Adaptive signal processing and correlational learning in mixed-signal VLSI

Portable electronic systems employ adaptive signal-processing algorithms to optimize their performance while facing severe constraints in power dissipation and circuit die-area. Software and custom-digital implementations of these algorithms are large and power hungry, while traditional analog and mixed-signal VLSI circuits are unable to provided the necessary accuracy due to limitations such as charge leakage, signal offsets, and device mismatch. This dissertation examines an important class of algorithms for adaptive signal processing and presents circuits and design techniques that enable large-scale analog and mixed-signal VLSI implementations that are compact, low-power, fast, and accurate.; I present a systems analysis of the Least-Mean-Squared (LMS) algorithm. I analyze the limitations of analog and digital VLSI arithmetic and determine their impact on the performance of the algorithm. I identify techniques that enable the automatic compensation of many of these effects using the adaptation intrinsic in the application, and point out those that require explicit compensation through calibration. I present examples of such calibration circuits and validate the approach through the implementation of three signal-processing systems. These systems use analog and mixed-signal arithmetic to achieve low-power operation and good performance, and use on-chip compensation to achieve good accuracy in the presence of nonlinearities and device mismatch. The resulting systems require two orders of magnitude less power and area than comparable digital solutions.