Evolving modular networks with genetic algorithms: application to nonlinear time series

Abstract: A key problem of modular neural networks is finding the optimal aggregation of the different subtasks (or modules) of the problem at hand. Functional networks provide a partial solution to this problem, since the inter‐module topology is obtained from domain knowledge (functional relationships and symmetries). However, the learning process may be too restrictive in some situations, since the resulting modules (functional units) are assumed to be linear combinations of selected families of functions. In this paper, we present a non‐parametric learning approach for functional networks using feedforward neural networks for approximating the functional modules of the resulting architecture; we also introduce a genetic algorithm for finding the optimal intra‐module topology (the appropriate balance of neurons for the different modules according to the complexity of their respective tasks). Some benchmark examples from nonlinear time‐series prediction are used to illustrate the performance of the algorithm for finding optimal modular network architectures for specific problems.     

基于异质信息网络的推荐模型

为了解决推荐系统的冷启动和稀疏性问题, 本文提出了一种基于异质信息网络的推荐模型. 传统的推荐方法无法在知识图谱表示学习中融入隐含的路径信息, 这样使得知识推荐系统性能较为一般. 本文提出的模型在异质信息网络中设置元路径, 通过图神经网络融入到知识图谱表示学习中. 再利用注意力网络连接推荐任务和知识图谱表示任务, 其可以学习两个任务之中潜在的特征, 并且能够增强推荐系统中被推荐项和知识图谱中实体的相互作用. 最后在推荐任务中进行用户点击率预测. 模型在公开数据集Book-Crossing和通过DBLP数据集构建的图谱上进行了实验. 最后结果表明, 模型在AUC, 召回率和F1值3个指标上均比其他算法有更好的表现.     

A comprehensive survey on optimizing deep learning models by metaheuristics

Deep neural networks (DNNs), which are extensions of artificial neural networks, can learn higher levels of feature hierarchy established by lower level features by transforming the raw feature space to another complex feature space. Although deep networks are successful in a wide range of problems in different fields, there are some issues affecting their overall performance such as selecting appropriate values for model parameters, deciding the optimal architecture and feature representation and determining optimal weight and bias values. Recently, metaheuristic algorithms have been proposed to automate these tasks. This survey gives brief information about common basic DNN architectures including convolutional neural networks, unsupervised pre-trained models, recurrent neural networks and recursive neural networks. We formulate the optimization problems in DNN design such as architecture optimization, hyper-parameter optimization, training and feature representation level optimization. The encoding schemes used in metaheuristics to represent the network architectures are categorized. The evolutionary and selection operators, and also speed-up methods are summarized, and the main approaches to validate the results of networks designed by metaheuristics are provided. Moreover, we group the studies on the metaheuristics for deep neural networks based on the problem type considered and present the datasets mostly used in the studies for the readers. We discuss about the pros and cons of utilizing metaheuristics in deep learning field and give some future directions for connecting the metaheuristics and deep learning. To the best of our knowledge, this is the most comprehensive survey about metaheuristics used in deep learning field.

     

基于多任务学习的人脸属性识别方法

针对传统深度卷积神经网络模型复杂、识别速度慢的问题,提出一种基于多任务学习的人脸属性识别方法。通过轻量化残差模块构建基础网络,根据属性类之间的关联关系设计共享分支网络,以大幅减少网络参数和计算开销。以多任务学习的方式联合优化各分支网络与基础网络的参数,利用关联属性间的共同特征实现人脸属性识别。采用带权重的交叉熵作为损失函数监督训练网络模型,改善正负样本数不均衡问题。在公开数据集CelebA上的实验结果表明,该方法的识别错误率低至8.45%,空间开销仅2.7 MB,在CPU上每幅图预测时间低至15ms,方便部署在资源有限的移动或便携式设备上,具有实际应用价值。     

