WEKO3
アイテム
スピンホール発振器を使用した脳型計算処理: 分類および予測タスクのための非線形磁化ダイナミクスのモデル化と活用
https://doi.org/10.18997/0002000488
https://doi.org/10.18997/00020004881bb06665-0840-4d80-ba66-1042218d2944
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
jou_k_386.pdf (6.9 MB)
|
|
| Item type | 学位論文 = Thesis or Dissertation(1) | |||||||
|---|---|---|---|---|---|---|---|---|
| 公開日 | 2024-04-09 | |||||||
| 資源タイプ | ||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_db06 | |||||||
| 資源タイプ | doctoral thesis | |||||||
| タイトル | ||||||||
| タイトル | Neuromorphic Computing with Spin Hall Oscillators: Modelling and Leveraging of nonlinear magnetization dynamics for classification and prediction computational tasks | |||||||
| 言語 | en | |||||||
| タイトル | ||||||||
| タイトル | スピンホール発振器を使用した脳型計算処理: 分類および予測タスクのための非線形磁化ダイナミクスのモデル化と活用 | |||||||
| 言語 | ja | |||||||
| 言語 | ||||||||
| 言語 | eng | |||||||
| 著者 |
Mohan, John Rex
× Mohan, John Rex
|
|||||||
| 抄録 | ||||||||
| 内容記述タイプ | Abstract | |||||||
| 内容記述 | The landscape of information technology has been profoundly reshaped by the emergence of Artificial Intelligence (AI) and Machine Learning (ML), catalyzing transformative shifts across industries and enhancing human interactions [1]. These advancements mark a shift from conventional information processing to a new era of intelligent computing. Central to this transformation is the capability of computers to analyze voluminous datasets, extracting significant insights that empower informed decision-making. Notably, OpenAI's ChatGPT exemplifies this paradigm shift, a language model adept at contextual comprehension and human-like responsiveness. Brain-inspired Artificial Neural Networks (ANNs) strive to emulate human brain information processing through multiple processing layer computational models that can learn representations of data at various levels of abstraction. Executed on Von Neumann architectures, these ANNs employ algorithms like backpropagation to fine-tune weights and replicate learning mechanisms [2]. Nonetheless, the journey of ANNs is obstructed by scalability challenges that demand innovative solutions to bridge the gap between artificial and biological intelligence. The limitations of the Von Neumann architecture, where processing and memory exist as distinct entities, constrain traditional ANN implementations, leading to processing power limitations and functional constraints. The renowned "von Neumann bottleneck" obstructs data-intensive operations, hindering parallelism and inducing inefficiencies in real-time data processing and AI inference. The evolution beyond Von Neumann architecture investigates different computing paradigms like neuromorphic, quantum, and unconventional methods. Spiking neural networks and memristors are two examples of neuromorphic devices that attempt to combine memory and processing to mimic the unified functionality of the human brain [3]. These devices simulate synapses and neurons found in biological systems, allowing for unified communication and parallel processing. The offloading of intensive computational tasks from the conventional computing architecture is where tailored neuromorphic components show promise for real-time computations. Such components have great potential for real-time computations and are ideal for memory-constrained gadgets like wearables, Internet of Things (IoT) devices, and embedded systems [4]. While ANNs excel in classification and pattern recognition tasks, incorporating dedicated neuro-inspired computing units mandates efficient signal processing, seamless Complementary Metal-Oxide-Semiconductor (CMOS) circuit integration, and adaptability with the existing machine learning algorithms. Integrating specialized computing elements with CMOS technology is pivotal in bridging the gap between conventional and unconventional computing paradigms. The implementation of specialized inference or feature extraction computing units holds the potential to significantly mitigate energy costs associated with feature mapping, a substantial proportion of current ANN expenditures [5]. The promise of spintronic devices, with their inherent nonlinear magnetization dynamics, as prospective candidates for neuromorphic hardware and unconventional computing components is compelling [6,7]. Spin torque oscillators, comprising spin transfer torque oscillators and spin Hall oscillators, showcase remarkable capabilities in classification and recognition tasks. This thesis investigates the realm of information processing capability of Spin Hall Oscillators (SHOs) using macrospin-level (micromagnetic) simulations. SHOs emerge as generators of high-frequency microwave signals and nonlinear magnetization dynamics, presenting opportunities in simple signal processing endeavors. The research aims to model SHO(s) as specialized computing component(s), adept at efficient signal processing, reduced computations, embracing real-time inference capabilities, and serving memory-constrained devices. Furthermore, the investigation extends into Reservoir Computing (RC) strategies, bolstering SHOs' information-handling prowess. To achieve these objectives, certain restrictions are imposed, guiding the course of the research:1. Designing computing components to offload computational complexity while