Accelerating the training of feedforward neural networks usinggeneralized Hebbian rules for initializing the internal representations

This paper presents an unsupervised learning scheme for initializing the internal representations of feedforward neural networks, which accelerates the convergence of supervised learning algorithms. It is proposed in this paper that the initial set of internal representations can be formed through a bottom-up unsupervised learning process applied before the top-down supervised training algorithm. The synaptic weights that connect the input of the network with the hidden units can be determined through linear or nonlinear variations of a generalized Hebbian learning rule, known as Oja's rule. Various generalized Hebbian rules were experimentally tested and evaluated in terms of their effect on the convergence of the supervised training process. Several experiments indicated that the use of the proposed initialization of the internal representations significantly improves the convergence of gradient-descent-based algorithms used to perform nontrivial training tasks. The improvement of the convergence becomes significant as the size and complexity of the training task increase.     

Supervised extended ART: A fast neural network classifier trained by combining supervised and unsupervised learning

A neural network classifier, called supervised extended ART (SEART), that incorporates a supervised mechanism into the extended unsupervised ART is presented here. It uses a learning theory called Nested Generalized Exemplar (NGE) theory. In any time, the training instances may or may not have desired outputs, that is, this model can handle supervised learning and unsupervised learning simultaneously. The unsupervised component finds the cluster relations of instances, and the supervised component learns the desired associations between clusters and classes. In addition, this model has the ability of incremental learning. It works equally well when instances in a cluster belong to different classes. Also, multi-category and nonconvex classifications can be dealt with. Besides, the experimental results are very encouraging.     

Cross-domain person re-identification by hybrid supervised and unsupervised learning

Although the single-domain person re-identification (Re-ID) method has achieved great accuracy, the dependence on the label in the same image domain severely limits the scalability of this method. Therefore, cross-domain Re-ID has received more and more attention. In this paper, a novel cross-domain Re-ID method combining supervised and unsupervised learning is proposed, which includes two models: a triple-condition generative adversarial network (TC-GAN) and a dual-task feature extraction network (DFE-Net). We first use TC-GAN to generate labeled images with the target style, and then we combine supervised and unsupervised learning to optimize DFE-Net. Specifically, we use labeled generated data for supervised learning. In addition, we mine effective information in the target data from two perspectives for unsupervised learning. To effectively combine the two types of learning, we design a dynamic weighting function to dynamically adjust the weights of these two approaches. To verify the validity of TC-GAN, DFE-Net, and the dynamic weight function, we conduct multiple experiments on Market-1501 and DukeMTMC-reID. The experimental results show that the dynamic weight function can improve the performance of the models, and our method is better than many state-of-the-art methods.

     

Deep and fast: Deep learning hashing with semi-supervised graph construction

Learning-based hashing methods are becoming the mainstream for approximate scalable multimedia retrieval. They consist of two main components: hash codes learning for training data and hash functions learning for new data points. Tremendous efforts have been devoted to designing novel methods for these two components, i.e., supervised and unsupervised methods for learning hash codes, and different models for inferring hashing functions. However, there is little work integrating supervised and unsupervised hash codes learning into a single framework. Moreover, the hash function learning component is usually based on hand-crafted visual features extracted from the training images. The performance of a content-based image retrieval system crucially depends on the feature representation and such hand-crafted visual features may degrade the accuracy of the hash functions. In this paper, we propose a semi-supervised deep learning hashing (DLH) method for fast multimedia retrieval. More specifically, in the first component, we utilize both visual and label information to learn an optimal similarity graph that can more precisely encode the relationship among training data, and then generate the hash codes based on the graph. In the second stage, we apply a deep convolutional network to simultaneously learn a good multimedia representation and a set of hash functions. Extensive experiments on five popular datasets demonstrate the superiority of our DLH over both supervised and unsupervised hashing methods.     

