The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-269

“Toward Microelectronics-Based Intelligent Systems Including Neural Network Understanding”

by Oscal Tzyh-Chiang Chen

August 1994

With the rapid progress of microelectronic technologies, an intelligent machine possessing skills of sensing, information processing, moving and thinking has been explored for many commercial, medical, and scientific applications. The high-speed and low-power features in VLSI design are pursued for many high-performance machines and portable devices. The medical prosthetic devices for cochlear implants and pacemakers for heart diseases are excellent examples. The various information processing schemes for medical images are also presented.

In an artificial neural network, the backpropagation learning method, quasi-Newton method, non-derivative quasi-Newton method, Gauss-Newton method, secant method and simulated Cauchy annealing method to obtain an optimized solution have been investigated. By using the quasi-Newton method, a three-layered feedforward network can successfully learn no-crossover trajectories. In a biologically-inspired neural network area, the nonlinear model of the functional properties of the hippocampal formation has been developed. The architecture of the proposed hardware implementation has a topology highly similar to the anatomical structure of the hippocampus, and the dynamical properties of its components are based on experimental characterization of individual hippocampal neurons. Recently, multimedia systems have received a lot of attention from the industry for consumer markets, publishing, education and entertainment. Data compression is an important scheme for multimedia systems to reduce data storage and transmission costs. A self-organization neural network architecture is used to implement vector quantization for image compression. A modified self-organization algorithm, which is based on the frequency-sensitive cost function and the centroid learning rule, is utilized to construct the codebooks. This self-organization method is quite efficient and can achieve near-optimal results.

A new adaptive vector quantization method based on the Gold-Washing method is also presented. The algorithm is shown to reach rate distortion function for memoryless sources. The computation power of the move-to-front vector quantizer can reach 40 billion operations per second at a system clock 100 MHz by using a 0.8 um CMOS technology. The algorithm and architecture for the Gold-Washing method can lead to the development of a high-speed image compressor with great local adaptivity, minimized complexity, and fairly good compression ratio. The work described in this dissertation has paved a way for further study of intelligent microelectronic systems.

To download the report in PDF format click here: USC-SIPI-269.pdf (20.2Mb)