minimizing memory utilization, 2. Seamless integration with conventional signal processing techniques to align with current computing architectures, 3. Ensuring real-time operation and suitability for memory-constrained devices to cater to diverse application scenarios. The study commences by showcasing SHOs' capability in classification tasks, adaptable for processing binary data inputs nonlinearly, enabling real-time feature extraction and classification. When combined with frequency domain filtering, input driven magnetization dynamics can be used to classify 4-bit binary digit patterns with a single floating-point output. This novel methodology, which eradicates the need for weight storage in the initial layer of computation, shows the capability of SHO's self-computation based on the order of inputs in the pattern. The methodology is applied to classify handwritten digit images from the Modified National Institute of Standards and Technology database [8]. In a simple linear regression model, the model achieves an accuracy of 83.1%, demonstrating the effectiveness of the SHO for real-time and on-device neuromorphic framework. Furthermore, the research also investigates the use of a single SHO in reservoir computing, a machine learning framework that uses recurrently connected nodes to effectively process sequential data. Memory capacity (MC) of a reservoir is a measure of the amount of data it can store and use over time [9]. It is important for a variety of reservoir computing tasks, such as time series prediction, nonlinear data transformation, and temporal pattern identification. We show that the reservoir's memory capacity and its use for temporal tasks are directly related. When SHO output magnetization dynamics include both transient state and limit cycle oscillations, the best reservoir computing results are obtained. The effectiveness of temporal tasks is revealed to be significantly correlated with reservoir memory capacity. The effect of input current pulse parameters on the memory capacity of SHOs is investigated. The results show an improvement trend with increasing pulse amplitude and width, peaking in the 4.5–5.0 range. Nonlinear Autoregressive Moving Average Mode 2 (NARMA2) time series prediction task and the three-bit parity task are used to test SHO's performance as a reservoir computing system, confirming a strong correlation between memory size and temporal task performance. Finally, the nonlinear dynamics of the magnetization, high-frequency oscillations, and cooperative behavior enabled by dipolar coupling of SHOs (dSHO) is investigated. The use of an array of dSHOs is a novel approach to enhance memory capacity in the spatial and temporal domains. Dipolar coupling introduces a cooperative behavior component, allowing interaction and storing and retrieving of complex temporal patterns. The systems' memory capacity can effectively be increased to 10 by using dSHOs for spatial domain extension, which also improves their ability to predict. Significantly, the approach substantially expedites large-scale data processing, speeds up prediction and classification. This accelerated functionality holds the promise of immediate decision-making in domains such as self-driving vehicles and financial predictions. In conclusion, the integration of Spin Hall Oscillators (SHOs) marks a pivotal point in computing by combining neuromorphic computing with existing computing architecture. This results in computing that is effective, flexible, and memory-efficient. To maximize the computing potential of SHOs, we concentrated on machine learning adaptability and the efficient signal processing capability of CMOS integration. By enabling more effective, adaptable systems that go in reservoir computing and beyond conventional approaches, these devices have the potential to fundamentally alter the computing landscape. Intelligent computing is made possible by the adaptability of SHOs to machine learning algorithms, enabling pattern recognition and decision-making in applications like image recognition, robotics, and autonomous vehicles. |
|||||||
| 目次 | ||||||||
| 内容記述タイプ | TableOfContents | |||||||
| 内容記述 | 1. Introduction| 2. Background: Neuromorphic Computing and Spintronics| 3. Materials and Methods| 4. Classification task using spin Hall oscillators – Self computing unit| 5. Spin Hall Oscillator for Reservoir Computing| 6. Enhancing information processing capability of SHOs – magnetic dipolar approach | |||||||
| 備考 | ||||||||
| 内容記述タイプ | Other | |||||||
| 内容記述 | 九州工業大学博士学位論文 学位記番号:情工博甲第386号 学位授与年月日:令和5年12月27日 | |||||||
| 学位授与番号 | ||||||||
| 学位授与番号 | 甲第386号 | |||||||
| 学位名 | ||||||||
| 学位名 | 博士（情報工学） | |||||||
| 学位授与年月日 | ||||||||
| 学位授与年月日 | 2023-12-27 | |||||||
| 学位授与機関 | ||||||||
| 学位授与機関識別子Scheme | kakenhi | |||||||
| 学位授与機関識別子 | 17104 | |||||||
| 言語 | ja | |||||||
| 学位授与機関名 | 九州工業大学 | |||||||
| 学位授与年度 | ||||||||
| 内容記述タイプ | Other | |||||||
| 内容記述 | 令和5年度 | |||||||
| 出版タイプ | ||||||||
| 出版タイプ | VoR | |||||||
| 出版タイプResource | http://purl.org/coar/version/c_970fb48d4fbd8a85 | |||||||
| アクセス権 | ||||||||
| アクセス権 | open access | |||||||
| アクセス権URI | http://purl.org/coar/access_right/c_abf2 | |||||||
| ID登録 | ||||||||
| ID登録 | 10.18997/0002000488 | |||||||
| ID登録タイプ | JaLC | |||||||
