The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Bio-inspired learning and hardware acceleration with emerging memories
Author:	Kulkarni, Shruti R.
View Online:	njit-etd2019-075 (xxii, 153 pages ~ 3.2 MB pdf)
Department:	Department of Electrical and Computer Engineering
Degree:	Doctor of Philosophy
Program:	Electrical Engineering
Document Type:	Dissertation
Advisory Committee:	Rajendran, Bipin (Committee chair) Akansu, Ali N. (Committee member) Misra, Durgamadhab (Committee member) Rotstein, Horacio G. (Committee member) Simeone, Osvaldo (Committee member)
Date:	2019-12
Keywords:	Crossbar arrays Hardware acceleration Neuromorphic computing Non volatile memory Spiking neural networks
Availability:	Unrestricted
Abstract:	Machine Learning has permeated many aspects of engineering, ranging from the Internet of Things (IoT) applications to big data analytics. While computing resources available to implement these algorithms have become more powerful, both in terms of the complexity of problems that can be solved and the overall computing speed, the huge energy costs involved remains a significant challenge. The human brain, which has evolved over millions of years, is widely accepted as the most efficient control and cognitive processing platform. Neuro-biological studies have established that information processing in the human brain relies on impulse like signals emitted by neurons called action potentials. Motivated by these facts, the Spiking Neural Networks (SNNs), which are a bio-plausible version of neural networks have been proposed as an alternative computing paradigm where the timing of spikes generated by artificial neurons is central to its learning and inference capabilities. This dissertation demonstrates the computational power of the SNNs using conventional CMOS and emerging nanoscale hardware platforms. The first half of this dissertation presents an SNN architecture which is trained using a supervised spike-based learning algorithm for the handwritten digit classification problem. This network achieves an accuracy of 98.17% on the MNIST test data-set, with about 4X fewer parameters compared to the state-of-the-art neural networks achieving over 99% accuracy. In addition, a scheme for parallelizing and speeding up the SNN simulation on a GPU platform is presented. The second half of this dissertation presents an optimal hardware design for accelerating SNN inference and training with SRAM (Static Random Access Memory) and nanoscale non-volatile memory (NVM) crossbar arrays. Three prominent NVM devices are studied for realizing hardware accelerators for SNNs: Phase Change Memory (PCM), Spin Transfer Torque RAM (STT-RAM) and Resistive RAM (RRAM). The analysis shows that a spike-based inference engine with crossbar arrays of STT-RAM bit-cells is 2X and 5X more efficient compared to PCM and RRAM memories, respectively. Furthermore, the STT-RAM design has nearly 6X higher throughput per unit Watt per unit area than that of an equivalent SRAM-based (Static Random Access Memory) design. A hardware accelerator with on-chip learning on an STT-RAM memory array is also designed, requiring $16$ bits of floating-point synaptic weight precision to reach the baseline SNN algorithmic performance on the MNIST dataset. The complete design with STT-RAM crossbar array achieves nearly 20X higher throughput per unit Watt per unit mm^2 than an equivalent design with SRAM memory. In summary, this work demonstrates the potential of spike-based neuromorphic computing algorithms and its efficient realization in hardware based on conventional CMOS as well as emerging technologies. The schemes presented here can be further extended to design spike-based systems that can be ubiquitously deployed for energy and memory constrained edge computing applications.