussion of algorithms for inference and learning in these models, including exact probabilistic calculations, sampling methods and mean field methods.