Simulation and Design of Biological and Biologically-Motivated Computing Systems
In life science, there is a great need in understandings of biological systems for therapeutics, synthetic biology, and biomedical applications. However, complex behaviors and dynamics of biological systems are hard to understand and design. In the mean time, the design of traditional computer architectures faces challenges from power consumption, device reliability, and process variations. In recent years, the convergence of computer science, computer engineering and life science has enabled new applications targeting the challenges from both engineering and biological fields. On one hand, computer modeling and simulation provides quantitative analysis and predictions of functions and behaviors of biological systems, and further facilitates the design of synthetic biological systems. On the other hand, bio-inspired devices and systems are designed for real world applications by mimicking biological functions and behaviors. This dissertation develops techniques for modeling and analyzing dynamic behaviors of biologically realistic genetic circuits and brain models and design of brain-inspired computing systems. The stability of genetic memory circuits is studied to understand its functions for its potential applications in synthetic biology. Based on the electrical-equivalent models of biochemical reactions, simulation techniques widely used for electronic systems are applied to provide quantitative analysis capabilities. In particular, system-theoretical techniques are used to study the dynamic behaviors of genetic memory circuits, where the notion of stability boundary is employed to characterize the bistability of such circuits. To facilitate the simulation-based studies of physiological and pathological behaviors in brain disorders, we construct large-scale brain models with detailed cellular mechanisms. By developing dedicated numerical techniques for brain simulation, the simulation speed is greatly improved such that dynamic simulation of large thalamocortical models with more than one million multi-compartment neurons and hundreds of synapses on commodity computer servers becomes feasible. Simulation of such large model produces biologically meaningful results demonstrating the emergence of sigma and delta waves in the early and deep stages of sleep, and suggesting the underlying cellular mechanisms that may be responsible for generation of absence seizure. Brain-inspired computing paradigms may offer promising solutions to many challenges facing the main stream Von Neumann computer architecture. To this end, we develop a biologically inspired learning system amenable to VLSI implementation. The proposed solution consists of a digitized liquid state machine (LSM) and a spike-based learning rule, providing a fully biologically inspired learning paradigm. The key design parameters of this liquid state machine are optimized to maximize the learning performance while considering hardware implementation cost. When applied to speech recognition of isolated word using TI46 speech corpus, the performance of the proposed LSM rivals several existing state-of-art techniques including the Hidden Markov Model based recognizer Sphinx-4.