基于多视图集成的网络表示学习算法

现有的网络表示学习算法主要为基于浅层神经网络的网络表示学习和基于神经矩阵分解的网络表示学习。基于浅层神经网络的网络表示学习又被证实是分解网络结构的特征矩阵。另外,现有的大多数网络表示学习仅仅从网络的结构学习特征,即单视图的表示学习;然而,网络本身蕴含有多种视图。因此,文中提出了一种基于多视图集成的网络表示学习算法(MVENR)。该算法摈弃了神经网络的训练过程,将矩阵的信息融合和分解思想融入到网络表示学习中。另外,将网络的结构视图、连边权重视图和节点属性视图进行了有效的融合,弥补了现有网络表示学习中忽略了网络连边权重的不足,解决了基于单一视图训练时网络特征稀疏的问题。实验结果表明,所提MVENR算法的性能优于网络表示学习中部分常用的联合学习算法和基于结构的网络表示学习算法,是一种简单且高效的网络表示学习算法。     

半监督属性网络表示学习方法

网络表示学习是一个重要的研究课题,其目的是将高维的属性网络表示为低维稠密的向量,为下一步任务提供有效特征表示。最近提出的属性网络表示学习模型SNE（Social Network Embedding）同时使用网络结构与属性信息学习网络节点表示,但该模型属于无监督模型,不能充分利用一些容易获取的先验信息来提高所学特征表示的质量。基于上述考虑提出了一种半监督属性网络表示学习方法SSNE（Semi-supervised Social Network Embedding）,该方法以属性网络和少量节点先验作为前馈神经网络输入,经过多个隐层非线性变换,在输出层通过保持网络链接结构和少量节点先验,学习最优化的节点表示。在四个真实属性网络和两个人工属性网络上,同现有主流方法进行对比,结果表明本方法学到的表示,在聚类和分类任务上具有较好的性能。     

An incremental neural learning framework and its application to vehicle diagnostics

This paper presents a framework for incremental neural learning (INL) that allows a base neural learning system to incrementally learn new knowledge from only new data without forgetting the existing knowledge. Upon subsequent encounters of new data examples, INL utilizes prior knowledge to direct its incremental learning. A number of critical issues are addressed including when to make the system learn new knowledge, how to learn new knowledge without forgetting existing knowledge, how to perform inference using both the existing and the newly learnt knowledge, and how to detect and deal with aged learnt systems. To validate the proposed INL framework, we use backpropagation (BP) as a base learner and a multi-layer neural network as a base intelligent system. INL has several advantages over existing incremental algorithms: it can be applied to a broad range of neural network systems beyond the BP trained neural networks; it retains the existing neural network structures and weights even during incremental learning; the neural network committees generated by INL do not interact with one another and each sees the same inputs and error signals at the same time; this limited communication makes the INL architecture attractive for parallel implementation. We have applied INL to two vehicle fault diagnostics problems: end-of-line test in auto assembly plants and onboard vehicle misfire detection. These experimental results demonstrate that the INL framework has the capability to successfully perform incremental learning from unbalanced and noisy data. In order to show the general capabilities of INL, we also applied INL to three general machine learning benchmark data sets. The INL systems showed good generalization capabilities in comparison with other well known machine learning algorithms.     

Evolutionary Learning of Modular Neural Networks with Genetic Programming

Evolutionary design of neural networks has shown a great potential as a powerful optimization tool. However, most evolutionary neural networks have not taken advantage of the fact that they can evolve from modules. This paper presents a hybrid method of modular neural networks and genetic programming as a promising model for evolutionary learning. This paper describes the concepts and methodologies for the evolvable model of modular neural networks, which might not only develop new functionality spontaneously, but also grow and evolve its own structure autonomously. We show the potential of the method by applying an evolved modular network to a visual categorization task with handwritten digits. Sophisticated network architectures as well as functional subsystems emerge from an initial set of randomly-connected networks. Moreover, the evolved neural network has reproduced some of the characteristics of natural visual system, such as the organization of coarse and fine processing of stimuli in separate pathways.     

