Review
Nanoscience & Nanotechnology
Abdourahman Khaireh-Walieh, Denis Langevin, Pauline Bennet, Olivier Teytaud, Antoine Moreau, Peter R. Wiecha
Summary: This article provides a practical guide for utilizing deep learning in solving inverse design problems in nanophotonics. It explores different iterative and direct deep learning techniques, evaluating their advantages and limitations. The tutorial aims to empower newcomers to leverage the potential of deep learning in their scientific pursuits through detailed Python examples and practical design guidelines.
Article
Mechanics
Max Daniels, Cedric Gerbelot, Florent Krzakala, Lenka Zdeborova
Summary: Signal recovery under generative neural network priors has emerged as a promising direction in statistical inference and computational imaging. This paper overcomes the limitation of assuming Gaussian i.i.d. weights for generative priors by establishing the state evolution of ML-AMP for random convolutional layers. The proof technique establishes a mapping between convolutional matrices and spatially coupled sensing matrices used in coding theory.
JOURNAL OF STATISTICAL MECHANICS-THEORY AND EXPERIMENT
(2023)
Article
Computer Science, Theory & Methods
Jing Wu, Lin Wang, Qiangyu Pei, Xingqi Cui, Fangming Liu, Tingting Yang
Summary: This article introduces HiTDL, a runtime framework for managing multiple DNNs provisioned following the hybrid approach at the edge. HiTDL aims to improve edge resource efficiency and guarantee SLA by optimizing the combined throughput of all co-located DNNs. It utilizes performance models and solves a multiple-choice knapsack problem to make throughput-optimal resource allocation decisions. Experimental results show a 4.3x improvement in overall throughput compared to the state-of-the-art.
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
(2022)
Article
Mechanics
Preetum Nakkiran, Gal Kaplun, Yamini Bansal, Tristan Yang, Boaz Barak, Ilya Sutskever
Summary: A study reveals a 'double-descent' phenomenon in modern deep learning tasks, where performance initially worsens and then improves as model size increases. The concept of effective model complexity is introduced to unify these phenomena and conjecture a generalized double descent based on this measure. Additionally, certain scenarios are identified where increasing the number of training samples, even quadrupling, can actually harm test performance.
JOURNAL OF STATISTICAL MECHANICS-THEORY AND EXPERIMENT
(2021)
Article
Nanoscience & Nanotechnology
Rohit Unni, Kan Yao, Xizewen Han, Mingyuan Zhou, Yuebing Zheng
Summary: The research introduces a tandem optimization model that combines a mixture density network (MDN) and a fully connected network to inversely design practical thin-film high reflectors. This model can retrieve the reflectance spectra of 20-layer thin-film structures and demonstrate improved designs with extended high-reflectance zones. By combining the efficiency advantage of DL with optimization-enabled performance improvement, efficient and on-demand inverse design for practical applications is enabled.
Article
Computer Science, Information Systems
Emna Baccour, Naram Mhaisen, Alaa Awad Abdellatif, Aiman Erbad, Amr Mohamed, Mounir Hamdi, Mohsen Guizani
Summary: Artificial intelligence has made significant breakthroughs in IoT applications and services, and the emergence of pervasive AI has expanded the role of ubiquitous IoT systems, from mainly data collection to executing distributed computations. This paper provides a comprehensive survey of the latest techniques and strategies for overcoming resource challenges in pervasive AI systems.
IEEE COMMUNICATIONS SURVEYS AND TUTORIALS
(2022)
Article
Chemistry, Multidisciplinary
Lakshmi Raju, Kyu-Tae Lee, Zhaocheng Liu, Dayu Zhu, Muliang Zhu, Ekaterina Poutrina, Augustine Urbas, Wenshan Cai
Summary: This research demonstrates the use of a deep learning framework to create an optimal design for a nonlinear metamaterial, maximizing its nonlinear effect. A nanolaminate metamaterial is used for validation, showing the effectiveness of the deep learning algorithm.
Article
Optics
Yan Teng, Chun Li, Shaochen Li, Yuhua Xiao, Ling Jiang
Summary: In this study, a method for designing terahertz random meta-surfaces based on deep Convolutional Neural Networks and genetic algorithms is proposed. The forward prediction model accurately predicts the reflection amplitude and phase response, with a calculation speed increased by 40,000 times compared to the Full-wave solver. By combining with genetic algorithms, the efficiency of the design is greatly improved, providing an efficient method for global optimization in complex designs.
OPTICS AND LASER TECHNOLOGY
(2023)
Article
Nanoscience & Nanotechnology
Raphael Pestourie, Wenjie Yao, Boubacar Kante, Steven G. Johnson
Summary: In this study, we propose a reciprocity technique to efficiently handle incoherent light in the design of optical devices. By reducing the number of scattering calculations from thousands to one Maxwell solve, we achieve a speedup of three orders of magnitude. This improvement allows for efficient inverse design and large-scale optimization of metasurfaces for applications such as light collimators and concentrators. We also demonstrate the impact of the angular distribution of incident light on the device's performance and present particularly promising designs for annular beams distributed only over nonzero angles.
