检索结果-内蒙古大学图书馆

A contribution to neuromorphic engineering: neuromodulation implemented through photochromic compounds maintained out of equilibrium by UV-visible radiation

引用

RENDICONTI LINCEI-SCIENZE FISICHE E NATURALI 2020年第1期31卷 39-52页

作者： Bartolomei, Beatrice Heron, B. Mark Gentili, Pier Luigi Univ Perugia Dept Chem Biol & Biotechnol Via Elce di Sotto 8 I-06123 Perugia Italy Univ Huddersfield Sch Appl Sci Dept Chem Sci Huddersfield HD1 3DH W Yorkshire England

Neuromodulation is the alteration of neuronal and synaptic properties in the context of neuronal circuits. It allows anatomically defined circuits to produce multiple outputs reconfiguring networks into different functional circuits. In this work, neuromodulation is mimicked using a solution of two direct photochromic compounds, a naphthopyran and a spirooxazine, which are models of phasic excitable neurons, sensitive to UV radiation. When the system, constituted by the two photochromic compounds, receives a UV signal, it behaves as a recurrent network with mutual inhibitory actions. The network responds to different UV wavelengths by changing its photo-excitability, synaptic strength, wiring of the circuit, and dynamics. These results contribute to the development of neuromorphic engineering. They will promote the design of artificial neural networks with a larger number of nodes, communicating through optical signals. These networks will be the essential ingredients of the new-generation brain-like computing machines complementing current electronic computers.

关键词： Far-from-equilibrium Neurons network Artificial Intelligence Feedback actions Autapses

来源：评论

学校读者我要写书评

暂无评论

End-to-End Delay Minimization based on Joint Optimization of DNN Partitioning and Resource Allocation for Cooperative Edge Inference

End-to-End Delay Minimization based on Joint Optimization of...

引用

IEEE Conference on Vehicular Technology (VTC)

作者： Xinrui Ye Yanzan Sun Dingzhu Wen Guanjin Pan Shunqing Zhang Key Laboratory of Specialty Fiber Optics and Optical Access Networks Shanghai University Shanghai China School of Information Science and Technology ShanghaiTech University Shanghai China

Cooperative inference in Mobile Edge computing (MEC), achieved by deploying partitioned Deep neural network (DNN) models between resource-constrained user equipments (UEs) and edge servers (ESs), has emerged as a promising paradigm. Firstly, we consider scenarios of continuous Artificial Intelligence (AI) task arrivals, like the object detection for video streams, and utilize a serial queuing model for the accurate evaluation of End-to-End (E2E) delay in cooperative edge inference. Secondly, to enhance the long-term performance of inference systems, we formulate a multi-slot stochastic E2E delay optimization problem that jointly considers model partitioning and multi-dimensional resource allocation. Finally, to solve this problem, we introduce a Lyapunov-guided Multi-Dimensional Optimization algorithm (LyMDO) that decouples the original problem into per-slot deterministic problems, where Deep Reinforcement Learning (DRL) and convex optimization are used for joint optimization of partitioning decisions and complementary resource allocation. Simulation results show that our approach effectively improves E2E delay while balancing long-term resource constraints.

关键词：

来源：评论

学校读者我要写书评

暂无评论

Astronomical image denoising based on Convolutional neural network

Astronomical image denoising based on Convolutional Neural N...

引用

IEEE International Conference on Intelligent computing and Information Systems (ICICIS)

作者： Ahmed. A. Elhakiem Tarek Elsaid Ghoniemy Gouda I. Salama Military Technical College Cairo Egypt

ISBN: (纸本)9781665411592

Astronomical images captured using optical telescopes usually suffer from severe noise effects which makes the denoising step inevitable for image analysis. This paper proposes a denoising framework for astronomical images based on Convolutional neural network (Astro U-net). The modified Astro U-net model has been learned in four ways, the first method is using astronomical images from the Hubble Space Telescope data set with three types of noise (dark noise, read- out noise, shot noise) added, the second method is learned using the same data set with the dark noise (dn) added only, the third method is using the same data set with the read-out noise (ron) overlaid, the fourth method is using the same data set with the shot noise (sn) added. The proposed framework for denoising the astronomical images is based on a fusion of the image that was improved by the model learned in the first method with the image that was improved by the three models that were learned by the second, third and fourth methods sequentially. Experimentally, the proposed framework shows a significant improvement in both the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) as compared to the Astro U-net model on different exposure time ratios.