Neural networks that learn from fuzzy if-then rules

An architecture for neural networks that can handle fuzzy input vectors is proposed, and learning algorithms that utilize fuzzy if-then rules as well as numerical data in neural network learning for classification problems and for fuzzy control problems are derived. The learning algorithms can be viewed as an extension of the backpropagation algorithm to the case of fuzzy input vectors and fuzzy target outputs. Using the proposed methods, linguistic knowledge from human experts represented by fuzzy if-then rules and numerical data from measuring instruments can be integrated into a single information processing system (classification system or fuzzy control system). It is shown that the scheme works well for simple examples     

基于精细神经元的类脑感知学习模型

大脑如何实现学习以及感知功能对于人工智能和神经科学领域均是一个重要问题.现有人工神经网络由于结构和计算机制与真实大脑相差较大,无法直接用于理解真实大脑学习以及处理感知任务的机理.树突神经元模型是一种对大脑神经元树突信息处理过程进行建模仿真的计算模型,相比人工神经网络更接近生物真实.使用树突神经网络模型处理学习感知任务对理解真实大脑的学习过程有重要作用.然而,现有基于树突神经元网络的学习模型大都局限于简化树突模型,无法完整建模树突的信号处理过程.针对这一问题,提出一种基于精细中型多棘神经元网络的学习模型,使得精细神经网络可以通过学习完成相应感知任务.实验表明,在经典的图像分类任务上,所提模型可以达到很好的分类性能.此外,精细神经网络对于噪声干扰有很强的鲁棒性.对网络特性进行进一步分析,发现学习后网络中的神经元表现出了刺激选择性这种神经科学中的经典现象,表明所提模型具有一定的生物可解释性,同时也表明刺激选择特性可能是大脑通过学习完成感知任务的一种重要特性.     

图神经网络前沿进展与应用

图结构数据是现实生活中广泛存在的一类数据形式.宏观上的互联网、知识图谱、社交网络数据,微观上的蛋白质、化合物分子等都可以用图结构来建模和表示.由于图结构数据的复杂性和异质性,对图结构数据的分析和处理一直是研究界的难点和重点.图神经网络(Graph Neural Network,GNN)是近年来出现的一种利用深度学习直接对图结构数据进行学习的框架,其优异的性能引起了学者高度的关注和深入的探索.通过在图中的节点和边上制定一定的策略,GNN将图结构数据转化为规范而标准的表示,并输入到多种不同的神经网络中进行训练,在节点分类、边信息传播和图聚类等任务上取得优良的效果.与其他图学习算法相比较,GNN能够学习到图结构数据中的节点以及边的内在规律和更加深层次的语义特征.由于具有对图结构数据强大的非线性拟合能力,因此在不同领域的图相关问题上,GNN都表现出更高的准确率和更好的鲁棒性.本文在现有GNN研究的基础上,首先概述了GNN的出现历程,并介绍了相关概念和定义.之后本文着重讨论和对比了GNN中的各种算法框架,包括核心思想、任务划分、学习方式、优缺点、适用范围、实现成本等.此外,本文对GNN算法在多个不同领域下的应用场景进行了详细的阐述,将GNN与其他图学习算法的优缺点作了联系和比较.针对存在的一些问题和挑战,本文勾画了GNN的未来方向和发展趋势,最后对全文进行了全面而细致的总结.     

