Search

Highlights • Development of a biological neural network model that allows long term simulation of brain activity. • Optimization of the model using multiple techniques that led to a speed-up of X200. • Presentation of alternative simulation frameworks for long term simulations., Simulations of neural networks can be used to study the direct effect of internal or external changes on brain dynamics. However, some changes are not immediate but occur on the timescale of weeks, months, or years. Examples include effects of strokes, surgical tissue removal, or traumatic brain injury but also gradual changes during brain development. Simulating network activity over a long time, even for a small number of nodes, is a computational challenge. Here, we model a coupled network of human brain regions with a modified Wilson-Cowan model representing dynamics for each region and with synaptic plasticity adjusting connection weights within and between regions. Using strategies ranging from different models for plasticity, vectorization and a different differential equation solver setup, we achieved one second runtime for one second biological time.

In the process of Human-robot skill transfer, we require the robot to reproduce the trajectory of teacher and expect that the robot can generalize the learned trajectory. For the trajectory after generalization, we expect that the robot arm can accurately track. However, because the model of the robot can not be accurately obtained, some researchers have proposed using a neural network to approximate the unknown term. The parameters of the traditional RBF neural network are usually selected through the empirical and trial-and-error method, which maybe biased and inefficient. In addition, due to the end-effector of the mechanical arm trajectory will be constantly changing according to the needs of the task, when the neural network of compact set cannot contain the whole input vector, the neural network cannot achieve the ideal approximation effect. In this paper, the broad neural network is used to approximate the unknown terms of the robot. This method can reuse the motion controller that has been learned and complete other motions in the robot operating space without relearning its weight parameters. In this paper, the effectiveness of the proposed method is proved by the ultrasound scanning task.

This paper proposes a hybrid architecture based on neural networks, fuzzy systems, and n-uninorms for solving pattern classification problems, termed as ENFS-Uni0 (short for evolving neuro-fuzzy system based on uni-nullneurons). The model can produce knowledge in an on-line (single-pass) and evolving learning context in a particular form of neuro-fuzzy rules representing the dependencies among input features through IF-THEN type relations. The rules antecedents are thereby realized through uni-nullneurons, which are constructed from n-uninorms, leading to the possibility to express both, AND- and OR-connections (and a mixture of these) among the single antecedent parts of a rule (and thus achieving an advanced interpretability aspect of the rules). The neurons’ evolution is done through an extended version of an autonomous data partition method (ADPA). On-line interpretation of the timely evolution of rules is addressed by (i) a concept for tracking the degree of changes of the rules over data stream samples, which may indicate experts/operators how much dynamics is in the process and may be used as a structural active learning component to request operator’s feedback in the case of significant changes and (ii) a concept for updating feature weights incrementally. These weights express the (possibly changing) impact degrees of features on the classification problem: features with low weights can be seen as unimportant and masked out when showing rules to an expert ( → rule length reduction). The rules’ consequents are represented by certainty vectors and are recursively updated by an indicator-based recursive weighted least squares (I-RWLS) approach (one RWLS estimator per class) where the weights are given through the neuron activation levels in order to gain stable local learning. The model proposed in this paper was successfully compared to related hybrid and evolving approaches in the literature for classifying binary and multi-class patterns. The results obtained by the proposed model show an outperformance of the related works in terms of higher accuracy trend lines over time, while offering a high degree of interpretability through coherent neuro-fuzzy rules to solve the classification problems.

With the rapid development of deep learning technology and improvement in computing capability, deep learning has been widely used in the field of hyperspectral image (HSI) classification. In general, deep learning models often contain many trainable parameters and require a massive number of labeled samples to achieve optimal performance. However, in regard to HSI classification, a large number of labeled samples is generally difficult to acquire due to the difficulty and time-consuming nature of manual labeling. Therefore, many research works focus on building a deep learning model for HSI classification with few labeled samples. In this article, we concentrate on this topic and provide a systematic review of the relevant literature. Specifically, the contributions of this paper are twofold. First, the research progress of related methods is categorized according to the learning paradigm, including transfer learning, active learning and few-shot learning. Second, a number of experiments with various state-of-the-art approaches has been carried out, and the results are summarized to reveal the potential research directions. More importantly, it is notable that although there is a vast gap between deep learning models (that usually need sufficient labeled samples) and the HSI scenario with few labeled samples, the issues of small-sample sets can be well characterized by fusion of deep learning methods and related techniques, such as transfer learning and a lightweight model. For reproducibility, the source codes of the methods assessed in the paper can be found at https://github.com/ShuGuoJ/HSI-Classification.git.

