Bio-inspired Hardware Architectures for Memory, Image Processing, and Control Applications

Abstract

Emerging technologies are expected to partially replace and enhance CMOS systems as the end of transistor scaling approaches. A particular type of emerging technology of interest is the variable resistance devices due to their scalability, non-volatile nature, and CMOS process compatibility. The goal of this dissertation is to present circuit and system level applications of CMOS and variable resistance devices with bio-inspired computation paradigms as the main focus. The summary of the results offered per chapter is as follows: In the first chapter of this thesis, an introduction to the work presented in the rest of this thesis and the model for the variable resistance device is provided. In the second chapter of this thesis, a crossbar memory architecture that utilizes a reduced constraint read-monitored-write scheme is presented. Variable resistance based crossbar memories are prime candidates to succeed the Flash as the mainstream nonvolatile memory due to their density, scalability, and write endurance. The proposed scheme supports multi-bit storage per cell and utilizes reduced hardware, aiming to decrease the feedback complexity and latency while still operating with CMOS compatible voltages. Additionally, a read technique that can successfully distinguish resistive states under the existence of resistance drift due to read/write disturbances in the array is presented. Derivations of analytical relations are provided to set forth a design methodology in selecting peripheral device parameters. In the third chapter of this thesis, an analog programmable resistive grid-based architecture mimicking the cellular connections of a biological retina in the most basic level, capable of performing various real time image processing tasks such as edge and line detections, is presented. Resistive grid-based analog structures have been shown to have advantages of compact area, noise immunity, and lower power consumption compared to their digital counterparts. However, these are static structures that can only perform one type of image processing task. The proposed unit cell structure employs 3-D confined resonant tunneling diodes called quantum dots for signal amplification and latching, and these dots are interconnected between neighboring cells through non-volatile continuously variable resistive elements. A method to program connections is introduced and verified through circuit simulations. Various diffusion characteristics, edge detection, and line detection tasks have been demonstrated through simulations using a 2-D array of the proposed cell structure, and analytical models have been provided. In the fourth chapter of this thesis, a bio-inspired hardware designed to solve the optimal control problem for general systems is presented. Adaptive Dynamic Programming algorithms provide means to approximate optimal control actions for linear and non-linear systems. Action-Critic Networks based approach is an efficient way to approximately evaluate the cost function and the optimal control actions. However, due to its computation intensiveness, this approach is usually implemented in high level programming languages run using general purpose processors. The presented hardware design is aimed at reducing the computation time and the hardware overhead by using the Heuristic Dynamic Programming algorithm which is a form of Adaptive Dynamic Programming. The proposed hardware operating at mere speed of 10 MHz yields 237 times faster learning rate in comparison to conventional software implementations running on fast processors such as the 1.2 GHz Intel Xeon processor.