关键词： Adaptation models PSNR Noise reduction optical computing Telescopes optical imaging Data models

来源：评论

学校读者我要写书评

暂无评论

NeurOLight: A Physics-Agnostic neural Operator Enabling Parametric Photonic Device Simulation

arXiv

引用

arXiv 2022年

作者： Gu, Jiaqi Gao, Zhengqi Feng, Chenghao Zhu, Hanqing Chen, Ray T. Boning, Duane S. Pan, David Z. The University of Texas Austin United States Massachusetts Institute of Technology United States

optical computing is an emerging technology for next-generation efficient artificial intelligence (AI) due to its ultra-high speed and efficiency. Electromagnetic field simulation is critical to the design, optimization, and validation of photonic devices and circuits. However, costly numerical simulation significantly hinders the scalability and turn-around time in the photonic circuit design loop. Recently, physics-informed neural networks have been proposed to predict the optical field solution of a single instance of a partial differential equation (PDE) with predefined parameters. Their complicated PDE formulation and lack of efficient parametrization mechanisms limit their flexibility and generalization in practical simulation scenarios. In this work, for the first time, a physics-agnostic neural operator-based framework, dubbed NeurOLight, is proposed to learn a family of frequency-domain Maxwell PDEs for ultra-fast parametric photonic device simulation. We balance the efficiency and generalization of NeurOLight via several novel techniques. Specifically, we discretize different devices into a unified domain, represent parametric PDEs with a compact wave prior, and encode the incident light via masked source modeling. We design our model with parameter-efficient cross-shaped NeurOLight blocks and adopt superposition-based augmentation for data-efficient learning. With these synergistic approaches, NeurOLight generalizes to a large space of unseen simulation settings, demonstrates 2-orders-of-magnitude faster simulation speed than numerical solvers, and outperforms prior neural network models by ∼54% lower prediction error with ∼44% fewer parameters. Our code is available at link. Copyright © 2022, The Authors. All rights reserved.

关键词： Photonic devices

来源：评论

学校读者我要写书评

暂无评论

ROBIN: A Robust optical Binary neural network Accelerator

引用

ACM TRANSACTIONS ON EMBEDDED computing SYSTEMS 2021年第5期20卷 57-57页

作者： Sunny, Febin P. Mirza, Asif Nikdast, Mahdi Pasricha, Sudeep Colorado State Univ Dept Elect & Comp Engn Ft Collins CO 80523 USA

Domain specific neural network accelerators have garnered attention because of their improved energy efficiency and inference performance compared to CPUs and GPUs. Such accelerators are thus well suited for resource-constrained embedded systems. However, mapping sophisticated neural network models on these accelerators still entails significant energy and memory consumption, along with high inference time overhead. Binarized neural networks (BNNs), which utilize single-bit weights, represent an efficient way to implement and deploy neural network models on accelerators. In this paper, we present a novel opticaldomain BNN accelerator, named ROBIN, which intelligently integrates heterogeneous microring resonator optical devices with complementary capabilities to efficiently implement the key functionalities in BNNs. We perform detailed fabrication-process variation analyses at the optical device level, explore efficient corrective tuning for these devices, and integrate circuit-level optimization to counter thermal variations. As a result, our proposed ROBIN architecture possesses the desirable traits of being robust, energy-efficient, low latency, and high throughput, when executing BNN models. Our analysis shows that ROBIN can outperform the best-known optical BNN accelerators and many electronic accelerators. Specifically, our energy-efficient ROBIN design exhibits energy-per-bit values that are similar to 4 x lower than electronic BNN accelerators and similar to 933 x lower than a recently proposed photonic BNN accelerator, while a performance-efficient ROBIN design shows similar to 3 x and similar to 25 x better performance than electronic and photonic BNN accelerators, respectively.