Recently, a common starting point for solving complex unsupervised image classification tasks is to use generic features, extracted with deep Convolutional Neural Networks (CNN) pretrained on a large and versatile dataset (ImageNet). However, in most research, the CNN architecture for feature extraction is chosen arbitrarily, without justification. This paper aims at providing insight on the use of pretrained CNN features for image clustering (IC). First, extensive experiments are conducted and show that, for a given dataset, the choice of the CNN architecture for feature extraction has a huge impact on the final clustering. These experiments also demonstrate that proper extractor selection for a given IC task is difficult. To solve this issue, we propose to rephrase the IC problem as a multi-view clustering (MVC) problem that considers features extracted from different architectures as different "views" of the same data. This approach is based on the assumption that information contained in the different CNN may be complementary, even when pretrained on the same data. We then propose a multi-input neural network architecture that is trained end-to-end to solve the MVC problem effectively. This approach is tested on nine natural image datasets, and produces state-of-the-art results for IC., Comment: 21 pages, 16 figures, 10 tables, preprint of our paper published in Neurocomputing

In this paper, we propose a method for obtaining sentence-level embeddings. While the problem of securing word-level embeddings is very well studied, we propose a novel method for obtaining sentence-level embeddings. This is obtained by a simple method in the context of solving the paraphrase generation task. If we use a sequential encoder-decoder model for generating paraphrase, we would like the generated paraphrase to be semantically close to the original sentence. One way to ensure this is by adding constraints for true paraphrase embeddings to be close and unrelated paraphrase candidate sentence embeddings to be far. This is ensured by using a sequential pair-wise discriminator that shares weights with the encoder that is trained with a suitable loss function. Our loss function penalizes paraphrase sentence embedding distances from being too large. This loss is used in combination with a sequential encoder-decoder network. We also validated our method by evaluating the obtained embeddings for a sentiment analysis task. The proposed method results in semantic embeddings and outperforms the state-of-the-art on the paraphrase generation and sentiment analysis task on standard datasets. These results are also shown to be statistically significant., Comment: This work is an extension of our COLING-2018 paper arXiv:1806.00807

Person reidentification (ReID) refers to the task of verifying the identity of a pedestrian observed from nonoverlapping views in a surveillance camera network. It has recently been validated that reranking can achieve remarkable performance improvements in person ReID systems. However, current reranking approaches either require feedback from users or suffer from burdensome computational costs. In this paper, we propose to exploit a density-adaptive smooth kernel technique to achieve efficient and effective reranking. Specifically, we adopt a smooth kernel function to formulate the neighbor relationships among data samples with a density-adaptive parameter. Based on this new formulation, we present two simple yet effective reranking methods, termed \emph{inverse} density-adaptive kernel based reranking (inv-DAKR) and \emph{bidirectional} density-adaptive kernel based reranking (bi-DAKR), in which the local density information in the vicinity of each gallery sample is elegantly exploited. Moreover, we extend the proposed inv-DAKR and bi-DAKR methods to incorporate the available extra probe samples and demonstrate that when and why these extra probe samples are able to improve the local neighborhood and thus further refine the ranking results. Extensive experiments are conducted on six benchmark datasets, including: PRID450s, VIPeR, CUHK03, GRID, Market-1501 and Mars. The experimental results demonstrate that our proposals are effective and efficient., 39 pages, 18 figures and 12 tables. This paper is an extended version of our preliminary work on ICPR 2018

Manual annotation of video segmentation datasets requires an immense amount of human effort, thus, reduction of human annotation costs is an active topic of research. While many papers deal with the propagation of masks through frames of a video, only a few results attempt to optimize annotation task selection. In this paper we present a deep learning based solution to the latter problem and train it using Reinforcement Learning. Our approach utilizes a modified version of the Dueling Deep Q-Network sharing weight parameters across the temporal axis of the video. This technique enables the trained agent to select annotation tasks from the whole video. We evaluate our annotation task selection method by means of a hierarchical supervoxel segmentation based mask propagation algorithm.

It is well known that weighted sum of node states plays an essential role in function implementation of neural networks. Therefore, this paper proposes a new weighted sum synchronization model for memristive neural networks. Unlike the existing synchronization models of memristive neural networks which control each network node to reach synchronization, the proposed model treats the networks as dynamic entireties by weighted sum of node states and makes the entireties instead of each node reach expected synchronization. In this paper, weighted sum complete synchronization and quasi-synchronization are both investigated by designing feedback controller and aperiodically intermittent controller, respectively. Meanwhile, a flexible control scheme is designed for the proposed model by utilizing some switching parameters and can improve anti-interference ability of control system. By applying Lyapunov method and some differential inequalities, some effective criteria are derived to ensure the synchronizations of memristive neural networks. Moreover, the error level of the quasi-synchronization is given. Finally, numerical simulation examples are used to certify the effectiveness of the derived results.

Facial landmarks are highly correlated with each other since a certain landmark can be estimated by its neighboring landmarks. Most of the existing deep learning methods only use one fully-connected layer called shape prediction layer to estimate the locations of facial landmarks. In this paper, we propose a novel deep learning framework named Multi-Center Learning with multiple shape prediction layers for face alignment. In particular, each shape prediction layer emphasizes on the detection of a certain cluster of semantically relevant landmarks respectively. Challenging landmarks are focused firstly, and each cluster of landmarks is further optimized respectively. Moreover, to reduce the model complexity, we propose a model assembling method to integrate multiple shape prediction layers into one shape prediction layer. Extensive experiments demonstrate that our method is effective for handling complex occlusions and appearance variations with real-time performance. The code for our method is available at https://github.com/ZhiwenShao/MCNet-Extension., This paper has been accepted by Neurocomputing

Feature selection plays an important role in reducing irrelevant and redundant features, while retaining the underlying semantics of selected ones. An effective feature selection method is expected to result in a significantly reduced subset of the original features without sacrificing the quality of problem-solving (e.g., classification). In this paper, a non-unique decision measure is proposed that captures the degree of a given feature subset being relevant to different categories. This helps to represent the uncertainty information in the boundary region of a granular model, such as rough sets or fuzzy-rough sets in an efficient manner. Based on this measure, the paper further introduce a differentiation entropy as an evaluator of feature subsets to implement a novel feature selection algorithm. The resulting feature selection method is capable of dealing with either nominal or real-valued data. Experimental results on both benchmark data sets and a real application problem demonstrate that the features selected by the proposed approach outperform those attained by state-of-the-art feature selection techniques, in terms of both the size of feature reduction and the classification accuracy.

