MURI Seminars

Seminars in October, 1996

Fixation has traditionally been thought to simplify visual motion analysis and to play a key role in vision-based motion control strategies. On the contrary, we show how fixation enhances the effects of noise in the estimates of motion and argue that guidance strategies based upon fixation result in un-natural driving behaviors.

We then go on to propose a sensor-based control strategy for driving a car along an unknown contour which allows greater flexibility in the design of the overall control behavior.

The MOBATEL-Program develops basic technologies for a system which is suggested for a number of applications in the Norwegian offshore oil exploration industry in coming years. The research of 11 doctoral candidates at NTNU explores different critical components of the system.

A key feature of Bayesian networks is that they provide a language that qualitatively represents the conditional independence properties of a probability distribution. This representation facilitates exploitation of these conditional independencies for compact specification of the distribution and for effective inference. It is well-known, however, that we cannot qualitatively capture certain independencies within the Bayesian network structure: independencies that hold only in certain contexts, i.e., given a specific assignment of values to certain variables. I will start by describing a formal notion of context-specific independence, based on regularities in the conditional probability tables (CPTs). I will present a technique for determining when such independences hold in a given network, and describe one particular qualitative representation scheme --- tree-structured CPTs --- for capturing context-based independence. I will then briefly describe how this representation of context-specific independence can be used to support effective inference algorithms, both exact and approximate.

In the second part of my talk, I will address the problem of learning Bayesian networks from raw data: A major problem in practice is the acquisition of Bayesian networks, which can be often expansive. Thus, there is a growing body of work on learning Bayesian networks from raw data, which is often readily available. I will claim that the explicit representation of context-specific independence leads to more effective learning procedures. This representation allows us to learn CPTs with a variable number of parameters, and thus, to induce models that emulate better the real complexity of the interactions observed in the data. I will describe the theoretical foundation and practical aspects of learning networks with tree-structured CPTs, as well as an empirical evaluation of the proposed method. This evaluation compares our approach with the standard learning procedure that does not represent context-specific independencies. This evaluation demonstrates that our procedure converges faster, in terms of number of training samples vs. generalization error, than the standard procedure. Our results also show that our procedure tends to learn more complex network structures (in terms of arcs), yet maintains less parameters.

This talk is based on joint work with Craig Boutilier of the University of British Columbia, Moises Goldszmidt of SRI International, and Daphne Koller of Stanford University.

Related papers:
- C. Boutilier, N. Friedman, M. Goldszmidt and D. Koller, ``Context-Specific Independence in Bayesian Networks''.

- N. Friedman and M. Goldszmidt, ``Learning Bayesian Networks with Local Structure''.

This is joint work with Attawith Sudsang and Narayan Srinivasa at the Beckman Institute.