Communications systems: Research in the area of communication systems is directed at problems of equalization, modulation, and coding, aimed in particular at aspects of wireless systems. One area of interest has been blind equalization and equal
ization of nonlinear channels using nonlinear signal processing techniques. Additionally, nonlinear receiver algorithms for interference suppression are of interest. For wireless channels, work in progress is concerned with new combinations of diversit
y, modulation, and coding techniques that yield good performance under severe fading and multipath conditions. Also of interest is the development of protocols for multiple-access in wireless systems. Another area that is of interest is that of video tr
ansmission over wireless channels.
Information theory, cryptography, combinatorics: Information-theoretic ideas and methods are being applied in a wide variety of problems including: coin-tossing games with applications to cryptography, finite-memory effects in storing data, enume
ration of folds in sequences (a biologically inspired problem with applications to RNA secondary structure), reconstruction of binary sequences transmitted over a binary erasure channel, and the determination of the fine structure of threshold functions
in random graphs.
Randomized algorithms: Learning in discrete structures, such as weight-constrained neural networks, may be intractable in the worst-case, but nevertheless, be tractable on average. The focus is on the development of efficient randomized procedure
s which reduce the average time complexity of learning.
Pattern recognition: Classical paradigms in pattern recognition use labeled examples with data labeled according to the pattern class of origin. While many learning procedures are provably efficient in the limit of an infinite number of examples, there i
s a relative paucity of results on their finite sample performance. Recent collaborative work with faculty at the University of Vermont and the California Institute of Technology has focused on the classical nearest neighbor algorithm for pattern classif
ication to characterize the intrinsic information-theoretic limitations caused by finite sample effects.
On another tack, while the classical learning paradigms utilize labeled examples, in practice, unlabeled or unidentified data are frequently more abundant and cheaper to acquire. Recent investigations have been directed toward characterizing the informat
ion content of labeled and unlabeled examples and identifying mechanisms by which expensive labeled examples can be traded off for (a large number of) inexpensive unlabeled examples. In this context, it is also sought to quantify the importance of side-in
formation to the learning process.
Learning complexity and regularization: Given a finite amount of data, how does one determine the computational model or network that generated it? Using new methods in statistics, investigations in progress seek to determine how to optimally
fit models to data and to determine the optimal duration of training.
Applications: Learning theory has a variety of applications in pattern classification, game theory, decision theory, finance, and econometrics. A typical application under investigation in finance casts the auditing and accounting problems in a
learning setting and seeks to provide a definitive answer to the question: Should auditors be responsible for detecting management fraud? This problem was motivated by the recent Savings and Loans debacle.
Areas of Research
Communication Theory
Communications Systems constitute a vital and active area of research concentration within the department, motivated by ongoing exciting developments leading to innovative new communications principles and technologies. Research activities revolve around
issues pertinent to communication theory, information theory, signal processing aspects of communications systems design, and performance evaluation of communication links and networks. Signal Processing
The general area of nonlinear signal processing and nonlinear dynamical models and chaos has formed another area of research focus. Nonlinear processing of signals is required in many applications, such as equalization of nonlinear communication channels
, interference cancellation, time-series prediction, impulsive noise rejection, and edge-preserving smoothing and filtering. Neural network structures are being investigated for use in such applications. In particular, the radial basis function network
(RBFN) has been found to be very useful for nonlinear multivariate function approximation in signal processing. Current research is concerned with the characterization of the basic capabilities of such networks, development and evaluation of adaptive alg
orithms or learning techniques for network parameters, extensions of network structures for enhanced performance in specific situations such as those involving complex data, and considerations of efficient implementation. Applications such as blind equal
ization of nonlinear channels are of interest as particularly appropriate applications of such processing networks. Computational Learning Theory
The complexity of learning has attracted substantial interest in recent years and a divers set of questions in this burgeoning area are being investigated ranging from an information-theoretic characterization of the learning process to the development of
algorithms for learning in particular structures. Associated Faculty
Associated Laboratories
Return to the EE Home Page