Facial expressions play an important role during communications, allowing information regarding the emotional state of an individual to be conveyed and inferred. Research suggests that automatic facial expression recognition is a promising avenue of enquiry in mental healthcare, as facial expressions can also reflect an individual's mental state. In order to develop user-friendly, low-cost and effective facial expression analysis systems for mental health care, this paper presents a novel deep convolution network based emotion analysis framework to support mental state detection and diagnosis. The proposed system is able to process facial images and interpret the temporal evolution of emotions through a new solution in which deep features are extracted from the Fully Connected Layer 6 of the AlexNet, with a standard Linear Discriminant Analysis Classifier exploited to obtain the final classification outcome. It is tested against 5 benchmarking databases, including JAFFE, KDEF,CK+, and databases with the images obtained ‘in the wild’ such as FER2013 and AffectNet. Compared with the other state-of-the-art methods, we observe that our method has overall higher accuracy of facial expression recognition. Additionally, when compared to the state-of-the-art deep learning algorithms such as Vgg16, GoogleNet, ResNet and AlexNet, the proposed method demonstrated better efficiency and has less device requirements. The experiments presented in this paper demonstrate that the proposed method outperforms the other methods in terms of accuracy and efficiency which suggests it could act as a smart, low-cost, user-friendly cognitive aid to detect, monitor, and diagnose the mental health of a patient through automatic facial expression analysis.

Hand gestures are quite suitable for space human-robot interaction (SHRI) because of their natural and convenient features. While the detection and localization of hands are the premise and foundation for SHRI based on hand gestures. But hand gestures are very complicated and hand sizes are very small in some images. These problems make the robust real-time hand detection and local- ization very difficult. In this paper, a feature-map-fused single shot multibox detector (FF-SSD) which is a deep learning network is designed to deal with the problems of hand detection and localization in SHRI. First, the background of the method is introduced in this paper, including an astronaut assistant robot platform, the difficulties of hand detection and localization, and introduction of the state-of-the-art deep learning networks for object detection and localiza- tion. Then, the FF-SSD is proposed for detecting and localizing hands especially pony-size hands. This network magentatakes into consideration both accuracy and speed with balanced performance. And in the experiment part, the FF-SSD is trained and tested on hand databases which include a homemade database and two public databases. At last, the superiority of the proposed method is demonstrated compared with the state-of-the-art methods.

The Denoising Autoencoder (DAE) enhances the flexibility of the data stream method in exploiting unlabeled samples. Nonetheless, the feasibility of DAE for data stream analytic deserves an in-depth study because it characterizes a fixed network capacity that cannot adapt to rapidly changing environments. Deep evolving denoising autoencoder (DEVDAN), is proposed in this paper. It features an open structure in the generative phase and the discriminative phase where the hidden units can be automatically added and discarded on the fly. The generative phase refines the predictive performance of the discriminative model exploiting unlabeled data. Furthermore, DEVDAN is free of the problem-specific threshold and works fully in the single-pass learning fashion. We show that DEVDAN can find competitive network architecture compared with state-of-the-art methods on the classification task using ten prominent datasets simulated under the prequential test-then-train protocol., This paper has been accepted for publication in Neurocomputing 2019. arXiv admin note: substantial text overlap with arXiv:1809.09081

In this paper, we address a challenging task in real-word applications, i.e., automatic reading recognition for pointer meters, called PRM. 1 This application is valuable in the fields of military, industry, and aerospace. However, the accuracy of recognizing the readings of pointer meters by machine vision is, oftentimes, affected by several factors, such as uneven illumination in each image, large range variation of illumination in different images, complex backgrounds, tilting of pointer meters, image blur, and scale change, resulting in the recognized readings with unacceptable accuracy. In this paper, a new robust approach to reading recognition of pointer meters is proposed. The proposed method consists of three main contributions: (1) constructing a novel deep learning algorithm in which the PrRoIPooling is used in lieu of the RoiAlign in the existing Mask-RCNN, (2) classifying the type of pointer meters while fitting the pointer binary mask, and (3) calculating the readings of pointer meters by the proposed angle method. In addition, we also report and release a new dataset for the community. Experiments show that the new algorithm can significantly improve the accuracy of the recognized readings of pointer meters, meanwhile, the proposed approach is also robust to the natural environments and computationally efficient.

