Deep Matrix Factorization Based on Convolutional Neural Networks for Image Inpainting

In this work, we formulate the image in-painting as a matrix completion problem. Traditional matrix completion methods are generally based on linear models, assuming that the matrix is low rank. When the original matrix is large scale and the observed elements are few, they will easily lead to over-fitting and their performance will also decrease significantly. Recently, researchers have tried to apply deep learning and nonlinear techniques to solve matrix completion. However, most of the existing deep learning-based methods restore each column or row of the matrix independently, which loses the global structure information of the matrix and therefore does not achieve the expected results in the image in-painting. In this paper, we propose a deep matrix factorization completion network (DMFCNet) for image in-painting by combining deep learning and a traditional matrix completion model. The main idea of DMFCNet is to map iterative updates of variables from a traditional matrix completion model into a fixed depth neural network. The potential relationships between observed matrix data are learned in a trainable end-to-end manner, which leads to a high-performance and easy-to-deploy nonlinear solution. Experimental results show that DMFCNet can provide higher matrix completion accuracy than the state-of-the-art matrix completion methods in a shorter running time.     

Particle Identification using Artificial Neural Networks at BESIII

A multilayered perceptrons neural network technique has been applied in the particle identification at BESIII. The networks are trained in each sub-detector level. The NN output of sub-detectors can be sent to a sequential network or be constructed as PDFs for a likelihood. Good muon-ID, electron-ID and hadron-ID are obtained from the networks by using the simulated Monte Carlo samples.     

Algorithm for Processing and Analysis of Raman Spectra using Neural Networks

The solution of the problem of processing of a large data set when analyzing Raman spectra of a gas mixture is considered. The algorithm is based on the artificial neural network. Conditions for the use of neural networks in solving practical problems of real-time analyzing spectra, including that for remote search for heavy hydrocarbons are determined. The algorithm speed is estimated using computer aids with sequential and parallel data processing.     

A novel design of Gaussian Wavelet Neural Networks for nonlinear Falkner-Skan systems in fluid dynamics

In this study, a design of integrated computational intelligent paradigm has been presented for numerical treatment of the one-dimensional boundary value problems represented with Falkner-Skan equations (FSE) by exploitation of Gaussian wavelet neural networks (GWNNs), genetic algorithms (GAs) and sequential quadratic programming (SQP), i.e., GWNN-GA-SQP. The GWNNs is used for mathematical modeling of the problem by constructing mean squared error based objective function while optimization of the cost function is initially conducted with efficacy of GAs as a global search and while fine tuning is performed with efficiency local search with SQP. The numerical results are obtained by proposed GWNN-GA-SQP for different FSEs arising in nonlinear regimes of computation fluid mechanics studies. A comparison of the results of proposed GWNN-GA-SQP stochastic numerical solver with reference state of the art solutions of Adams method establishes the accuracy, convergence and stability, which further endorsed through statistics on multiples runs. The T-Paired test is also applied to validate the effectiveness of the proposed GWNN-GA-SQP algorithm for solving nonlinear FSEs.     

Parallel Subtraction using LCLV and Lenslet Array in Optical Neural Networks

Basing on the method of wavelength and beam deviation encoding for implementing a bipolar optical neural networks, the subtraction between the positive and negative stimulation of a neuron can be converted to the addition by the normal and in verse states of a liquid crystal light value (LCLV). The summation can be performed by using a lenslet to focus the two light spots which represent the two stimulations. The pre liminary experimental results demonstrate the feasibility of the presental method.     

神经网络型光计算

本文综述了近年来在神经网络型光计算方面国外的研究情况,简单地介绍了Hopfield神经网络模型,较详细地介绍了该模型的几种光学模拟方法,包括一维处理和二维处理。     

Probabilistic Models with Deep Neural Networks

Recent advances in statistical inference have significantly expanded the toolbox of probabilistic modeling. Historically, probabilistic modeling has been constrained to very restricted model classes, where exact or approximate probabilistic inference is feasible. However, developments in variational inference, a general form of approximate probabilistic inference that originated in statistical physics, have enabled probabilistic modeling to overcome these limitations: (i) Approximate probabilistic inference is now possible over a broad class of probabilistic models containing a large number of parameters, and (ii) scalable inference methods based on stochastic gradient descent and distributed computing engines allow probabilistic modeling to be applied to massive data sets. One important practical consequence of these advances is the possibility to include deep neural networks within probabilistic models, thereby capturing complex non-linear stochastic relationships between the random variables. These advances, in conjunction with the release of novel probabilistic modeling toolboxes, have greatly expanded the scope of applications of probabilistic models, and allowed the models to take advantage of the recent strides made by the deep learning community. In this paper, we provide an overview of the main concepts, methods, and tools needed to use deep neural networks within a probabilistic modeling framework.     

