Bernhard E. Boser - US grants

The proposed research is directed towards finding classification algorithms that are trained from examples and achieve good generalization performance. A new training procedure is investigated that adjusts the complexity of the classifier to the problem at hand in order to optimize the tradeoff between perfect memorization and good generalization. The new algorithm applies to a large variety of classification functions, including polynomial classifiers, two layer neural networks, or radial basis functions. Initial experiments on optical character recognition problems demonstrate both its performance relative to other learning algorithms, and its efficiency regarding time and memory requirements. Topics for investigation include the estimation of the VC dimension of the resulting classifiers, and the study of the relationship between VC dimension and generalization performance on a wide variety of problems in order to obtain a robust algorithm that performs well even when no expert knowledge in the training procedure to further improve performance will be considered. the ability of the algorithm to find atypical, misclassified, or meaningless, or misclassified patterns will be investigated to automatically clean databases.

Gray The rapid evolution of LSI and VLSI digital technologies has resulted in much wider utilization of digital techniques for signal processing and control applications than in the past. As the level of integration of such systems continues to increase, it has become more and more desirable to implement the Analog/Digital (A/D) conversion function on the VLSI control or signal processing device itself. Great progress has been made in the past ten years in the implementation of analog functions of various types of MOS LSI technology, using NMOS initially, then CMOS as that technology matured, and most recently BiCMOS technology. This research is directed towards new architectures and new approaches to monolithic A/D conversion which more closely approach the fundamental limits imposed by the technology used to implement them. Topics of emphasis include self-calibration techniques for high-speed, high resolution pipeline A/D converters, properties of parallel pipeline, or "systolic" A/D converters, micropower high speed A/D converters using charge transfer signal processing, optimum decoding algorithms and stability and scaling in oversampled A/D converters, approaches to very high speed oversampled A/D converters, and performance limits for the electromechanical sigma-delta A/D converter used for readout purposes in sensor applications.

Present medical tests are largely performed in conditioned laboratories by highly trained personel with access to stable power and a constant supply of reagents. In many situations, however, it is valuable to have point-of-analysis or point-of-care capabilities. For example, environmental factors may best be evaluated onsite to avoid transportation of samples, to provide a real-time warning, when working in third world countries, or when working in challenging and hostile environments. The long term aim of this project is to develop a paradigm for a fully integrated platform for performing complex protein assays in the field. The system consists of a programmable microfluidic sample processing device called a cellular automaton coupled to a novel Hall-effect electrical sensor. This analysis system will be developed to detect PAH-protein adducts that mutagenic and carcenogenic challenges. However, the development of this general platform for immunodetection of targets will also find wide application in biomedical research, pathogen and infectious disease detection. Our overall goal will be addressed by the following specific aims: 1. A microfluidic automaton will be developed that is a programmable microfluidic device for performing fluidic and sample preparation operations at the nanoliter scale. This device will be developed and tested to perform the operations necessary to do basic optically detected ELISA analyzes of PAH-protein adducts. 2. In parallel, the Boser group will design and fabricate a set of advanced Hall Sensor chips that are optimized for mating with the automaton and detecting the immunoassay process. 3. The next step will demonstrate the integration of the automaton sample preparation with the Hall sensor for the detection of PAH-protein adducts. Once the assays are demonstrated to be compatible, the Hall sensor will be directly coupled with the automaton to form a prototype integrated analysis system. In the final aim a fully integrated electronic interface and low-cost disposable microfluidic cartridges and instrument will be developed to perform sample preprocessing, incubation, mixing with reagents, and detection of the analyte automatically in the cartridge. This project leverages microfluidic sample preparation and state-of-the-art detection with the PAH assay developed in Project 1 of this proposal. These technologies will lead to devices and methods for early detection and better information that will lead to more accurate diagnosis and targeted treatments.

ARI-MA: Collaborative Research: High Z Materials for Nuclear Detection: Synergy of Growth, Characterization and Defect Physics for Room Temperature Devices


The objective of this research is to develop novel high density thin films for application as room temperature nuclear detection devices and to understand the basic charge collection mechanisms that would govern their performance. The approach is a synergistic coupling of crystal growth, materials characterization and device physics. The goal is to combine expertise in materials science, physics and engineering for potential breakthroughs in the application of complex oxides with high stopping power to achieve cost-effective nuclear detection capability.
The intellectual merit of the work will be the development of a fundamental framework for the identification of new candidate materials, coupled with development of the characterization tools and miniaturized electronics required for next generation devices. The work will leverage significant recent breakthroughs in the field of complex functional oxide materials with tunable electronic properties.
The broader impact will be in the contribution to national infrastructure for nuclear scanning capability and in the development of human capital. Graduate and undergraduate students will be educated in this highly interdisciplinary field and student recruitment will leverage existing programs to enhance diversity at UC Berkeley and the Naval Postgraduate School (NPS). An outreach component is also included in collaboration with the NPS Center for Homeland Defense and Security. This will provide high-quality materials to educate local and state officials, and other first responders, on the availability, limitations and potential impact of current and future technologies for nuclear threat detection.