Ansari M.S. Non-Linear Feedback Neural Networks: VLSI Implementations and Applications

New Delhi: Springer India, 2014. — 201 p. — (Studies in Computational Intelligence 508). — ISBN 978-81-322-1562-2Recurrent neural networks, which are essentially ANNs employing feedback interconnections between the neurons, were extensively explored after the concept of Lyapunov (or ‘energy’) functions, as a means of understanding the complex dynamics of such networks, was introduced by Hopfield and Tank. Their architecture, called the Hopfield Network, was implementable in hardware, and although it became very popular, many limitations like convergence to infeasible solutions and the requirement of a large number of neurons and interconnection weights were revealed when attempts were made to apply it to practical applications. These drawbacks warranted the exploration of alternative neural network architectures which are amenable to hardware realizations. The Nonlinear Synapse Neural Network (NOSYNN) has been proposed as one such alternative which alleviates the problems that plagued the Hopfield Network.
This book deals with VLSI implementations and applications of the NOSYNN type of nonlinear feedback neural networks. These networks have been shown to be better performing than their Hopfield Neural Network (HNN)-based counterparts, in the sense that their convergence to the exact solution is fast and guaranteed. This improvement in the performance is due to underlying difference in the nature of feedback between the HNN and the NOSYNN. While the HNN employs linear feedback (typically implemented using resistors), the NOSYNN employs nonlinear feedback (typically implemented using voltage-mode comparators). This difference in hardware also carries over to a difference in the dynamical properties of the two networks and makes the energy functions of the two networks vastly different. While the HNN has a quadratic form of the energy function, the NOSYNN has transcendental terms in its energy function which account for better and faster convergence characteristics.Neural Networks
Applications of Neural Networks
Hardware for Neural Networks
Outline of Contents
Organization of the ChaptersHopfield Neural Network
Nonlinear Synapse Neural Network
Chosen Problems: Description and Applications
Overview of Relevant LiteratureVoltage-Mode Neural Network for the Solution of Linear EquationsSolving Linear Equations Using the Hopfield Neural Network
NOSYNN-Based Neural Circuit for Solving Linear Equations
Hardware Simulation Results
Hardware Implementation
Low-Voltage CMOS-Compatible Linear Equation Solver
Comparison with Existing Works
Discussion on VLSI Implementation IssuesMixed-Mode Neural Circuit for Solving Linear EquationsMixed-Mode Neural Network for Solving Linear Equations
Hardware Simulation Results
Digitally-Controlled DVCC
DC-DVCC Based Linear Equation Solver
Hardware Simulation Results
Performance Evaluation
VLSI Implementation IssuesNon-Linear Feedback Neural Circuits for Linear and Quadratic ProgrammingNon-Linear Feedback Neural Network for Linear Programming
Non-Linear Feedback Neural Network for Solving QPP
Discussion on Energy Function
Issues in Actual Implementation
Comparison with Existing Works
Mixed-Mode Neural Circuits for LPP and QPPOTA-Based Implementations of Mixed-Mode Neural CircuitsOTA-Based Linear Equation Solver
Improved OTA-Based Linear Equation Solver
OTA-Based Graph Colouring Neural Network
OTA-Based Neural Network for Ranking
Linear Programming Using OTAs
OTA-Based QPP SolverFurther ReadingAppendix A. Mixed-Mode Neural Network for Graph Colouring
Appendix B. Mixed-Mode Neural Network for Ranking