A5100386741- Neural Networks and Reservoir Computing
- Neural Networks and Applications
- Quantum Computing Algorithms and Architecture
- Optical Network Technologies
- Photonic Crystals and Applications
- Random lasers and scattering media
- Advanced Statistical Methods and Models
- Multimodal Machine Learning Applications
- Quantum and electron transport phenomena
- Advanced Materials and Semiconductor Technologies
- Machine Learning in Materials Science
- Animal Vocal Communication and Behavior
- Parallel Computing and Optimization Techniques
- Topological and Geometric Data Analysis
- Gyrotron and Vacuum Electronics Research
- Advanced MRI Techniques and Applications
- Solar Radiation and Photovoltaics
- Quantum Chromodynamics and Particle Interactions
- NMR spectroscopy and applications
- Radio Wave Propagation Studies
- Indoor and Outdoor Localization Technologies
- Electromagnetic Simulation and Numerical Methods
- Plant and animal studies
- Domain Adaptation and Few-Shot Learning
- Quantum Information and Cryptography
Massachusetts Institute of Technology
2016-2021
Meta (United States)
2020
North University of China
2020
State Key Laboratory of Nuclear Physics and Technology
2019
Peking University
2019
IIT@MIT
2019
China University of Labor Relations
2009
Xi'an University of Technology
2007
Symbolic regression is a powerful technique that can discover analytical equations describe data, which lead to explainable models and generalizability outside of the training data set. In contrast, neural networks have achieved amazing levels accuracy on image recognition natural language processing tasks, but are often seen as black-box difficult interpret typically extrapolate poorly. Here we use network-based architecture for symbolic called Equation Learner (EQL) network integrate it...
Deep learning is known to be data-hungry, which hinders its application in many areas of science when data sets are small. Here, we propose use transfer methods migrate knowledge between different physical scenarios and significantly improve the prediction accuracy artificial neural networks trained on a small set. This method can help reduce demand for expensive by making additional inexpensive data. First, demonstrate that, predicting transmission from multilayer photonic film, relative...
Abstract The inability of conventional electronic architectures to efficiently solve large combinatorial problems motivates the development novel computational hardware. There has been much effort toward developing application-specific hardware across many different fields engineering, such as integrated circuits, memristors, and photonics. However, unleashing potential requires algorithms which optimally exploit their fundamental properties. Here, we present Photonic Recurrent Ising Sampler...
We present a novel recurrent neural network (RNN)–based model that combines the remembering ability of unitary evolution RNNs with gated to effectively forget redundant or irrelevant information in its memory. achieve this by extending restricted orthogonal gating mechanism similar unit reset gate and an update gate. Our is able outperform long short-term memory, units, vanilla on several long-term-dependency benchmark tasks. empirically show both lack forget. This plays important role RNNs....
Abstract The prediction and design of photonic features have traditionally been guided by theory-driven computational methods, spanning a wide range direct solvers optimization techniques. Motivated enormous advances in the field machine learning, there has recently growing interest developing complementary data-driven methods for photonics. Here, we demonstrate several predictive generative approaches characterization inverse crystals. Concretely, built data set 20,000 two-dimensional...
We propose a method to use artificial neural networks approximate light scattering by multilayer nanoparticles. find the network needs be trained on only small sampling of data in order simulation high precision. Once is trained, it can simulate such optical processes orders magnitude faster than conventional simulations. Furthermore, used solve nanophotonic inverse design problems using back-propogation - where gradient analytical, not numerical.
Obtaining a full view and complete information of the surrounding dynamics is great significance for plethora applications in sensing, imaging, navigation, orientation. However, conventional spatial spectrum methods heavily rely on priori knowledge with trial‐and‐error solution fashion, leading to challenge estimate volatile scenarios. Inspired by mechanism jumping spider (Salticidae), here universal detection approach driven an intelligent antenna array, usage amplitude‐only as inputs,...
Studying the longitudinally polarized fraction of ${W}^{\ifmmode\pm\else\textpm\fi{}}{W}^{\ifmmode\pm\else\textpm\fi{}}$ scattering at LHC is crucial to examine unitarization mechanism vector boson amplitude through Higgs and possible new physics. We apply here for first time a deep neural network classification extract longitudinal fraction. Based on fast simulation implemented with Delphes framework, significant improvement from found be achievable robust over all dijet mass region. A...
<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Objective:</i> Magnetic Resonance Spectroscopy (MRS) is an important technique for biomedical detection. However, it challenging to accurately quantify metabolites with proton MRS due serious overlaps of metabolite signals, imperfections because non-ideal acquisition conditions, and interference strong background signals mainly from macromolecules. The most popular method, LCModel, adopts...
Stacking long short-term memory (LSTM) cells or gated recurrent units (GRUs) as part of a neural network (RNN) has become standard approach to solving number tasks ranging from language modeling text summarization. Although LSTMs and GRUs were designed model long-range dependencies more accurately than conventional RNNs, they nevertheless have problems copying recalling information the distant past. Here, we derive phase-coded representation state, Rotational Unit Memory (RUM), that unifies...
Abstract At present, it is difficult to detect the photovoltaic (PV) hot spots and recognition efficiency low. In this paper, an improved Single Shot MultiBox Detector (SSD) algorithm was designed for PV spot detection. The used MobileNet network replace VGG16 convolutional neural structure in original SSD. This a depthwise separable convolution structure. Using feature extraction can reduce number of parameters achieve purpose speeding up network. experimental results show that array with...
We present the Photonic Recurrent Ising Sampler (PRIS), an algorithm tailored for photonic parallel networks, that can sample distributions of arbitrary problems. The PRIS finds ground state general problems and probes critical exponents universality classes.
Meta learning methods have found success when applied to few shot classification problems, in which they quickly adapt a small number of labeled examples. Prototypical representations, each representing particular class, been importance this setting, as provide compact form convey information learned from the However, these prototypes are just one method information, and narrow their scope ability classify unseen We propose implementation contextualizers, generalizable that given examples...
Deep learning is known to be data-hungry, which hinders its application in many areas of science when datasets are small. Here, we propose use transfer methods migrate knowledge between different physical scenarios and significantly improve the prediction accuracy artificial neural networks trained on a small dataset. This method can help reduce demand for expensive data by making additional inexpensive data. First, demonstrate that predicting transmission from multilayer photonic film,...
During a prolonged time execution, deep recurrent neural networks suffer from the so-called long-term dependency problem due to their connection. Although Long Short-Term Memory (LSTM) provide temporary solution this problem, they have inferior memory capabilities which limit applications. We use recent approach for network model implementing unitary matrix in its connection deal with dependencies, without affecting abilities. The is capable of high technical results, but insufficient...
We demonstrate an integrated silicon photonic Markov Chain Monte Carlo sampler capable of high-probability convergence to the ground state various 4-spin Ising graphs. Robustness getting trapped in local minima is enhanced by experimental system noise.
We present the Photonic Recurrent Ising Sampler (PRIS), an algorithm tailored for photonic parallel networks, that can sample distributions of arbitrary problems. The PRIS finds ground state general problems and probes critical exponents universality classes. © 2019 Author(s)
Lightning electromagnetic pulse is excitated and may cause instantaneous high voltage or current in electronic equipments through the coupling, which will lead to malfunction even complete destruction. Research on lightning process fast precise calculation of field great significance. A two dimensional cylindrical coordinate novel FDTD (NFDTD) algorithm with scheme proposed this paper order obtain results more precision than traditional (TFDTD) for vertically stratified ground when vertical...