一种基于前向无监督卷积神经网络的人脸表示学习方法

当前基于深度卷积神经网络的人脸表示学习方法需要利用海量的有标注的人脸数据。在实际应用中,精确标注人脸的身份非常困难。因此,提出了一种基于前向无监督卷积神经网络的人脸表示学习方法。其中,基于K-means聚类获取训练样本虚拟标签,再利用线性判别分析进行卷积核学习。提出的网络结构简单有效,训练阶段不需要反向传递,训练速度显著优于有监督的深度卷积神经网络。实验结果表明,提出的方法在真实条件下的人脸数据集LFW和经典的Feret数据集上取得了优于当前主流的无监督特征学习方法和局部特征描述子的性能。     

Unsupervised Dense Light Field Reconstruction with Occlusion Awareness

Light field (LF) reconstruction is a fundamental technique in light field imaging and has applications in both software and hardware aspects. This paper presents an unsupervised learning method for LF‐oriented view synthesis, which provides a simple solution for generating quality light fields from a sparse set of views. The method is built on disparity estimation and image warping. Specifically, we first use per‐view disparity as a geometry proxy to warp input views to novel views. Then we compensate the occlusion with a network by a forward‐backward warping process. Cycle‐consistency between different views are explored to enable unsupervised learning and accurate synthesis. The method overcomes the drawbacks of fully supervised learning methods that require large labeled training dataset and epipolar plane image based interpolation methods that do not make full use of geometry consistency in LFs. Experimental results demonstrate that the proposed method can generate high quality views for LF, which outperforms unsupervised approaches and is comparable to fully‐supervised approaches.     

Noise injection into inputs in sparsely connected Hopfield andwinner-take-all neural networks

In this paper, we show that noise injection into inputs in unsupervised learning neural networks does not improve their performance as it does in supervised learning neural networks. Specifically, we show that training noise degrades the classification ability of a sparsely connected version of the Hopfield neural network, whereas the performance of a sparsely connected winner-take-all neural network does not depend on the injected training noise.     

Improved competitive learning neural networks for network intrusion and fraud detection

In this research, we propose two new clustering algorithms, the improved competitive learning network (ICLN) and the supervised improved competitive learning network (SICLN), for fraud detection and network intrusion detection. The ICLN is an unsupervised clustering algorithm, which applies new rules to the standard competitive learning neural network (SCLN). The network neurons in the ICLN are trained to represent the center of the data by a new reward-punishment update rule. This new update rule overcomes the instability of the SCLN. The SICLN is a supervised version of the ICLN. In the SICLN, the new supervised update rule uses the data labels to guide the training process to achieve a better clustering result. The SICLN can be applied to both labeled and unlabeled data and is highly tolerant to missing or delay labels. Furthermore, the SICLN is capable to reconstruct itself, thus is completely independent from the initial number of clusters.To assess the proposed algorithms, we have performed experimental comparisons on both research data and real-world data in fraud detection and network intrusion detection. The results demonstrate that both the ICLN and the SICLN achieve high performance, and the SICLN outperforms traditional unsupervised clustering algorithms.     

A novel ensemble learning approach to unsupervised record linkage

Record linkage is a process of identifying records that refer to the same real-world entity. Many existing approaches to record linkage apply supervised machine learning techniques to generate a classification model that classifies a pair of records as either match or non-match. The main requirement of such an approach is a labelled training dataset. In many real-world applications no labelled dataset is available hence manual labelling is required to create a sufficiently sized training dataset for a supervised machine learning algorithm. Semi-supervised machine learning techniques, such as self-learning or active learning, which require only a small manually labelled training dataset have been applied to record linkage. These techniques reduce the requirement on the manual labelling of the training dataset. However, they have yet to achieve a level of accuracy similar to that of supervised learning techniques. In this paper we propose a new approach to unsupervised record linkage based on a combination of ensemble learning and enhanced automatic self-learning. In the proposed approach an ensemble of automatic self-learning models is generated with different similarity measure schemes. In order to further improve the automatic self-learning process we incorporate field weighting into the automatic seed selection for each of the self-learning models. We propose an unsupervised diversity measure to ensure that there is high diversity among the selected self-learning models. Finally, we propose to use the contribution ratios of self-learning models to remove those with poor accuracy from the ensemble. We have evaluated our approach on 4 publicly available datasets which are commonly used in the record linkage community. Our experimental results show that our proposed approach has advantages over the state-of-the-art semi-supervised and unsupervised record linkage techniques. In 3 out of 4 datasets it also achieves comparable results to those of the supervised approaches.     

