Neil J. Bershad
Professor Emeritus, Electrical Engineering and Computer Science
The Henry Samueli School of Engineering
The Henry Samueli School of Engineering
PH.D., Rensselaer Polytechnic Institute
University of California, Irvine
Engineering Tower 416C
Mail Code: 2625
Irvine, CA 92697
Engineering Tower 416C
Mail Code: 2625
Irvine, CA 92697
Research Interests
Communication and Information Theory, Digital Signal Processing
Academic Distinctions
Appointments
Research Abstract
Analyses of the Stochastic Behavior of Adaptive Signal Processing Algorithms
Investigator: N.J. Bershad
The least mean squares (LMS) algorithm has found wide application in situations when the statistics of the input processes are unknown, partially known, or time-varying. These include such applications as noise cancellation, line enhancement, adaptive echo cancellation, unknown channel modeling and identification, and spatial processing such as adaptive beamforming and side-lobe cancellation. The basic algorithm is described by discrete-time continuous-valued variables. When the LMS algorithm is implemented digitally, significant and non-trivial differences in the behavior of the algorithm appear. The researcher has developed analytical models that predict accurately the differences in behavior between the analog and digital implementations.
The primary reasons for the widespread use of the LMS adaptive algorithm is its simplicity of implementation and relatively well-understood behavior as compared to other adaptive algorithms. However, LMS transient behavior is slower than more-complicated algorithms such as recursive least squares (RLS). Therefore, in applications where algorithm convergence speed is critical, the RLS algorithm often is suggested as a better alternative to LMS. On the other hand, there are many practical applications, such as tracking, where fast algorithm convergence does not necessarily imply good performance. Comparison of the tracking performance of LMS and RLS for several important problems has shown that LMS often performs better than RLS.
Stochastic Modeling and Analysis of Neural Networks
Investigator: N.J. Bershad
Research Assistant: J. Vaughn
Neural networks are defined by their architecture and training rules. However, they also can be viewed as generalized non-linear adaptive filters. Stochastic models, originally developed for studying adaptive filters, are being investigated and applied to analyzing the steady-state and transient behavior of single- and multi-layer neural networks. In particular, networks that use the Rosenblatt and the back-propagation training algorithms already have been analyzed successfully.
Investigator: N.J. Bershad
The least mean squares (LMS) algorithm has found wide application in situations when the statistics of the input processes are unknown, partially known, or time-varying. These include such applications as noise cancellation, line enhancement, adaptive echo cancellation, unknown channel modeling and identification, and spatial processing such as adaptive beamforming and side-lobe cancellation. The basic algorithm is described by discrete-time continuous-valued variables. When the LMS algorithm is implemented digitally, significant and non-trivial differences in the behavior of the algorithm appear. The researcher has developed analytical models that predict accurately the differences in behavior between the analog and digital implementations.
The primary reasons for the widespread use of the LMS adaptive algorithm is its simplicity of implementation and relatively well-understood behavior as compared to other adaptive algorithms. However, LMS transient behavior is slower than more-complicated algorithms such as recursive least squares (RLS). Therefore, in applications where algorithm convergence speed is critical, the RLS algorithm often is suggested as a better alternative to LMS. On the other hand, there are many practical applications, such as tracking, where fast algorithm convergence does not necessarily imply good performance. Comparison of the tracking performance of LMS and RLS for several important problems has shown that LMS often performs better than RLS.
Stochastic Modeling and Analysis of Neural Networks
Investigator: N.J. Bershad
Research Assistant: J. Vaughn
Neural networks are defined by their architecture and training rules. However, they also can be viewed as generalized non-linear adaptive filters. Stochastic models, originally developed for studying adaptive filters, are being investigated and applied to analyzing the steady-state and transient behavior of single- and multi-layer neural networks. In particular, networks that use the Rosenblatt and the back-propagation training algorithms already have been analyzed successfully.
Link to this profile
https://faculty.uci.edu/profile/?facultyId=2321
https://faculty.uci.edu/profile/?facultyId=2321
Last updated
02/22/2002
02/22/2002