Richard H. Tsai

A Vlsi Neural Network Processor Based on Hippocampal Model

The hippocampal region of the brain system can be analyzed with the nonlinear system modeling approach. The input-output relationship of the neural units is best represented by the kernel functions of different complexities. The modeling expression of the first and second order kernels are computed in analog current-mode instead of digital data processing in order to fully explore massively parallel processing capability of the neural networks. Two distinct methods are utilized: the table-look-up approach and the model-based approach. The former can achieve high accuracy but consumes large silicon area while the latter saves silicon area and maintains moderately high accuracy. Circuit-level simulation results and experimental data from two test structures are presented.

A state-constrained model for cellular nonlinear network optimization

A G/sub m/C-style state constrained neuron (SCN) model for the design of processors in analog recurrent neural networks such as Hopfield neural networks, cellular nonlinear networks for combinatorial optimization is described. The unconstrained neurons which have the free state variable, could be stable at any arbitrary point in the solution space or trapped by un-intentional effects. These may introduce errors. For the unconstrained network, the solution could be different from the expected one due to the discrepancy in the energy function of the network and the objective function to be optimized. In addition, if the state variable is limited by some neighboring saturated transistors, un-desirable results may be obtained. The G/sub m/C-style SCN model can ensure the convergence of the network and avoid discrepancy between the energy function of the network and the objective function. The state resistor is also eliminated in the G/sub m/C model so that high cell-density can be achieved. Simulation results show that the proposed model is effective in significantly reducing optimization error.

A hippocampal model implementation using VLSI table-look-up and model-based approaches

The hippocampus region of the brain system performs cognitive functions of learning and memory. VLSI design of a hippocampal model has been proposed and two mixed analog-digital chips which use analog circuits for parallel computing and digital circuits for interconnection were designed. The design with one extensive hippocampal neuron has been sent for fabrication. One of the objectives of this research effort is to incorporate adequate biological constraints into the microelectronic chips and systems.

One step closer to a bionic brain

Contribution toward a bionic brain is no longer an impossible dream. Several research teams are making attempts in the right direction all over the world. Analog VLSI design of a hippocampal neuron model is an important step toward construction of more complex microelectronic devices to facilitate the study of important brain regions. Gradually, selected functions of brain regions can be replaced by the bionic chips. The microchips will not only cure the illness of the brain but also enhance its ability.

Design of a programmable pulse-coded neural processor for hippocampal region

Rapid advances in submicron microelectronic technologies have made possible the integration of millions of transistors on a single silicon chip for complex signal processing and control functions. The hippocampal region of the brain system can be analyzed with a nonlinear system modeling approach. The input-output relationship of the neural units is represented by the kernel functions of various complexities. The modeling expressions of the first and second order kernels are computed in analog current-mode instead of digital format in order to fully explore the massively parallel processing capability of the neural networks. A programmable pulse-coded neural network based on the hippocampal kernel functions can process the pulse information efficiently. The model-based approach saves silicon area and achieves adequate accuracy level. Circuit-level simulation results and experimental data are also presented.

Advances in efficient optical links to enhance desktop multimedia processor systems

Rapid progresses in computers, telecommunication, and consumer electronics research have stimulated the emergence of powerful multimedia systems that will have profound impacts, on scientific, industrial, business endeavors, and even our daily lives. Continuous advances in deep-submicron microelectronic fabrication technologies have made possible the integration of millions of transistors on a single silicon microchip. However, one of the anticipated major technological breakthroughs will be the efficient and reliable use of optical links to increase the communication bandwidth and quality of electronic systems. Optical links can be used not only at the machine-to-machine level, but also at the board-to-board and chip-to-chip levels. The architecture, design, and associated measurement results are presented regarding the use of optical interconnections to enhance desktop multimedia computing systems.

VLSI neural network implementation of a hippocampal model

As an important portion of the brain system, the hippocampus performs cognitive functions of learning and memory. VLSI implementation of a hippocampal model has been proposed and a mixed analog-digital chip which uses analog circuits for parallel computing and digital circuit for interconnection was designed. The design with one extensive hippocampal neuron has been sent for fabrication. One of the objectives of this research effort is to incorporate adequate biological constraints as possible into the hardware design.