Although supervised deep representation learning has attracted enormous attentions across areas of pattern recognition and computer vision, little progress has been made towards unsupervised deep representation learning for image clustering. In this paper, we propose a deep spectral analysis network for unsupervised representation learning and image clustering. While spectral analysis is established with solid theoretical foundations and has been widely applied to unsupervised data mining, its essential weakness lies in the fact that it is difficult to construct a proper affinity matrix and determine the involving Laplacian matrix for a given dataset. In this paper, we propose a SA-Net to overcome these weaknesses and achieve improved image clustering by extending the spectral analysis procedure into a deep learning framework with multiple layers. The SA-Net has the capability to learn deep representations and reveal deep correlations among data samples. Compared with the existing spectral analysis, the SA-Net achieves two advantages: (i) Given the fact that one spectral analysis procedure can only deal with one subset of the given dataset, our proposed SA-Net elegantly integrates multiple parallel and consecutive spectral analysis procedures together to enable interactive learning across different units towards a coordinated clustering model; (ii) Our SA-Net can identify the local similarities among different images at patch level and hence achieves a higher level of robustness against occlusions. Extensive experiments on a number of popular datasets support that our proposed SA-Net outperforms 11 benchmarks across a number of image clustering applications., Comment: arXiv admin note: text overlap with arXiv:2009.05235

The Convolutional Neural Networks (CNNs), in domains like computer vision, mostly reduced the need for handcrafted features due to its ability to learn the problem-specific features from the raw input data. However, the selection of dataset-specific CNN architecture, which mostly performed by either experience or expertise is a time-consuming and error-prone process. To automate the process of learning a CNN architecture, this paper attempts at finding the relationship between Fully Connected (FC) layers with some of the characteristics of the datasets. The CNN architectures, and recently datasets also, are categorized as deep, shallow, wide, etc. This paper tries to formalize these terms along with answering the following questions. (i) What is the impact of deeper/shallow architectures on the performance of the CNN w.r.t. FC layers?, (ii) How the deeper/wider datasets influence the performance of CNN w.r.t. FC layers?, and (iii) Which kind of architecture (deeper/ shallower) is better suitable for which kind of (deeper/ wider) datasets. To address these findings, we have performed experiments with three CNN architectures having different depths. The experiments are conducted by varying the number of FC layers. We used four widely used datasets including CIFAR-10, CIFAR-100, Tiny ImageNet, and CRCHistoPhenotypes to justify our findings in the context of the image classification problem. The source code of this research is available at https://github.com/shabbeersh/Impact-of-FC-layers., This paper is accepted for publication in Neurocomputing Journal

In this paper, we focus on synchronization issue of coupled multistable memristive neural networks (CMMNNs) with time delay under multiple stable equilibrium states. First, we build delayed CMMNNs consisting of one master subnetwork without controller and N−1 identical slave subnetworks with controllers, and every subnetwork has n nodes. Moreover, this paper investigates multistability of delayed CMMNNs with continuous nonmonotonic piecewise linear activation function (PLAF) owning 2 r + 2 corner points. By using the theorems of differential inclusion and fixed point, sufficient conditions are derived such that master subnetwork of CMMNNs can acquire ( r + 2 ) n exponentially stable equilibrium points, stable periodic orbits or hybrid stable equilibrium states. Then, this paper proposes hybrid multisynchronization of delayed CMMNNs related with various external inputs under multiple stable equilibrium states for the first time. There exist ( r + 2 ) n hybrid multisynchronization manifolds in CMMNNs with different initial conditions and external inputs. Finally, two numerical simulations are given to illustrate the effectiveness of the obtained results.

This paper addresses two issues toward practical use of wearable electroencephalogram (EEG) measurement devices. Ocular (eye movement and blink) artifacts often contaminate EEGs and deteriorate the performance of EEG-based brain–computer interfaces (BCIs). Although wearable consumer-grade EEG devices with single electrode allow users to operate BCIs conveniently in daily lives, it remains a challenging issue to attenuate ocular artifacts from single-channel measurements without spatial information. Existing ocular artifact reduction methods are, however, not simple enough for single-channel EEG data in the sense that they require an additional reference channel and/or pre-processing for artifact segment detection. Another issue is how to assess the performance of artifact reduction; the existing studies have used their own datasets that are not accessible from other researchers. Then, this paper makes two major contributions. (1) This paper proposes a novel ocular artifact reduction method, multi-scale dictionary learning (MSDL), which operates under single-channel measurements and without artifact segment detection. (2) We also develop a semi-simulation setting for quantitative evaluation with a publicly available EEG dataset. In particular, we employed BCI Competition IV Dataset 2a, on which the proposed method was compared with state-of-art methods. The proposed technique showed the best performance for recovering artifact-reduced waveforms from single-channel data compared to the other artifact reduction methods. The Matlab scripts for semi-simulation data generation and single-channel artifact reduction are available on GitHub .

In online-learning, the data is incrementally received and the distributions from which it is drawn may keep changing over time. This phenomenon is widely known as concept drift. Such changes may affect the generalization of a learned model to future data. This problem may be exacerbated by the form of the drift itself changing over time. Quantitative measures to describe and analyze the concept drift have been proposed in previous work. A description composed from these measures is called a concept drift map. We believe that these maps could be useful for guiding how much knowledge in the old model should be forgotten. Therefore, this paper presents an adaptive online learning model that uses a concept drift map to regulate the forgetting factor of an extreme learning machine. Specifically, when a batch of new instances are labeled, the distribution of each class on each attribute is firstly estimated, and then it is compared with the distribution estimated in the previous batch to calculate the magnitude of concept drift, which is further used to regulate the forgetting factor and to update the learning model. Therefore, the novelty of this paper lies in that a quantitative distance metric between two distributions constructed on continuous attribute space is presented to construct concept drift map which can be further associated with the forgetting factor to make the learning model adapt the concept drift. Experimental results on several benchmark stream data sets show the proposed model is generally superior to several previous algorithms when classifying a variety of data streams subject to drift, indicating its effectiveness and feasibility.

