A Biomimetic Approach to Non-Linear Signal Processing in Ultra Low Power Analog Circuits

The human brain has perfected the task of performing complex functions, such as learning, memory, and cognition, in energy and area efficient manners, and therefore, it is attractive to model and replicate this performance in biomimetic systems. Applications for which such models are essential include implantable electronics that replace certain brain functions damaged by diseases or injury, and neuromorphic architectures for novel high-speed, low power computing. To properly mimic a neural function, it is important to determine the optimal level of abstraction because there are trade-offs between capturing biological complexities and the scalability and efficiency of the model. Additionally, the hardware implementation of that model must consume little power to minimize heat dissipation in order to avoid tissue damage and to allow for scaling to a large number of components. The Laguerre Expansion of Volterra (LEV) model is a promising approach since it can capture the spatio-temporal nonlinearities of a neural system at several abstraction levels. Similar to the brain, it uses basic processing strategies and operations like amplification, filtering, delay, and redundancy to produce complex functionalities. This doctoral research addresses the hardware challenges of a biomimetic system and proposes a novel method for its practical implementation. The LEV model is realized using analog subthreshold CMOS signal processing units, which are more power and area efficient than the digital counterparts. Several challenges in subthreshold analog design are addressed in this work. First, application-specific analog circuits are neither easily programmable nor flexible. Second, in the subthreshold regime mismatch and process variations lead to large differences in identically-sized transistors. The brain performs complex functions despite facing similar limitations by using redundancy and learning rules that adjust cell properties. Guided by these principles the analog subthreshold circuits are designed to be digitally programmable for coefficient and time constant adjustments. Also, since Laguerre basis functions of different orders are not completely orthogonal, a redundancy is created that helps reproduce the composite waveform more accurately. Calibration and training techniques are proposed in this work to accomplish the required precision. These techniques are verified through numerical simulations, and physical implementation and measurement of a system fabricated in a 0.13 &mgr;m CMOS process. This work demonstrates the utility of these optimized low-power subthreshold analog circuits, aided by calibration, and uses them to implement the LEV model of a single-input, single-output spike system to replicate the signal transformations of a single neuron. A foundation is laid for an easily scalable system useful for multi-input, multi-output (MIMO) models, to be used in systems such as a hippocampus prosthesis for memory restoration or a large scale neuromorphic computer.