Query-based learning for aerospace applications

Models of real-world applications often include a large number of parameters with a wide dynamic range, which contributes to the difficulties of neural network training. Creating the training data set for such applications becomes costly, if not impossible. In order to overcome the challenge, one can employ an active learning technique known as query-based learning (QBL) to add performance-critical data to the training set during the learning phase, thereby efficiently improving the overall learning/generalization. The performance-critical data can be obtained using an inverse mapping called network inversion (discrete network inversion and continuous network inversion) followed by oracle query. This paper investigates the use of both inversion techniques for QBL learning, and introduces an original heuristic to select the inversion target values for continuous network inversion method. Efficiency and generalization was further enhanced by employing node decoupled extended Kalman filter (NDEKF) training and a causality index (CI) as a means to reduce the input search dimensionality. The benefits of the overall QBL approach are experimentally demonstrated in two aerospace applications: a classification problem with large input space and a control distribution problem.     

基于强化学习的股票预测系统的研究与设计

股票市场是金融分析领域中重要而困难的问题。股票数据的分析和预测具有重大的理论意义和诱人的应用价值。BP神经网络在目前的股票预测系统中应用广泛,但是作为有导师的学习系统,BP神经网络必须要求提供相关的经验数据才能正常运行。对此本文提出了一种基于强化学习BP算法应用于股票预测系统,通过强化学习体系来实现体统的自学习,通过网络集成来达到初始数据的预处理,提高系统的泛化能力,在实际应用中取的较好的效果。     

Training recurrent neural networks using a hybrid algorithm

This paper proposes a new hybrid approach for recurrent neural networks (RNN). The basic idea of this approach is to train an input layer by unsupervised learning and an output layer by supervised learning. In this method, the Kohonen algorithm is used for unsupervised learning, and dynamic gradient descent method is used for supervised learning. The performances of the proposed algorithm are compared with backpropagation through time (BPTT) on three benchmark problems. Simulation results show that the performances of the new proposed algorithm exceed the standard backpropagation through time in the reduction of the total number of iterations and in the learning time required in the training process.     

多标记学习自编码网络无监督维数约简

多标记学习是针对一个实例同时与一组标签相关联而提出的一种机器学习框架,是该领域研究热点之一,降维是多标记学习一个重要且具有挑战性的工作。针对有监督的多标记维数约简方法,提出一种无监督自编码网络的多标记降维方法。首先,通过构建自编码神经网络,对输入数据进行编码和解码输出;然后,引入稀疏约束计算总体成本,使用梯度下降法进行迭代求解;最后,通过深度学习训练获得自编码网络学习模型,提取数据特征实现维数约简。实验中使用多标记算法ML-kNN做分类器,在6个公开数据集上与其他4种方法对比。实验结果表明,该方法能够在不使用标记的情况下有效提取特征,降低多标记数据维度,稳定提高多标记学习性能。     

A Hybrid Training Algorithm for Feedforward Neural Networks

Feedforward neural networks (FNN) have been proposed to solve complex problems in pattern recognition, classification and function approximation. Despite the general success of learning methods for FNN, such as the backpropagation (BP) algorithm, second-order algorithms, long learning time for convergence remains a problem to be overcome. In this paper, we propose a new hybrid algorithm for a FNN that combines unsupervised training for the hidden neurons (Kohonen algorithm) and supervised training for the output neurons (gradient descent method). Simulation results show the effectiveness of the proposed algorithm compared with other well-known learning methods.     