Currently, knowledge discovery in databases is an essential first step when identifying valid, novel and useful patterns for decision making. There are many real-world scenarios, such as bankruptcy prediction, option pricing or medical diagnosis, where the classification models to be learned need to fulfill restrictions of monotonicity (i.e. the target class label should not decrease when input attributes values increase). For instance, it is rational to assume that a higher debt ratio of a company should never result in a lower level of bankruptcy risk. Consequently, there is a growing interest from the data mining research community concerning monotonic predictive models. This paper aims to present an overview of the literature in the field, analyzing existing techniques and proposing a taxonomy of the algorithms based on the type of model generated. For each method, we review the quality metrics considered in the evaluation and the different data sets and monotonic problems used in the analysis. In this way, this paper serves as an overview of monotonic classification research in specialized literature and can be used as a functional guide for the field.

Twin Support Vector Machines (TWSVMs) have emerged as an efficient alternative to Support Vector Machines (SVM) for learning from imbalanced datasets. The TWSVM learns two non-parallel classifying hyperplanes by solving a couple of smaller sized problems. However, it is unsuitable for large datasets, as it involves matrix operations. In this paper, we discuss a Twin Neural Network (Twin NN) architecture for learning from large unbalanced datasets. The objective functions of the networks in the Twin NN are designed to realize the idea of the Twin SVM with non-parallel decision boudaries for the respective classes, while also being able to reduce model complexity. The Twin NN optimizes the feature map, allowing for better discrimination between classes. The paper also discusses an extension of the Twin NN for multiclass datasets. This architecture trains as many neural networks as the number of classes, and has the additional advantage that it does not have any hyper-parameter which requires tuning. Results presented in the paper demonstrate that the Twin NN generalizes well and scales well on large unbalanced datasets.

Feature selection is an important step in building classifiers for high-dimensional data problems, such as EEG classification for BCI applications. This paper proposes a new wrapper method for feature selection, based on a multi-objective evolutionary algorithm, where the representation of the individuals or potential solutions, along with the breeding operators and objective functions, have been carefully designed to select a small subset of features that has good generalization capability, trying to avoid the over-fitting problems that wrapper methods usually suffer. A novel feature ranking procedure is also proposed in order to analyze the stability of the proposed wrapper method. Four different classification schemes have been applied within the proposed wrapper method in order to evaluate its accuracy and stability for feature selection on a real motor imagery dataset. Experimental results show that the wrapper method presented in this paper is able to obtain very small subsets of features, which are quite stable and also achieve high classification accuracy, regardless of the classifiers used., Project TIN2015-67020-P (Spanish “Ministerio de Economía y Competitividad”), European Regional Development Funds (ERDF)

In this paper, we study the problem of matrix recovery, which aims to restore a target matrix of authentic samples from grossly corrupted observations. Most of the existing methods, such as the well-known Robust Principal Component Analysis (RPCA), assume that the target matrix we wish to recover is low-rank. However, the underlying data structure is often non-linear in practice, therefore the low-rankness assumption could be violated. To tackle this issue, we propose a novel method for matrix recovery in this paper, which could well handle the case where the target matrix is low-rank in an implicit feature space but high-rank or even full-rank in its original form. Namely, our method pursues the low-rank structure of the target matrix in an implicit feature space. By making use of the specifics of an accelerated proximal gradient based optimization algorithm, the proposed method could recover the target matrix with non-linear structures from its corrupted version. Comprehensive experiments on both synthetic and real datasets demonstrate the superiority of our method.

It is a big challenge of computer vision to make machine automatically describe the content of an image with a natural language sentence. Previous works have made great progress on this task, but they only use the global or local image feature, which may lose some important subtle or global information of an image. In this paper, we propose a model with 3-gated model which fuses the global and local image features together for the task of image caption generation. The model mainly has three gated structures. 1) Gate for the global image feature, which can adaptively decide when and how much the global image feature should be imported into the sentence generator. 2) The gated recurrent neural network (RNN) is used as the sentence generator. 3) The gated feedback method for stacking RNN is employed to increase the capability of nonlinearity fitting. More specially, the global and local image features are combined together in this paper, which makes full use of the image information. The global image feature is controlled by the first gate and the local image feature is selected by the attention mechanism. With the latter two gates, the relationship between image and text can be well explored, which improves the performance of the language part as well as the multi-modal embedding part. Experimental results show that our proposed method outperforms the state-of-the-art for image caption generation., Comment: 35 pages, 7 figures, magazine

In recent years, convolutional neural networks (CNNs) have shown great performance in various fields such as image classification, pattern recognition, and multi-media compression. Two of the feature properties, local connectivity and weight sharing, can reduce the number of parameters and increase processing speed during training and inference. However, as the dimension of data becomes higher and the CNN architecture becomes more complicated, the end-to-end approach or the combined manner of CNN is computationally intensive, which becomes limitation to CNN’s further implementation. Therefore, it is necessary and urgent to implement CNN in a faster way. In this paper, we first summarize the acceleration methods that contribute to but not limited to CNN by reviewing a broad variety of research papers. We propose a taxonomy in terms of three levels, i.e. structure level, algorithm level, and implementation level, for acceleration methods. We also analyze the acceleration methods in terms of CNN architecture compression, algorithm optimization, and hardware-based improvement. At last, we give a discussion on different perspectives of these acceleration and optimization methods within each level. The discussion shows that the methods in each level still have large exploration space. By incorporating such a wide range of disciplines, we expect to provide a comprehensive reference for researchers who are interested in CNN acceleration.

