Optimizing Training Methodologies for Low-Latency and Energy-Efficient Neuromorphic Computing Systems

open_access
Open Access
Author:
Mahmoodi Takaghaj, Sanaz
Graduate Program:
Computer Science and Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
February 18, 2025
Committee Members:
Chitaranjan Das, Program Head/Chair
Dezhe Jin, Outside Unit & Field Member
John Sampson, Chair & Dissertation Advisor
Robert Collins, Major Field Member
Abhinav Verma, Major Field Member
Keywords:
Neuromorphic systems
Spiking Neural Networks (SNNs)
Rouser
Exemplar LCA
Sparse coding
Vision Transformers (ViT)
PointLCA-Net
Event-based datasets
Energy-efficient AI
Abstract:
In recent years, the field of neuromorphic systems has gained significant traction due to the growing demand for AI applications in energy-constrained autonomous systems, such as unmanned aerial vehicles (UAVs) and robotics. Neuromorphic computing is inspired by the energy-efficient processing and communication mechanisms of biological Spiking Neural Networks (SNNs). SNNs are computational models that emulate the behavior of neurons in the brain, offering promising prospects for energy-efficient and parallel data processing. However, the inherent temporal nature and spike-based communication of these neurons pose challenges in training these networks, limiting their scaling and broader adoption. Unlike ANNs, which use continuous activation amplitudes to extract features from data, SNNs transmit data through discrete spikes. With the remarkable success of error backpropagation in the training of Artificial Neural Networks (ANNs), the current approach to training SNNs revolves around using surrogate gradients and backpropagation. Although these training algorithms are actively being investigated, their computational intensity leads to significant energy and memory consumption, raising questions about the efficacy of SNNs. In this dissertation, we explore new training methods for SNNs that focus on their stateful dynamics, thresholding behavior, and sparsity through spiking and inhibition. In the first facet of this research, we improved the training of SNNs using error backpropagation, while simultaneously gaining insights into the dynamics of learning within SNNs. Conventional SNNs rely on pre-determined threshold values, limiting their ability to adapt to various datasets, dynamic inputs, and altering network performance. By incorporating the Rouser threshold adjustment mechanism, we have improved the adaptability and responsiveness of SNNs, allowing them to better accommodate varying input patterns and optimize their learning capabilities. This has resulted in better accuracy and faster convergence than a fixed-threshold feed-forward SNN. In order to tackle the high-energy/memory demands and computational challenges posed by training SNNs with backpropagation, the second aspect of this dissertation involves leveraging sparse coding and the Exemplar Locally Competitive Algorithm (LCA). Exemplar LCA is a sparse coding technique, inspired by the observed sparsity in neural activity in visual cortex, efficiently encodes input data using only a subset of spiking neurons. By integrating LCA into the training of SNNs within Exemplar LCA-Decoder, we reduced the computational demands and memory requirements associated with gradient calculations and error backpropagation. Our experiments showed a test accuracy of 98.57\% in the MNIST dataset, exceeding the previously reported accuracy of 96.69\% achieved using the (original) LCA. We further enhanced LCA-Decoder by using features extracted from pre-trained Convolutional Neural Network (CNN) models to improve the accuracy on additional datasets, including CIFAR-10, CIFAR-100, and ImageNet-1K. Our results show the highest reported top-1 test accuracy for SNNs in the ImageNet-1K and CIFAR-100 datasets, surpassing previous benchmarks. Specifically, we achieved a record top-1 accuracy of 80.75\% in ImageNet-1K(ILSVRC2012 validation set) and 79.32\% in CIFAR-100. We also achieved an accuracy of 94.98\% on CIFAR-10. The recent success of Transformers in Large Language Models (LLMs) has sparked significant interest in attention mechanisms and transformer architectures. In this dissertation, we explore vision transformers (ViT) and leverage self-attention representations in ViT-LCA. We also demonstrated how ViT-LCA can be implemented on memory crossbars to enable energy-efficient neuromorphic deployment. Our experiments show that ViT-LCA achieves higher accuracy on ImageNet dataset while consuming significantly less energy than other spiking vision transformer counterparts. We also achieved the highest reported SNN accuracy of 95.63\% on CIFAR-10 and 81.80\% on CIFAR-100. Finally, with PointLCA-Net, we leverage PointNets to extend Exemplar LCA-Decoder for neuromorphic (event-based) datasets, which are composed of sparse spatio-temporal recordings. These sparse spatio-temporal signals can be treated as point clouds, representing the space and time at which these events occur. With PointLCA-Net, we leverage PointNets to extract robust features from input point clouds and utilize the efficiency of the Exemplar LCA Decoder to encode and decode these features. PointLCA-Net achieves an accuracy of 93.41\% in the DVS128 dataset while reducing energy consumption by approximately 92\% compared to other spiking neural networks applied to point clouds, making it more efficient and well-suited for energy-constrained applications. The results of this dissertation have the potential to advance neuromorphic computing with spiking neural networks, enabling their adoption across various domains, including robotics, UAVs, and other fields that require energy-efficient, real-time processing of sensory data.