弱监督分层深度学习的车辆识别算法

针对已有分类器在结构形式和训练方法的不足,构建了一个以二维深度置信网络（2D deep belief networks,2D DBN）为架构的弱监督分层深度学习车辆识别算法。首先,将传统一维的深度置信网络(Deep belief networks,DBN)扩展成2D-DBN,并构建相应分类器结构,从而能够直接以二维图像像素矩阵作为输入; 其次,在传统无监督训练的目标函数中,引入了一个具有适当权重的判别度正则化项,将原有无监督训练转化为带有较弱监督性的弱监督训练方式,从而使提取的特征较传统无监督特征更具判别性。多组对比实验表明,本文所提算法在识别率等指标上要优于已有深度学习算法。     

A reinforcement learning approach based on the fuzzy min-max neural network

The fuzzy min-max neural network constitutes a neural architecture that is based on hyperbox fuzzy sets and can be incrementally trained by appropriately adjusting the number of hyperboxes and their corresponding volumes. Two versions have been proposed: for supervised and unsupervised learning. In this paper a modified approach is presented that is appropriate for reinforcement learning problems with discrete action space and is applied to the difficult task of autonomous vehicle navigation when no a priori knowledge of the enivronment is available. Experimental results indicate that the proposed reinforcement learning network exhibits superior learning behavior compared to conventional reinforcement schemes.     

Self-supervised segmentation using synthetic datasets via L-system

Vessel segmentation plays a crucial role in the diagnosis of many diseases, as well as assisting surgery.With the development of deep learning, many segmentation methods have been proposed, and the results have become more and more accurate. However, most of these methods are based on supervised learning, which require a large amount of labeled data as training data. To overcome this shortcoming, unsupervised and self-supervisedmethods have also received increasing attention. In this paper, we generate a synthetic training datasets through L-system, and utilize adversarial learning to narrow the distribution difference between the generated data and the real data to obtain the ultimate network. Our method achieves state-of-the-art (SOTA) results on X-ray angiography artery disease (XCAD) by a large margin of nearly 10.4%.     

Unsupervised Learning for Graph Matching

Graph matching is an essential problem in computer vision that has been successfully applied to 2D and 3D feature matching and object recognition. Despite its importance, little has been published on learning the parameters that control graph matching, even though learning has been shown to be vital for improving the matching rate. In this paper we show how to perform parameter learning in an unsupervised fashion, that is when no correct correspondences between graphs are given during training. Our experiments reveal that unsupervised learning compares favorably to the supervised case, both in terms of efficiency and quality, while avoiding the tedious manual labeling of ground truth correspondences. We verify experimentally that our learning method can improve the performance of several state-of-the art graph matching algorithms. We also show that a similar method can be successfully applied to parameter learning for graphical models and demonstrate its effectiveness empirically.     

A unifying methodology for the evaluation of neural network models on novelty detection tasks

An important issue in data analysis and pattern classification is the detection of anomalous observations and its influence on the classifier’s performance. In this paper, we introduce a novel methodology to systematically compare the performance of neural network (NN) methods applied to novelty detection problems. Initially, we describe the most common NN-based novelty detection techniques. Then we generalize to the supervised case, a recently proposed unsupervised novelty detection method for computing reliable decision thresholds. We illustrate how to use the proposed methodology to evaluate the performances of supervised and unsupervised NN-based novelty detectors on a real-world benchmarking data set, assessing their sensitivity to training parameters, such as data scaling, number of neurons, training epochs and size of the training set.     

Growing radial basis neural networks: merging supervised andunsupervised learning with network growth techniques

This paper proposes a framework for constructing and training radial basis function (RBF) neural networks. The proposed growing radial basis function (GRBF) network begins with a small number of prototypes, which determine the locations of radial basis functions. In the process of training, the GRBF network gross by splitting one of the prototypes at each growing cycle. Two splitting criteria are proposed to determine which prototype to split in each growing cycle. The proposed hybrid learning scheme provides a framework for incorporating existing algorithms in the training of GRBF networks. These include unsupervised algorithms for clustering and learning vector quantization, as well as learning algorithms for training single-layer linear neural networks. A supervised learning scheme based on the minimization of the localized class-conditional variance is also proposed and tested. GRBF neural networks are evaluated and tested on a variety of data sets with very satisfactory results.