关键词： Silicon photonics binarized neural networks inference acceleration design optimization

来源：评论

学校读者我要写书评

暂无评论

Analytical Prediction of Scattering Properties of Spheroidal Dust Particles With Machine Learning

引用

GEOPHYSICAL RESEARCH LETTERS 2022年第9期49卷 e2021GL097548-e2021GL097548页

作者： Chen, Xi Wang, Jun Gomes, Joe Dubovik, Oleg Yang, Ping Saito, Masanori Univ Iowa Iowa Technol Inst Dept Chem & Biochem Engn Iowa City IA 52242 USA Univ Iowa Ctr Global & Reg Environm Res Iowa City IA 52242 USA Univ Iowa Dept Phys & Astron Iowa City IA 52242 USA Univ Lille CNRS Lab dOpt Atmospher Lille France Texas A&M Univ Dept Atmospher Sci College Stn TX USA Texas A&M Univ Dept Oceanog College Stn TX USA Texas A&M Univ Dept Phys & Astron College Stn TX USA

A neural network (NN) model is trained with a database widely used in the aerosol remote sensing community to rapidly predict the single-scattering optical properties of spheroidal dust particles. Analytical solutions for their Jacobians with respect to microphysical properties are derived based on the functional form of the NN. The Jacobian predictions are improved by adding Jacobians from a linearized T-matrix model into the training. Out-of-database testing implies that NN-based predictions perform better than the business-as-usual method that interpolates optical properties from the database. Independent validation further demonstrates the efficacy of the NN-based predictions by reducing computational costs while maintaining accuracy. This work represents the first use of machine learning-based function approximation to computationally expedite the application of the existing spheroidal dust properties database;the resultant NN model can be implemented in atmospheric models and satellite retrieval algorithms with high accuracy, computational efficiency, and the rigor of analytical solutions. Plain Language Summary Dust particles affect both solar and terrestrial radiative transfer, but whether they cool or warm the climate is currently an open question in the literature. Accurate estimation of dust scattering and absorption properties, while critical for climate studies, is hindered by the fact that dust particles have irregular shapes and large size ranges;hence, no single method can be applied for all particle sizes and shapes. Often, a comprehensive look-up table of these properties is created by combining multiple methods. The application of such databases, however, is cumbersome and inaccurate due to the need for multi-variable interpolation. Furthermore, the look-up-table approach lacks the mathematical rigor needed to determine the sensitivity (Jacobians) of the single-scattering properties to the dust size, shape, and refractive index that are needed in

关键词： nonspherical dust particle machine learning scattering linearization and Jacobians neural network T-matrix

来源：评论

学校读者我要写书评

暂无评论

Resource planning on elastic optical networks using traffic matrix prediction

引用

AEU-INTERNATIONAL JOURNAL OF ELECTRONICS AND COMMUNICATIONS 2021年 134卷 153615-153615页

作者： Mesquita, Leonardo A. J. Assis, Karcius D. R. Pinho, Luana V. B. C. Almeida Jr, Raul C. Simas Filho, Eduardo F. Alencar, Marcelo S. Univ Fed Bahia Dept Elect & Comp Engn Salvador BA Brazil Univ Fed Pernambuco Elect & Syst Dept Photon Grp Recife PE Brazil

The rapidly increasing data demand from current Internet services, such as cloud computing and high-quality video streaming, is raising the pressure on network operators to provide reliable high-speed connections while keeping costs low. Without having to rely on the bandwidth expansion of optical waveguides or modulation level efficiency, one of the ways to increase spectral efficiency of a network is by the optimal utilization of already existing resources. For elastic optical networks (EON) this can be solved by the Routing, Modulation and Spectrum Assignment (RMSA). However, another problem arises, due to the nature of computer network traffic, that makes it difficult to assess the longevity of physical components and to plan a network to serve a maximum amount of bandwidth demand for predetermined time span. To help alleviating these problems, this paper presents an investigation of a novel technique for resource planning and consume estimation based on simulations. First, traffic matrix prediction is made using recurrent neural networks (RNN), and then simulations are ran to estimate the consumed bandwidth for a point in time. For the traffic forecasting, real traffic data has been used to train the RNN, obtained from an anonymous dataset containing the traffic history of the ABILENE and GEANT networks from the years 2004-2005, which is a relatively recent available dataset. Furthermore, heuristic algorithms are proposed to apply the RMSA to the predicted traffic, in a quasi-optimal manner, in order to minimize the use of network resources for future traffic, by releasing available capacity in the existing optical fiber links in an incremental traffic approach. The results show that the proposed RNN models are indeed capable of predicting the traffic matrices a month ahead, with low mean squared error. RMSA simulations were also performed using heuristics in order to estimate the consumed bandwidth in the context of EON networks, revealing that, for the propos

关键词： optical networks Routing Modulation and spectrum assignment Traffic matrix prediction Artificial neural networks Multiperiod network planning

