Adaptive Analog VLSI Signal Processing and Neural Networks
Jeffery Don Dugger · UPT. Syiah Kuala University Library (Syiah Kuala University) · 2003
Research presented in this thesis provides \ta substantial leap from the study of interesting \tdevice physics to fully adaptive analog networks \tand lays a solid foundation for future development \tof large-scale, compact, low-power adaptive parallel \tanalog computation systems. The investigation described here started with \tobservation of this potential learning capability \tand led to the first derivation and characterization of \tthe floating-gate pFET correlation learning rule. Starting with two synapses sharing the same error signal, \twe progressed from phase correlation experiments \tthrough correlation experiments involving harmonically related sinusoids, \tculminating in learning the Fourier series coefficients \tof a square wave cite{kn:Dugger2000}. Extending these earlier two-input node experiments to the general case \tof correlated inputs required dealing with \tweight decay naturally exhibited by the learning rule. We introduced a source-follower floating-gate synapse \tas an improvement over our earlier source-degenerated floating-gate synapse \tin terms of relative weight decay cite{kn:Dugger2004}. A larger network of source-follower floating-gate synapses was fabricated \tand an FPGA-controlled testboard was designed and built. This more sophisticated system provides an excellent \tframework for exploring applications to multi-input, multi-node \tadaptive filtering applications. Adaptive channel equalization provided \ta practical test-case illustrating the use \tof these adaptive systems in solving real-world problems. The same system could easily be applied to noise and echo cancellation \tin communication systems and system identification tasks in \toptimal control problems. We envision the commercialization of these adaptive analog VLSI \tsystems as practical products within a couple of years.