Optical Fourier Transformation for Free-Space Convolutional Neural Networks: Physics, Algorithms, System Optimization, and Machine Learning Acceleration
Open AccessThe rapid advancement of artificial intelligence and deep learning has led to a growing demand for efficient and high-performance computing architectures. In this dissertation, we explore the utilization of optical Fourier transformation for accelerating the performance of free-space convolutional neural networks (CNNs). By leveraging the properties of optical systems, including low latency, energy efficiency, and the potential computing advantages, we investigate the application of optical Fourier transformation in the realm of CNNs.The dissertation begins by elucidating the fundamental principles of optical Fourier transformation and its relevance in free-space convolutional neural networks. Through a comprehensive exploration of the physics underlying coherent light propagation and interference, we design a powerful tool for neural network computation according to the theoretical foundations for optical Fourier transformation.Building upon the theoretical framework, we delve into the development of novel algorithms tailored for optical Fourier transformation in free-space CNNs. We investigate the design and optimization of algorithms for efficient Fourier domain modulation, leveraging the convolution theorem to accelerate the computation of convolutional layers. Additionally, we explore machine learning acceleration techniques that exploit the parallel processing capabilities inherent in optical systems.To ensure practical implementation, we address the challenges of system optimization and performance enhancement. We investigate methods for achieving precise wavefront shaping and alignment, overcoming limitations imposed by diffraction and interference. Furthermore, we explore techniques for system integration, including the integration of optical Fourier transformation with conventional electronic hardware, facilitating seamless hybrid computing architectures.Through extensive simulations and experimental validations, we demonstrate the feasibility and efficacy of the proposed optical Fourier transformation techniques for free-space convolutional neural networks. Our results showcase significant improvements in computational speed, energy efficiency, and parallelism compared to traditional electronic systems.The findings presented in this dissertation contribute to the broader field of optical computing and accelerate the advancement of free-space convolutional neural networks. By harnessing the power of optical Fourier transformation, we unlock new opportunities for high-performance, energy-efficient neural network architectures with applications in image recognition, pattern analysis, and deep learning.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.