来源：评论

学校读者我要写书评

暂无评论

neural networks for computing and denoising the continuous nonlinear Fourier spectrum in focusing nonlinear Schrodinger equation

引用

SCIENTIFIC REPORTS 2021年第1期11卷 1-15页

作者： Sedov, Egor V. Freire, Pedro J. Seredin, Vladimir V. Kolbasin, Vladyslav A. Kamalian-Kopae, Morteza Chekhovskoy, Igor S. Turitsyn, Sergei K. Prilepsky, Jaroslaw E. Aston Univ Aston Inst Photon Technol Birmingham B4 7ET W Midlands England Novosibirsk State Univ Novosibirsk 630090 Russia Natl Tech Univ Kharkiv Polytech Inst UA-61102 Kharkiv Ukraine Federal Res Ctr Informat & Computat Technol Novosibirsk 630090 Russia

We combine the nonlinear Fourier transform (NFT) signal processing with machine learning methods for solving the direct spectral problem associated with the nonlinear Schrodinger equation. The latter is one of the core nonlinear science models emerging in a range of applications. Our focus is on the unexplored problem of computing the continuous nonlinear Fourier spectrum associated with decaying profiles, using a specially-structured deep neural network which we coined NFT-Net. The Bayesian optimisation is utilised to find the optimal neural network architecture. The benefits of using the NFT-Net as compared to the conventional numerical NFT methods becomes evident when we deal with noise-corrupted signals, where the neural networks-based processing results in effective noise suppression. This advantage becomes more pronounced when the noise level is sufficiently high, and we train the neural network on the noise-corrupted field profiles. The maximum restoration quality corresponds to the case where the signal-to-noise ratio of the training data coincides with that of the validation signals. Finally, we also demonstrate that the NFT b-coefficient important for optical communication applications can be recovered with high accuracy and denoised by the neural network with the same architecture.

关键词：

来源：评论

学校读者我要写书评

暂无评论

Deep Learning-Assisted Design of Integrated $2\times 2$ Linear optical Processors

Deep Learning-Assisted Design of Integrated $2\times 2$ Line...

引用

Asia Communications and Photonics Conference and Exhibition (ACP)

作者： Simei Mao Lirong Cheng Faisal Nadeem Khan H. Y. Fu Tsinghua-Berkeley Shenzhen Institute & Tsinghua Shenzhen International Graduate School Tsinghua University Shenzhen China

We propose a joint deep neural network that both instantaneously predicts the transfer matrices of Q R -code patterned optical processors and inversely designs optical processors for target transfer matrix with over 8... 详细信息

ISBN: (纸本)9781665410953

关键词： Integrated optics Training Program processors optical design neural networks optical computing Predictive models

来源：评论

学校读者我要写书评

暂无评论

Lightweight Image Super-Resolution with Multi-Residual Feature Aggregation network

SSRN

引用

SSRN 2022年

作者： Qin, Jiayi Yang, Xiaomin College of Electronics and Information Engineering Sichuan University Sichuan Chengdu610065 China School of Aeronautics & Astronautics Sichuan University Sichuan Chengdu610065 China

Recently, single image super-resolution (SISR) methods using deep convolution neural networks (CNNs) have achieved remarkable performance. Especially, lightweight networks have received unprecedented attention because of their broad application prospects. However, existing methods for lightweight SR lack the adequate utilization of hierarchical features, which weakens the representation ability of the network. To alleviate this issue, we propose an effective and accurate multi-residual feature aggregation network (MrFAN) for SISR. Specifically, we design a multi-residual block (MRB) to boost the representation ability of the network. By adopting the multi-residual learning (MRL) strategy, MRB can efficiently improve reconstruction results with slight computational resources, compared with the ordinary residual block (RB). To use the hierarchical features sufficiently, we construct a multi-residual aggregation block (MRAB) with a multi-residual feature aggregation (MRFA) strategy. Utilizing the MRFA, informative high-frequency features, extracted by the hierarchical feature fusion (HRFF), are refined through hierarchical feature aggregation (HFA) at deep layers. Finally, we build our MrFAN by densely stacking MRABs and introducing double-residual learning (DRL) strategy into the network at the global level. Extensive experiments demonstrate that the MrFAN is superior to the state-of-the-art SISR models while taking up less computing resources. © 2022, The Authors. All rights reserved.

关键词： optical resolving power

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：