This paper studies the consensus of multi-agent systems with faults and mismatches under switched topologies using a delta operator method. Since faults and mismatches can result in failure of the consensus even for a fixed topology with a spanning tree, how to reach a consensus is a complicated and challenging problem under such circumstances especially when part topologies have no spanning tree. Although some works studied the influence of faults and mismatches on the consensus, there is little work on reaching a consensus for the multi-agent systems with faults and mismatches. In this paper, we introduce the delta operator to unify the consensus analysis for continuous, discrete, or sampled systems under one framework. We develop the theories on the delta operator systems first and then apply theories of the delta operator systems to the consensus problems. By converting the consensus problems into stability problems, we investigate and prove consensus and the associated conditions for systems 1) without any fault, 2) with a known fault, and 3) with unknown faults, under switching topologies with matching or mismatching coefficients. Numerical examples are provided and validate the effectiveness of the theoretical results.

This work addresses the coordination problem of multiple robots with the goal of finding specific hazardous targets in an unknown area and dealing with them cooperatively. The desired behaviour for the robotic system entails multiple requirements, which may also be conflicting. The paper presents the problem as a constrained bi-objective optimization problem in which mobile robots must perform two specific tasks of exploration and at same time cooperation and coordination for disarming the hazardous targets. These objectives are opposed goals, in which one may be favored, but only at the expense of the other. Therefore, a good trade-off must be found. For this purpose, a nature-inspired approach and an analytical mathematical model to solve this problem considering a single equivalent weighted objective function are presented. The results of proposed coordination model, simulated in a two dimensional terrain, are showed in order to assess the behaviour of the proposed solution to tackle this problem. We have analyzed the performance of the approach and the influence of the weights of the objective function under different conditions: static and dynamic. In this latter situation, the robots may fail under the stringent limited budget of energy or for hazardous events. The paper concludes with a critical discussion of the experimental results., Comment: 40 pages

Activation function is crucial to the recent successes of deep neural networks. In this paper, we first propose a new activation function, Multiple Parametric Exponential Linear Units (MPELU), aiming to generalize and unify the rectified and exponential linear units. As the generalized form, MPELU shares the advantages of Parametric Rectified Linear Unit (PReLU) and Exponential Linear Unit (ELU), leading to better classification performance and convergence property. In addition, weight initialization is very important to train very deep networks. The existing methods laid a solid foundation for networks using rectified linear units but not for exponential linear units. This paper complements the current theory and extends it to the wider range. Specifically, we put forward a way of initialization, enabling training of very deep networks using exponential linear units. Experiments demonstrate that the proposed initialization not only helps the training process but leads to better generalization performance. Finally, utilizing the proposed activation function and initialization, we present a deep MPELU residual architecture that achieves state-of-the-art performance on the CIFAR-10/100 datasets. The code is available at https://github.com/Coldmooon/Code-for-MPELU .

Pedestrian detection and tracking have become an important field in the computer vision research area. This growing interest, started in the last decades, might be explained by the multitude of potential applications that could use the results of this research field, e.g. robotics, entertainment, surveillance, care for the elderly and disabled, and content-based indexing. In this survey paper, vision-based pedestrian detection systems are analysed based on their field of application, acquisition technology, computer vision techniques and classification strategies. Three main application fields have been individuated and discussed: video surveillance, human-machine interaction and analysis. Due to the large variety of acquisition technologies, this paper discusses both the differences between 2D and 3D vision systems, and indoor and outdoor systems. The authors reserved a dedicated section for the analysis of the Deep Learning methodologies, including the Convolutional Neural Networks in pedestrian detection and tracking, considering their recent exploding adoption for such a kind systems. Finally, focusing on the classification point of view, different Machine Learning techniques have been analysed, basing the discussion on the classification performances on different benchmark datasets. The reported results highlight the importance of testing pedestrian detection systems on different datasets to evaluate the robustness of the computed groups of features used as input to classifiers.

Many real-life problems involve dealing with multiple objectives. For example, in network routing the criteria may consist of energy consumption, latency, and channel capacity, which are in essence conflicting objectives. As in many problems there may be multiple (conflicting) objectives, there usually does not exist a single optimal solution. In those cases, it is desirable to obtain a set of trade-off solutions between the objectives. This problem has in the last decade also gained the attention of many researchers in the field of reinforcement learning (RL). RL addresses sequential decision problems in initially (possibly) unknown stochastic environments. The goal is the maximization of the agent's reward in an environment that is not always completely observable. The purpose of this special issue is to obtain a broader picture on the algorithmic techniques at the confluence between multi-objective optimization and reinforcement learning. The growing interest in multi-objective reinforcement learning (MORL) was reflected in the quantity and quality of submissions received for this special issue. After a rigorous review process, seven papers were accepted for publication, and they reflect the diversity of research being carried out within this emerging field of research. The accepted papers consider many different aspects of algorithmic design and the evaluation and this editorial puts them in a unified framework.

