|
|
MIT Electrical Engineering and Computer Science
EECS Event |
Wednesday, April 25, 2001
1:15 PM (refreshments 1:00)
Grier Room, Room 34-401B
EECS Special Seminar
Abstract
Title: Signal Processing in Non-Stationary Environments
Abstract: Statistical modeling has provided the foundation for signal detection, filtering and identification from noisy, distorted and incomplete information. Adaptive processing, wherein model parameters are adapted based on processed information has been successfully employed for unknown and slowly time-varying environments. However, the underlying procedure here of iterative adaptation and processing breaks down for unknown, dynamic and highly non-stationary environments, a problem that gains significance in a wide range of applications such as mobile communications, noise reduction and underwater acoustics.
As I will show, the rate of variation fundamentally limits the model resolution and reliability, which in turn impacts the overall performance. Consequently, I am led to the question of how to robustly extract relevant information from finite noisy data to maximize performance. To address this question, I will present a new principle for statistical modeling & adaptation, wherein I recursively adapt in a smaller dimensional space of models in such a way that the estimates are guaranteed (with high confidence) to remain at an ìoptimallyî achievable distance (model-resolution) from the unknown, underlying process. I will show that the rate of variation induces a fundamental tradeoff between model-complexity, resolution, confidence level, and performance. I will motivate this evolution of robust signal processing in the context of the development of the cabin-communication system and discuss other potential applications.