Multi-view Contrastive Enhanced Heterogeneous Graph Structure Learning

As a learning method of heterogeneous graph representation, heterogeneous graph neural networks can effectively extract complex structural and semantic information from heterogeneous graphs, and perform excellently in node classification and link prediction tasks to provide strong support for the representation and analysis of knowledge graphs. Due to the existence of some noisy interactions or missing interactions in the heterogeneous graphs, the heterogeneous graph neural network incorporates erroneous neighbor features, thus affecting the overall performance of the model. To solve the above problems, in this paper we proposes a heterogeneous graph structure learning model enhanced by multi-view contrast. Firstly, the semantic information in the heterogeneous graph is maintained by the meta-path, and the similarity graph is generated by calculating the feature similarity among the nodes under each meta-path, which is fused with the meta-path graph to optimize the graph structure. By contrasting the similarity graph and meta-path graph as multiple views, the graph structure is optimized without supervision information, and the dependence on supervision signals is eliminated. Finally, for addressing the problem that the learning ability of the neural network model is insufficient at the initial training stage and there are often erroneous interactions in the generated graph structure, we design a progressive graph structure fusion method. Through incremental weighted addition of meta-path graphs and similarity graphs, the weight of similarity graphs in the fusion is changed. This not only prevents erroneous interactions from being introduced in the initial training stage but also achieves the purpose of employing the interactions in similarity graphs to suppress interference interactions or complete missing interactions, which leads to the optimized heterogeneous structure. Meanwhile, node classification and node clustering are selected as the verification tasks of graph structure learning. The experimental results on four real heterogeneous graph datasets prove that the proposed learning method is feasible and effective. Compared with the optimal comparison model, the performance of this model has been significantly improved under both tasks.     

A hybrid classification algorithm based on coevolutionary EBFNN and domain covering method

A new hybrid scheme of the elliptical basis function neural network (EBFNN) model combined with the cooperative coevolutionary algorithm (Co-CEA) and domain covering method is presented for multiclass classification tasks. This combination of the Co-CEA EBFNN (CC-EBFNN) and the domain covering method is proposed to enhance the predictive capability of the estimated model. The whole training process is divided into two stages: the evolutionary process, and the heuristic structure refining process. First, the initial hidden nodes of the EBFNN model are selected randomly in the training samples, which are further partitioned into modules of hidden nodes with respect to their class labels. Subpopulations are initialized on modules, and the Co-CEA evolves all subpopulations to find the optimal EBFNN structural parameters. Then the heuristic structure refining process is performed on the individual in the elite pool with the special designed constructing and pruning operators. Finally, the CC-EBFNN model is tested on six real-world classification problems from the UCI machine learning repository, and experimental results illustrate that the EBFNN model can be estimated in fewer evolutionary trials, and is able to produce higher prediction accuracies with much simpler network structures when compared with conventional learning algorithms.     

A cooperative constructive method for neural networks for pattern recognition

In this paper, we propose a new constructive method, based on cooperative coevolution, for designing automatically the structure of a neural network for classification. Our approach is based on a modular construction of the neural network by means of a cooperative evolutionary process. This process benefits from the advantages of coevolutionary computation as well as the advantages of constructive methods. The proposed methodology can be easily extended to work with almost any kind of classifier.The evaluation of each module that constitutes the network is made using a multiobjective method. So, each new module can be evaluated in a comprehensive way, considering different aspects, such as performance, complexity, or degree of cooperation with the previous modules of the network. In this way, the method has the advantage of considering not only the performance of the networks, but also other features.The method is tested on 40 classification problems from the UCI machine learning repository with very good performance. The method is thoroughly compared with two other constructive methods, cascade correlation and GMDH networks, and other classification methods, namely, SVM, C4.5, and k nearest-neighbours, and an ensemble of neural networks constructed using four different methods.     

基于矩阵分解的DeepWalk链路预测算法

现有的链路预测方法的数据来源主要是基于邻居、路径和随机游走的方法,使用的是节点相似性假设或者最大似然估计,尚缺少基于神经网络的链路预测研究。基于神经网络的一些研究表明,基于神经网络的DeepWalk网络表示学习算法可以更加有效地挖掘到网络中的结构特征,已有研究证明DeepWalk等同于分解目标矩阵。因此,提出了一种基于矩阵分解的DeepWalk链路预测算法（LPMF）。该算法首先基于矩阵分解的DeepWalk算法分解得到网络的表示向量;然后通过余弦相似度计算每对节点之间的相似度,构建目标网络的相似度矩阵;最后利用相似度矩阵,在三个真实的引文网络中进行链路预测实验。实验结果表明,提出的链路预测算法性能优于现存的20余种链路预测算法。这充分表明了LPMF能够有效地挖掘网络中节点之间的结构关联性,而且在实际网络的链路预测中能够发挥出较为优异的性能。     