The outbreak and rapid spread of coronavirus disease 2019 (COVID-19) has had a huge impact on the lives and safety of people around the world. Chest CT is considered an effective tool for the diagnosis and follow-up of COVID-19. For faster examination, automatic COVID-19 diagnostic techniques using deep learning on CT images have received increasing attention. However, the number and category of existing datasets for COVID-19 diagnosis that can be used for training are limited, and the number of initial COVID-19 samples is much smaller than the normal’s, which leads to the problem of class imbalance. It makes the classification algorithms difficult to learn the discriminative boundaries since the data of some classes are rich while others are scarce. Therefore, training robust deep neural networks with imbalanced data is a fundamental challenging but important task in the diagnosis of COVID-19. In this paper, we create a challenging clinical dataset (named COVID19-Diag) with category diversity and propose a novel imbalanced data classification method using deep supervised learning with a self-adaptive auxiliary loss (DSN-SAAL) for COVID-19 diagnosis. The loss function considers both the effects of data overlap between CT slices and possible noisy labels in clinical datasets on a multi-scale, deep supervised network framework by integrating the effective number of samples and a weighting regularization item. The learning process jointly and automatically optimizes all parameters over the deep supervised network, making our model generally applicable to a wide range of datasets. Extensive experiments are conducted on COVID19-Diag and three public COVID-19 diagnosis datasets. The results show that our DSN-SAAL outperforms the state-of-the-art methods and is effective for the diagnosis of COVID-19 in varying degrees of data imbalance.

K-nearest neighbor (KNN) based weighted multi-class twin support vector machines (KWMTSVM) is a novel multi-class classification method. In this paper, we propose a novel least squares version of KWMTSVM called LS-KWMTSVM by replacing the inequality constraints with equality constraints and minimized the slack variables using squares of 2-norm instead of conventional 1-norm. This simple modification leads to a very fast algorithm with much better results. The modified primal problems in the proposed LS-KWMTSVM solves only two systems of linear equations whereas two quadratic programming problems (QPPs) need to solve in KWMTSVM. The proposed LS-KWMTSVM, same as KWMTSVM, employed the weight matrix in the objective function to exploit the local information of the training samples. To exploit the inter class information, we use weight vectors in the constraints of the proposed LS-KWMTSVM. If any component of vectors is zero then the corresponding constraint is redundant and thus we can avoid it. Elimination of redundant constraints and solving a system of linear equations instead of QPPs makes the proposed LS-KWMTSVM more robust and faster than KWMTSVM. The proposed LS-KWMTSVM, commensurate as the KWMTSVM, all the training data points into a “1-versus-1-versus-rest” structure, and thus our LS-KWMTSVM generate ternary output {-1,0,+1} which helps to deal with imbalance datasets. Numerical experiments on several UCI and KEEL imbalance datasets(with high imbalance ratio) clearly indicate that the proposed LS-KWMTSVM has better classification accuracy compared with other baseline methods but with remarkably less computational time. This work was supported by Science and Engineering Research Board (SERB) under Early Career Research Award Grant No. ECR/2017/000053 and Ramanujan fellowship Grant No. SB/S2/ RJN-001/2016. This work was also supported by Council of Scientific & Industrial Research (CSIR), New Delhi, INDIA under Extra Mural Research (EMR) Scheme Grant No. 22(0751)/17/ EMR-II.

The use of hierarchical networks in motor pattern generation in hexapod robots and in insect movement provides ways to understand and control both systems. In this paper we consider the Central Pattern Generator (CPG) of insect movement consisting of six coupled neurons developed by Ghigliazza and Holmes (2004) that produces the global leg coordination pattern. We provide a detailed study of the possible gaits generated by the CPG through different numerical techniques recently developed for the study of small networks, which allow us to consider the complete model without any simplification. We combine the analysis of isolated neurons (using several three dimensional parameter spaces) to give a roadmap of the dynamics of the neurons involved, lateral phase lag plots to show the convergence and transitions towards particular patterns and a quasi-Monte Carlo sweeping method to describe different pattern routes in the parametric phase space. In all of our studies we have found the same final result: most of the observed patterns follow routes that lead to the stable tripod gait. We obtain several routes made of symmetrical patterns, but despite considering mainly a symmetric leg CPG we detect non-symmetrical patterns that provide (minor) routes present in the model. The bifurcations of the main pattern routes detected in the parameter space are studied in detail using continuation techniques. This study reveals the bifurcations that create and destroy the different routes and how for large values of some parameters only the tripod gait is present. Based on the symmetric case, a preliminary study of asymmetric configurations is done revealing the robustness of the already located patterns and routes. Due to the relevance of the tripod gait, and since more parameters are included in this study, we introduce an algorithm to locate the limit of the tripod gait in the parameter space, which shows that there is a large three parametric region where the tripod pattern is ubiquitous and highly dominant in rapidly moving insect regimes.

SCARA robot is one of the most popularly used robots in industry. The obstacle avoidance feature of multiple SCARA robot collaboration is essential and prominent, which can be used to support multiple robots to accomplish not only more sophisticated tasks but also more efficient than individual robot. This paper mainly focuses on studying the problem of simultaneous multi-robot coordination and obstacle avoidance. A cooperative kinematic control problem of multiple robot manipulators, collision avoidance is taken into account to be the primary task as an inequality constraint and trajectory planning task is considered to be the secondary objective as to ensure the priority of safety, is described as a quadratic programming (QP) problem. Then, a recurrent neural network (RNN) based dynamic controller is designed to solve the formulated QP problem recursively. The convergence of the designed neural network is proved through Lyapunov analysis. With three SCARA planar robots, the effectiveness of the proposed controller is validated through numerical simulations. As observed in the results, when the minimal distance between robots is less than the setting safety distance, the collision avoidance strategy reacts to impel robots to avoid collision, which achieves the primary objective for obstacle avoidance; otherwise, the robot performs the desired trajectory tracking task.