Review
Nanoscience & Nanotechnology
Qizhou Wang, Maksim Makarenko, Arturo Burguete Lopez, Fedor Getman, Andrea Fratalocchi
Summary: Nanophotonics inverse design is a rapidly growing field that focuses on defining complex optical functionalities while using machines to search for material and geometry configurations in sub-wavelength structures. The emergence of deep learning has started to play a significant role in this area.
Article
Nanoscience & Nanotechnology
Yubin Gao, Qikai Chen, Sijie Pian, Yaoguang Ma
Summary: Due to its sub-wavelength resolution, compact size, and ease of integration, flat optics has become a crucial aspect in diffractive metalens, structured-light generation, holograms, nonlocal metasurfaces, and other applications. Inverse design has emerged as an essential approach in flat optics, allowing designers to focus on high-level optical functionalities while leaving the task of finding conforming structures to machines. This article outlines the main advances of inverse design in flat optics, from basic theories and algorithms to further developments and diversified applications.
PHOTONICS AND NANOSTRUCTURES-FUNDAMENTALS AND APPLICATIONS
(2022)
Article
Optics
Jiankai Xiong, Jiaqing Shen, Yuan Gao, Yingshi Chen, Jun-Yu Ou, Qing Huo Liu, Jinfeng Zhu
Summary: This article presents a divide-and-conquer deep learning model for the design of plasmonic stack metamaterials (PSMs). The model demonstrates significant reduction in prediction error and training parameters in the forward network, supporting powerful inverse design from spectra to PSM structures. Additionally, a flexible tool based on free customer definition is developed for real-time design of metamaterials with various circuit-analog functions.
LASER & PHOTONICS REVIEWS
(2023)
Article
Nanoscience & Nanotechnology
Yang Deng, Simiao Ren, Jordan Malof, Willie J. Padilla
Summary: Deep learning has been widely applied to solve inverse problems in artificial electromagnetic materials (AEMs). Deep inverse models have achieved impressive results surpassing capabilities of other approaches. This article provides an overview of the process, discusses important issues, and presents an outlook for future development in this field.
PHOTONICS AND NANOSTRUCTURES-FUNDAMENTALS AND APPLICATIONS
(2022)
Article
Computer Science, Hardware & Architecture
Qinglin Yang, Xiaofei Luo, Peng Li, Toshiaki Miyazaki, Wenfeng Shen, Weiqin Tong
Summary: This paper proposes collaborative inference among mobile devices to share computation workloads and accelerate processing speed by batching inference tasks on GPUs. An algorithm based on PSO is designed for efficient collaboration, as well as a distributed algorithm to address the challenge of collecting global network information and running centralized algorithms. Extensive simulations show that the collaborative inference scheme effectively reduces inference time for mobile deep learning applications.
MOBILE NETWORKS & APPLICATIONS
(2021)
Article
Engineering, Electrical & Electronic
Bonan Zhang, Lung-Yen Chen, Naveen Verma
Summary: This work extends the S-DDHR model to enable statistical computation models for hardware energy efficiency, by incorporating the statistical distribution of hardware variations for model-parameter learning. The approach aims to develop accurate and composable abstractions of computations, enabling scalable hardware-generalized training.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I-REGULAR PAPERS
(2021)
Article
Nanoscience & Nanotechnology
Deniz Mengu, Yifan Zhao, Anika Tabassum, Mona Jarrahi, Aydogan Ozcan
Summary: This study presents the use of diffractive optical networks to perform permutation operations with a large number of input-output connections. By using deep learning techniques, the capacity of the diffractive optical network to approximate permutation operations increases with the number of diffractive layers and trainable transmission elements in the system. The authors also addressed challenges related to physical alignment and output efficiency by designing misalignment tolerant diffractive designs. The proposed diffractive permutation networks have potential applications in security, image encryption, and data processing, and can serve as channel routing and interconnection panels in wireless networks.
Article
Chemistry, Multidisciplinary
Bijie Bai, Heming Wei, Xilin Yang, Tianyi Gan, Deniz Mengu, Mona Jarrahi, Aydogan Ozcan
Summary: This paper presents a diffractive optical network that performs data-class-specific transformations between the input and output fields-of-view (FOVs) using optical methods. The visual information of objects is encoded into the amplitude, phase, or intensity of the optical field at the input and processed by a data-class-specific diffractive network. The output patterns are optically encrypted using preassigned transformation matrices, and the original input images can be recovered by applying the correct decryption key.
ADVANCED MATERIALS
(2023)
Article
Engineering, Electrical & Electronic
Yuhang Li, Yi Luo, Bijie Bai, Aydogan Ozcan
Summary: Imaging through diffusive media is a challenging problem that typically requires digital computers for image reconstruction. In this study, we propose an alternative method using diffractive neural networks to see through random, unknown phase diffusers. Through detailed analysis, we observed a trade-off between image reconstruction fidelity and distortion reduction capability of the diffractive network. We also discovered that training the network with a variety of random diffusers and introducing misalignments improved its generalization performance.
IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS
(2023)
Article
Multidisciplinary Sciences
Jingxi Li, Tianyi Gan, Yifan Zhao, Bijie Bai, Che-Yung Shen, Songyu Sun, Mona Jarrahi, Aydogan Ozcan
Summary: The first demonstration of unidirectional imagers is reported, presenting polarization-insensitive and broadband uni-directional imaging based on successive diffractive layers that are linear and isotropic. The diffractive unidirectional imager maintains its functionality over a large spectral band and works under broadband illumination. The diffractive unidirectional imaging using structured materials will have applications in security, defense, telecommunications, and privacy protection.
Article
Engineering, Biomedical
Tairan Liu, Yuzhu Li, Hatice Ceylan Koydemir, Yijie Zhang, Ethan Yang, Merve Eryilmaz, Hongda Wang, Jingxi Li, Bijie Bai, Guangdong Ma, Aydogan Ozcan
Summary: An automated plaque assay leveraging lens-free holographic imaging and deep learning rapidly and accurately detects the cell-lysing events caused by viral replication.
NATURE BIOMEDICAL ENGINEERING
(2023)
Article
Chemistry, Multidisciplinary
Yuhang Li, Jingxi Li, Yifan Zhao, Tianyi Gan, Jingtian Hu, Mona Jarrahi, Aydogan Ozcan
Summary: A universal polarization transformer is demonstrated that can synthesize various complex-valued polarization scattering matrices, providing a solution for controlled synthesis of optical fields with nonuniform polarization distributions.
ADVANCED MATERIALS
(2023)
Article
Engineering, Electrical & Electronic
Hanlong Chen, Luzhe Huang, Tairan Liu, Aydogan Ozcan
Summary: The application of deep learning techniques has significantly improved holographic imaging capabilities, with enhanced phase recovery and image reconstruction. This study introduces eFIN, a deep neural network that utilizes pixel super-resolution and image autofocusing for hologram reconstruction. Experimental results demonstrate the superior image quality and external generalization of eFIN, which achieves a wide autofocusing range and accurately predicts hologram axial distances. The network also enables 3x pixel super-resolution and improves the space-bandwidth product of reconstructed images by 9-fold.
IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS
(2023)
Review
Optics
Xurong Li, Jingxi Li, Yuhang Li, Aydogan Ozcan, Mona Jarrahi
Summary: This article reviews the development of terahertz imaging technologies and discusses different types of hardware and computational imaging algorithms. It explores opportunities for capturing various image data and briefly introduces the prospects and challenges for future high-throughput terahertz imaging systems.
LIGHT-SCIENCE & APPLICATIONS
(2023)
Article
Computer Science, Artificial Intelligence
Luzhe Huang, Hanlong Chen, Tairan Liu, Aydogan Ozcan
Summary: GedankenNet is a self-supervised learning model that achieves image reconstruction without the need for labelled or experimental training data, demonstrating superior generalization on hologram reconstruction tasks.
NATURE MACHINE INTELLIGENCE
(2023)
Article
Optics
Jingxi Li, Tianyi Gan, Bijie Bai, Yi Luo, Mona Jarrahi, Aydogan Ozcan
Summary: Large-scale linear operations are crucial for complex computational tasks, and optical computing offers advantages in terms of speed, parallelism, and scalability. The deep-learning-based design of a broadband diffractive neural network enables the performance of a large group of complex-valued linear transformations. By assigning different illumination wavelengths to each transformation, a single diffractive network can execute multiple linear transformations simultaneously or sequentially. This technology allows for the approximation of unique linear transforms with negligible errors, and the spectral multiplexing capability can be increased by increasing the number of diffractive neurons.
ADVANCED PHOTONICS
(2023)
Article
Automation & Control Systems
Md Sadman Sakib Rahman, Aydogan Ozcan
Summary: This study demonstrates for the first time a time-lapse image classification scheme using a diffractive network, improving classification accuracy and generalization performance by leveraging the lateral movements of the input objects and/or the diffractive network. Numerical exploration reveals a blind testing accuracy of 62.03% on the optical classification of objects from the CIFAR-10 dataset using time-lapse diffractive networks, achieving the highest inference accuracy so far. Time-lapse diffractive networks will be widely beneficial for spatiotemporal analysis of input signals using all-optical processors.
ADVANCED INTELLIGENT SYSTEMS
(2023)
Article
Mathematical & Computational Biology
Michael John Fanous, Nir Pillar, Aydogan Ozcan
Summary: Traditional staining methods have drawbacks, while computational virtual staining using deep learning techniques has emerged as a powerful solution. Virtual staining can be combined with neural networks to correct microscopy aberrations and enhance resolution, significantly improving sample preparation and imaging in biomedical microscopy.
FRONTIERS IN BIOINFORMATICS
(2023)