Publication Date

12-2016

Date of Final Oral Examination (Defense)

11-2-2016

Type of Culminating Activity

Dissertation

Degree Title

Doctor of Philosophy in Electrical and Computer Engineering

Department

Electrical and Computer Engineering

Major Advisor

John Chiasson, Ph.D.

Major Advisor

Vishal Saxena, Ph.D.

Advisor

Hao Chen, Ph.D.

Abstract

Human society is now facing grand challenges to satisfy the growing demand for computing power, at the same time, sustain energy consumption. By the end of CMOS technology scaling, innovations are required to tackle the challenges in a radically different way. Inspired by the emerging understanding of the computing occurring in a brain and nanotechnology-enabled biological plausible synaptic plasticity, neuromorphic computing architectures are being investigated. Such a neuromorphic chip that combines CMOS analog spiking neurons and nanoscale resistive random-access memory (RRAM) using as electronics synapses can provide massive neural network parallelism, high density and online learning capability, and hence, paves the path towards a promising solution to future energy-efficient real-time computing systems. However, existing silicon neuron approaches are designed to faithfully reproduce biological neuron dynamics, and hence they are incompatible with the RRAM synapses, or require extensive peripheral circuitry to modulate a synapse, and are thus deficient in learning capability. As a result, they eliminate most of the density advantages gained by the adoption of nanoscale devices, and fail to realize a functional computing system.

This dissertation describes novel hardware architectures and neuron circuit designs that synergistically assemble the fundamental and significant elements for brain-inspired computing. Versatile CMOS spiking neurons that combine integrate-and-fire, passive dense RRAM synapses drive capability, dynamic biasing for adaptive power consumption, in situ spike-timing dependent plasticity (STDP) and competitive learning in compact integrated circuit modules are presented. Real-world pattern learning and recognition tasks using the proposed architecture were demonstrated with circuit-level simulations. A test chip was implemented and fabricated to verify the proposed CMOS neuron and hardware architecture, and the subsequent chip measurement results successfully proved the idea.

The work described in this dissertation realizes a key building block for large-scale integration of spiking neural network hardware, and then, serves as a step-stone for the building of next-generation energy-efficient brain-inspired cognitive computing systems.

Download

Share

COinS