A programmable neural processor for pulse-coded hippocampal signal

Rapid advances in submicron microelectronic technologies have made possible the integration of millions of transistors on a single silicon chip for complex signal processing and control functions. The hippocampal region of the brain system can be analyzed with a nonlinear system modeling approach. The input-output relationship of the neural units is represented by the kernel functions of various complexities. The modeling expressions of the first and second order kernels are computed in analog current-mode instead of digital format in order to fully explore massively parallel processing capability of the neural networks. A programmable pulse-coded neural network based on the hippocampal kernel functions can process the pulse information efficiently. The model-based approach saves silicon area and achieves adequate accuracy level. Circuit-level simulation results and experimental data are also presented.

ContentBased Image RetrievalResearch Issues 1995, NEC Research Institute, Inc. All rights reserved.

This chapter surveys the research problems and directions in the field of image retrieval by content-based query. The central problem is to find high-speed methods for finding which images within a large collection best match a given image template. To be of practical use, a solution should be able to do detailed analyses at a rate of about 1000 images per second, which, when coupled with descriptor-based search techniques to select candidates for analysis, will enable effective search of libraries containing millions of images. An attractive approach is to perform the search on compressed versions of the images. Reduction in data volume reduces I/O time, and should reduce processing time as well. Fourier-domain compression and wavelet compression are the major directions for this research. The advantage of Fourier-domain compression is that it provides an efficient computational means to determine which position within an image best matches a specified pattern. However, the cost of performing Fourier transforms is a computational bottleneck, forcing researchers to consider other alternatives that may yield faster searches. Wavelet compression has the advantage that correlation peaks can be detected in the wavelet domain, thus eliminating the need to invert the wavelet transform. But searching within a wavelet transform is more costly than in a Fourier domain. One might obtain the best of both transforms by combining Fourier transforms with wavelet functions. For both wavelet and Fourier-domain computations, architectural support can reduce access-time delays to near zero on critical paths within the memory hierarchy because all of the accesses are known in advance. Memory-access time is the main source of performance degradation in the memory hierarchy. It may be possible to use the knowledge of future reference patterns to schedule dat a movement at various levels of the memory hierarchy in order to reduce the access time delay to zero for the critical accesses. Ideally, image streams will arrive at an arithmetic unit just in time to be processed, and the results will be returned to an external disk (or other form of memory) just as the write head reaches the region where the data are to be written. The gains from both architectural support and processing of compressed images may yield the performance improvement required to meet our goal.

Intelligent Speech/Audio Processing for Multimedia Applications

Intelligent speech and audio processing can provide efficient and smart interfaces for various multimedia applications. Generally, speech is the most natural form of human communication. Audio and music can enhance our emotional impacts and promote interest in multimedia applications. A successful interactive multimedia system must have the capabilities of speech and audio compression, text-to-speech conversion, speech understanding, and music synthesis. The main purpose of speech and audio compression is to provide cost-effective storage or to minimize transmission costs. Text-to-speech converts linguistic information stored as data or text into speech for the applications of talking terminals, alarm systems, and audiotext services. Speech understanding systems make it possible for people to interact with computers using human speech. Its success relies on the integration of a wide variety of speech technologies, including acoustic, lexical, syntactic, semantic, and pragmatic analyses. The applications of music processing for multimedia were mostly realized by means of the combination of music, graphics, video, and other media. Since musical sounds and compositions can be precisely specified and controlled by a computer, we can easily create artificial orchestras, performers, and composers. Nowadays, multimedia systems have become more sophisticated with the advances made in computer and microelectronic technologies. Many applications require efficient processing of speech and audio for interactive presentations and integration with other types of media. The application-specific hardwares are proposed to meet the high-speed, low-cost, lightweight, and low-power requirements. The design example of a speech recognition processor and system for voice-control applications is introduced. The industrial standards and commercial products of speech and audio processing ar e also summarized in this chapter.