多层校正的无监督领域自适应方法

目的 目前深度神经网络已成功应用于众多机器学习任务,并展现出惊人的性能提升效果。然而传统的深度网络和机器学习算法都假定训练数据和测试数据服从的是同一分布,而这种假设在实际应用中往往是不成立的。如果训练数据和测试数据的分布差异很大,那么由传统机器学习算法训练出来的分类器的性能将会大大降低。为了解决此类问题,提出了一种基于多层校正的无监督领域自适应方法。方法 首先利用多层校正来调整现有的深度网络,利用加法叠加来完美对齐源域和目标域的数据表示;然后采用多层权值最大均值差异来适应目标域,增加网络的表示能力;最后提取学习获得的域不变特征来进行分类,得到目标图像的识别效果。结果 本文算法在Office-31图像数据集等4个数字数据集上分别进行了测试实验,以对比不同算法在图像识别和分类方面的性能差异,并进行准确度测量。测试结果显示,与同领域算法相比,本文算法在准确率上至少提高了5%,在应对照明变化、复杂背景和图像质量不佳等干扰情况时,亦能获得较好的分类效果,体现出更强的鲁棒性。结论 在领域自适应相关数据集上的实验结果表明,本文方法具备一定的泛化能力,可以实现较高的分类性能,并且优于其他现有的无监督领域自适应方法。     

Additive Composition of Supervised Self-Organizing Maps

The learning of complex relationships can be decomposed into several neural networks. The modular organization is determined by prior knowledge of the problem that permits to split the processing into tasks of small dimensionality. The sub-tasks can be implemented with neural networks, although the learning examples cannot be used anymore to supervise directly each of the networks. This article addresses the problem of learning in a modular context, developing in particular additive compositions. A simple rule allows defining efficient training, and combining, for example, several Supervised-SOM networks. This technique is important because it introduces interesting generalizations in many modular compositions, permitting data fusion or sequential combinations of neural networks. This revised version was published online in August 2006 with corrections to the Cover Date.     

类脑超大规模深度神经网络系统

大规模神经网络展现出强大的端到端表示能力和非线性函数的无限逼近能力,在多个领域表现出优异的性能,成为一个重要的发展方向.如自然语言处理(NLP)模型GPT,经过几年的发展,目前拥有1750亿网络参数,在多个NLP基准上到达最先进性能.然而,按照现有的神经网络组织方式,目前的大规模神经网络难以到达人脑生物神经网络连接的规...     

A learning automata-based algorithm for determination of the number of hidden units for three-layer neural networks

There is no method to determine the optimal topology for multi-layer neural networks for a given problem. Usually the designer selects a topology for the network and then trains it. Since determination of the optimal topology of neural networks belongs to class of NP-hard problems, most of the existing algorithms for determination of the topology are approximate. These algorithms could be classified into four main groups: pruning algorithms, constructive algorithms, hybrid algorithms and evolutionary algorithms. These algorithms can produce near optimal solutions. Most of these algorithms use hill-climbing method and may be stuck at local minima. In this article, we first introduce a learning automaton and study its behaviour and then present an algorithm based on the proposed learning automaton, called survival algorithm, for determination of the number of hidden units of three layers neural networks. The survival algorithm uses learning automata as a global search method to increase the probability of obtaining the optimal topology. The algorithm considers the problem of optimization of the topology of neural networks as object partitioning rather than searching or parameter optimization as in existing algorithms. In survival algorithm, the training begins with a large network, and then by adding and deleting hidden units, a near optimal topology will be obtained. The algorithm has been tested on a number of problems and shown through simulations that networks generated are near optimal.