Co-Design of Novel Devices for Neuromorphic Edge Computing
Сохранить в:
| Опубликовано в:: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Главный автор: | |
| Опубликовано: |
ProQuest Dissertations & Theses
|
| Предметы: | |
| Online-ссылка: | Citation/Abstract Full Text - PDF |
| Метки: |
Нет меток, Требуется 1-ая метка записи!
|
| Краткий обзор: | Due to the proliferation of AI models powered by deep neural networks (DNNs), computational resource usage for neural network training and inference has doubled every 3 months. The gap between available hardware resources and software requirements has driven the need for new devices, architectures, and algorithms beyond the confines of traditional von Neumann architecture. In particular, deploying these models and abilities to edge computing remains a challenge. One approach for computationally efficient computing involves the use of analog crossbar arrays, which leverages Ohm's Law to accelerate vector matrix multiplications. The devices that compose crossbar arrays can be referred to as artificial synapses, and are analogous to the weights of a neural network. At the outputs of the crossbar array, analog devices can act as effective activations for neural networks due to the ability to leverage complex behaviors for artificial neurons and dendrites. Additionally, the exploration of novel devices also enables co-design of architectures and algorithms that fit the physical characteristics of the device.This dissertation details the development of domain wall-magnetic tunnel junctions (DW-MTJs), graphene-Nafion transistors, and stochastic MTJs for DNN, Bayesian neural network (BNN), and stochastic computing applications. Domain-wall magnetic tunnel junctions are thoroughly characterized through micromagnetic simulations as artificial synapses, showing linear and symmetric behavior when notches are engineered into the track. The quantization effect of the notches is alleviated by leveraging the stochasticity of SOT-driven DW dynamics, showing that DW-MTJ devices can be effective artificial synapses for online learning. Additionally, by shaping the track, stream learning neural network simulations were performed based on experimental measurements to show that metaplasticity can combat catastrophic forgetting in continually learning systems. DW-MTJ devices have also been shown to be effective as artificial neurons. Edgy-relaxed behavior was shown through micromagnetic and analytical simulations and was beneficial for datasets with repeated instances. Additionally, experimental data from a stochastically switching DW-MTJ neuron device was used to simulate a spiking neural network (SNN), showing that noise resilience can be engineered into a hardware network using stochastic devices. Experimental data of fan integrate-and-fire (IF) and leaky integrate-and-fire (LIF) neuron was also used to respectively show higher generalizability and operation as a no-reset artificial neuron. For BNNs, the Bayes-MTJ was proposed, where a magneto-ionic or ferroelectric layer is engineered to the interface with the free layer of a MTJ to induce changes in the noise range of a stochastic superparamagnetic tunnel junction. This device is then integrated directly into crossbar arrays. With sufficient noise tunability, shown through a temperature-dependent Landau-Lifshitz-Gilbert (LLG) macrospin model, the hardware BNN was shown to represent a software BNN to almost ideal accuracies, which allows networks to characterize noise distributions. Additionally, work was done using the developed temperature-dependent LLG model to characterize VCMA, SOT, and stochastic write-driven MTJ random bitstream generators. All three methods were shown to be effective for generating high speed, energy efficient, and high quality bitstreams according to two NIST randomness tests. They were also shown to be effective for applications of generating arbitrary distributions and simulated annealing.For the graphene-Nafion ionic devices, they were characterized extensively for synaptic characteristics, along with endurance over 107 cycles, temperature resilience, and energy efficiency. Microscale devices were shown to have unique synaptic characteristics that allowed the composed crossbar arrays simulated in CrossSim to intrinsically perform an effect similar to weight normalization, enabling accuracies higher than ideal synapses on harder image classification tasks. By engineering an additional gate to the graphene-Nafion devices, artificial dendrite behavior emulating alpha and gaussian kernel behaviors was shown to allow spiking neural networks (SNNs) simulated through the Norse framework to be resilient to low spiking activity and reduced the energy usage of the SNN.This dissertation describes novel devices and co-design applications that leverage material-intrinsic behaviors for a broad range of neuromorphic applications, constructing a framework for computationally efficient analog edge computing. |
|---|---|
| ISBN: | 9798270235482 |
| Источник: | ProQuest Dissertations & Theses Global |