Bootstrap Percolation in Living Neural Networks

Recent experimental studies of living neural networks reveal that their global activation induced by electrical stimulation can be explained using the concept of bootstrap percolation on a directed random network. The experiment consists in activating externally an initial random fraction of the neurons and observe the process of firing until its equilibrium. The final portion of neurons that are active depends in a non linear way on the initial fraction. The main result of this paper is a theorem which enables us to find the final proportion of the fired neurons, in the asymptotic case, in the case of random directed graphs with given node degrees as the model for interacting network. This gives a rigorous mathematical proof of a phenomena observed by physicists in neural networks.     

Winsorization for Robust Bayesian Neural Networks

With the advent of big data and the popularity of black-box deep learning methods, it is imperative to address the robustness of neural networks to noise and outliers. We propose the use of Winsorization to recover model performances when the data may have outliers and other aberrant observations. We provide a comparative analysis of several probabilistic artificial intelligence and machine learning techniques for supervised learning case studies. Broadly, Winsorization is a versatile technique for accounting for outliers in data. However, different probabilistic machine learning techniques have different levels of efficiency when used on outlier-prone data, with or without Winsorization. We notice that Gaussian processes are extremely vulnerable to outliers, while deep learning techniques in general are more robust.     

Discontinuity and Complex Dynamics of Neural Networks

We focus on the discontinuity of a neural network model with diluted and clipped synaptic connections (±l only). The exact evolution rule of the average firing rate becomes a discontinuous piece-wise nonlinear map when very simple functions of dynamical threshold are introduced into the network. Complex dynamics is observed.     

Optimization of Multilayer Photonic Structures using Artificial Neural Networks to Obtain a Target Optical Response

JETP Letters - A new deep machine learning method is proposed for the task of selecting the parameters of a multilayer photonic structure to obtain a target optical spectrum of the reflection...     

非线性振子的矩阵元及其经典近似

对于势能为V(x)=1/2 mω2x2+λx4的非线性谐振子,不能用微扰论对经典方程进行求解.这里利用海森伯对应原理,由量子力学的矩阵元得到了非线性振子的经典解,从而对于非线性振子的性质有了进一步的理解.     

Neural Networks for Estimating Speculative Attacks Models

Currency crises have been analyzed and modeled over the last few decades. These currency crises develop mainly due to a balance of payments crisis, and in many cases, these crises lead to speculative attacks against the price of the currency. Despite the popularity of these models, they are currently shown as models with low estimation precision. In the present study, estimates are made with first- and second-generation speculative attack models using neural network methods. The results conclude that the Quantum-Inspired Neural Network and Deep Neural Decision Trees methodologies are shown to be the most accurate, with results around 90% accuracy. These results exceed the estimates made with Ordinary Least Squares, the usual estimation method for speculative attack models. In addition, the time required for the estimation is less for neural network methods than for Ordinary Least Squares. These results can be of great importance for public and financial institutions when anticipating speculative pressures on currencies that are in price crisis in the markets.     

Densely Connected Neural Networks for Nonlinear Regression

Densely connected convolutional networks (DenseNet) behave well in image processing. However, for regression tasks, convolutional DenseNet may lose essential information from independent input features. To tackle this issue, we propose a novel DenseNet regression model where convolution and pooling layers are replaced by fully connected layers and the original concatenation shortcuts are maintained to reuse the feature. To investigate the effects of depth and input dimensions of the proposed model, careful validations are performed by extensive numerical simulation. The results give an optimal depth (19) and recommend a limited input dimension (under 200). Furthermore, compared with the baseline models, including support vector regression, decision tree regression, and residual regression, our proposed model with the optimal depth performs best. Ultimately, DenseNet regression is applied to predict relative humidity, and the outcome shows a high correlation with observations, which indicates that our model could advance environmental data science.     

Phase Space Prediction Model Using Neural Networks

A new nonlinear prediction technique is proposed by feedforward neural network, the learning algorithmfor network is a chaotic one. A timc-delay embedding is used to reconstruct the underlying attractor, the predictionmodel is based on the time evolution of the topological neighboring in the phase space, the spatial neighbors are chosenby the rate of exponential divergence of close trajectory. The model is tested for the Mackey-Glass delay equation andLorentz equations, good results are obtained for the prediction.     

Strongly Diluted Neural Networks with Correlation-Causing Stimuli

The strongly diluted neural networks are studied rigorously under correlation-causing stimuli for any time step with a Gaussian distribution of stabilities. For the second time step, the analytical formulas of retrieval qualities are derived by taking the correlation between random variables into consideration. Numerical results show that the correlation deteriorates the improvement of the retrieval quality by persistent stimuli in comparison with the case neglecting it, but it is, at least for t = 2, not strong enough to cause essential discrepancy between the retrieval qualities for the networks with non-correlation- and correlation-causing stimuli respectively.