Sensory information is detected and transformed by sensory neural networks before reaching higher levels of processing. These networks need to perform significant processing tasks while being compatible with the following levels. Lateral inhibition is a mechanism of local neuronal interaction that produces significant global properties. This book discusses those sensory neural networks influenced by nonlinear lateral inhibition. It features biological bases of lateral inhibition models, computational properties of these models that stress their short term adaptive behavior, their relation to recent activity in neural networks and connectionist systems, their use for image processing applications, and their application to motion detection. Descriptions from different technologies of analog hardware implementations of these classes of networks are described and results from implementations that corroborate theoretical analysis and show technologically desirable applications are presented. The book also uses nonlinear mathematical techniques to analyze temporal and spatial behavior of models presented within the text. Sensory Neural Networks: Lateral Inhibition is an interdisciplinary work that will prove useful to neural network theorists, biologists, circuit designers, and vision scientists.
Table of Contents
INTRODUCTION, BIOLOGICAL BASES. A BRIEF HISTORY OF LATERAL INHIBITION. CENTER-SURROUND ORGANIZATION AND RECEPTIVE FIELDS. ADAPTIVE LATERAL INHIBITION. EVIDENCE FOR MULTIPLICATIVE LATERAL INHIBITION. MULTIPLICATIVE LATERAL INHIBITION: A DERIVATION. QUANTUM ABSORPTION EVENTS AND ADAPTIVE CHANGES IN LOW-PASS FILTERING. COMPUTATIONAL PROPERTIES. RELATION TO OTHER MODELS. Additive Models. PROPERTIES RELATED TO VISION. Variable "Connection Strengths". Automatic Gain Control. Some Psychophysical Properties. Coding of the Intensity. NONLINEAR LATERAL INHIBITION AND IMAGE PROCESSING. INTRODUCTION. BACKGROUND. IMPLEMENTATION. RESULTS AND DISCUSSION. MODELING THE PROCESSING AND PERCEPTION OF VISUAL MOTION. INTRODUCTION. FEATURE MATCHING SCHEMES. INTENSITY BASED SCHEMES. Global Models. Local Models. SHUNTING LATERAL INHIBITORY MODELS. Directional Selectivity With MLINNs. Multiplicative Inhibitory Motion Detectors. Response Characteristics of MIMDs. DISCUSSION. ELECTRON REALIZATION. NEURAL NETWORK IMPLEMENTATIONS. Associative Memory Chips. Learning Chips. Sensory Neural Networks. DESIGN FRAMEWORK FOR ANALOG IMPLEMENTATION. Input Sources. Cell Body and Temporal Characteristics. Shunting Recurrent Circuitry. Shunting Non-Recurrent Circuitry. Summing Circuitry and Excitatory and Inhibitory. Inhibitory Connections. Sigmoidal Nonlinearity. Interface Circuitry. THE CHOICE OF TECHNOLOGY. Gallium Arsenide Framework. IMPLEMENTATION OF SHUNTING NETWORKS WITH FET TECHNOLOGIES. ANALOG IMPLEMENTATION. NONLINEAR DEVICE CHARACTERISTICS. IMPLEMENTATION OF NONSATURATING FETs. IMPLEMENTATION OF INTERCONNECTIONS. LIMITATIONS TO PERFORMANCE. RESULTS OF SPECIFIC IMPLEMENTATION. ACCURACY. LEVEL-SHIFTING AND THE DYNAMIC RANGE. RESPONSE TO UNIFORM INPUT. RANGE COMPRESSION AND DATA COMPRESSION. POINT-SOURCE RESPONSE. Intensity Dependence. Boundary Effects. TUNABILITY OF SENSITIVITY. SPATIAL EDGE RESPONSE. Intensity Dependence. DIRECTIONAL SELECTIVITY. DISCUSSION. NONLINEAR MATHEMATICAL DESCRIPTION. STABILITY AND CONTENT ADDRESSABILITY. Stability of the Implemented Network. RELATION TO ADAPTIVE RESONANCE. VOLTERRA-WIENER SERIES EXPANSION. TEMPORAL KERNELS. First Order Temporal Kernel. Second Order Temporal Kernel. Discussion. SPATIAL KERNELS. Zero Order Kernel. First Order Spatial Kernel. Second Order Spatial Kernel. Higher Order Spatial Kernels. CLASSIFICATION PROPERTIES. CONCLUSIONS. BIBLIOGRAPHY.
Professor Bahram Nabet has been a senior visiting scientist in Brazil and Italy, ASEE Summer Faculty Fellow at the Naval Research Lab in Washington DC, and has been on sabbatical leave at Telcordia Applied Research Inc. and Dexxon Group Inc.