 |
|
|
|
|
|
|
|
|
|
|
|
|
 |
|
|||||||||||
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
|
 |
 |
|
||||||||||
 |
|
|||||||||||
 |
|
|||||||||||
 |
 |
|
||||||||||
| Seminar Schedule (Seminar Organizer: Prof. Ziv Bar-Joseph) | |
Machine Learning Lunchtime Chats
|
Fall 2004 Seminar
Date: December 3, 2004 Time: 12:00 PM - 1:00 PM (Pizza, while it lasts!) Location: 4623 Wean Hall Speaker: Anthony Brockwell Assistant Professor, Statistics Title: Generalizing the Hidden Markov Model Abstract: The hidden Markov model has been used heavily in a number of areas, including speech recognition and molecular biology, to name only two. The fundamental idea is to include two components in the model, a hidden process, and a sequence of observations which depend (probabilistically) on corresponding elements of the hidden process. While very useful, the model is restrictive in the sense that it requires the hidden process, and usually the observations as well, to take values in a discrete space. We discuss a generalization of the hidden Markov model known as a "generalized state-space model". Although these models have existed in the literature for some time, they are not widely used, due to a lack of associated algorithms. (The recently-developed "particle filter" is an exception, but has certain practical drawbacks.) To address this problem, we develop a simulation-based estimator of the likelihood for generalized state-space models. We then examine results in two case studies: one where we model behavior of a neuron in a monkey's motor cortex, and the other where we fit a model to a time series of share prices.