Computational dentistry uses computerized methods and mathematical models for dental image analysis. One of the fundamental problems in computational dentistry is accurate tooth instance segmentation in high-resolution mesh data of intra-oral scans (IOS). This paper presents a new computational model based on deep neural networks, called Mask-MCNet, for end-to-end learning of tooth instance segmentation in 3D point cloud data of IOS. The proposed Mask-MCNet localizes each tooth instance by predicting its 3D bounding box and simultaneously segments the points that belong to each individual tooth instance. The proposed model processes the input raw 3D point cloud in its original spatial resolution without employing a voxelization or down-sampling technique. Such a characteristic preserves the finely detailed context in data like fine curvatures in the border between adjacent teeth and leads to a highly accurate segmentation as required for clinical practice (e.g. orthodontic planning). The experiments show that the Mask-MCNet outperforms state-of-the-art models by achieving 98% Intersection over Union (IoU) score on tooth instance segmentation which is very close to human expert performance.

[EN] The electric vehicle (EV) is gradually being introduced in cities. The impact of this introduction is less due, among other reasons, to the lack of charging infrastructure necessary to satisfy the demand. In today¿s cities there is no adequate infrastructure and it is necessary to have action plans that allow an easy deployment of a network of EV charging points in current cities. These action plans should try to place the EV charging stations in the most appropriate places for optimizing their use. According to this, this paper presents an agent-oriented approach that analyses the different configurations of possible locations of charging stations for the electric vehicles in a specific city. The proposed multi-agent system takes into account data from a variety of sources such as social networks activity and mobility information in order to estimate the best configurations. The proposed approach employs a genetic algorithm (GA) that tries to optimize the possible configurations of the charging infrastructure. Additionally, a new crossover method for the GA is proposed considering this context., This work was partially supported by MINECO/FEDER RTI2018-095390-B-C31 and MODINVECI project of the Spanish government. Vicent Botti and Jaume Jordan are funded by UPV PAID-06-18 project. Jaume Jordan is funded by grant APOSTD/2018/010 of GVA-FSE

Scene text detection plays a vital role for scene text understanding, but arbitrary-shaped text detection remains a significant challenge. To extract discriminative features, most recent state-of-the-art methods adopt heavy networks, resulting in parameter redundancy and inference inefficiency. For accurate and efficient scene text detection, in this paper we propose a novel recurrent feature refinement network (RFRN). RFRN, as a recurrent segmentation framework, contains a recurrent path augmentation that refines the previous feature maps as inner states, which not only helps improve the segmentation quality, but also fully facilitates the reuse of parameters and low computational cost. During testing, RFRN discards redundant prediction procedures for efficient inference, and achieves a good balance between speed and accuracy of inference. We conduct experiments on four challenging scene text benchmarks, CTW1500, Total-Text, ICDAR2015 and ICDAR2017-MLT, which include curved texts and multi-oriented texts with complex background. The results show that the proposed RFRN achieves competitive performance on detection accuracy while maintaining computational efficiency.

Knowledge distillation (KD) is widely applied in the training of efficient neural network. A compact model, which is trained to mimic the representation of a cumbersome model for the same task, generally obtains a better performance compared with being trained with the ground truth label. Previous KD-based works mainly focus on two aspects: (1) designing various feature representation for knowledge transfer; (2) introducing different training mechanism such as progressive learning or adversarial learning. In this paper, we revisit the standard KD and observe that training with teacher’s logits might suffer from incorrect and uncertain supervision. To tackle these problems, we propose two novel approaches to deal with incorrect logits and uncertain logits respectively, which are called Logits Adjustment (LA) and Dynamic Temperature Distillation (DTD). To be specific, LA rectifies the incorrect logits according to ground truth label and certain rules. While DTD treats the temperature of KD as a dynamic sample wise parameter rather than a static and global hyper-parameter, which actually notes the uncertainty for each sample’s logits. With iteratively updating the sample wise temperature, the student model could pay more attention on the samples that confuse the teacher model. Experiments on CIFAR-10/100, CINIC-10 and Tiny ImageNet verify that the proposed methods yield encouraging improvement compared with the standard KD. Furthermore, considering the simple implementations, LA and DTD can be easily attached to many KD-based frameworks and bring improvements without extra cost of training time and computing resources.

Detail-and-geometry richness is essential to bas-relief modelling. However, existing image-based and model-based bas-relief modelling techniques commonly suffer from detail monotony or geometry loss. In this paper, we introduce a new bas-relief modelling framework for detail abundance with visual attention based mask generation and geometry preservation, which benefits from our two key contributions. For detail richness, we propose a novel semantic neural network of normal transfer to enrich the texture styles on bas-reliefs. For geometry preservation, we introduce a normal decomposition scheme based on Domain Transfer Recursive Filter (DTRF). Experimental results demonstrate that our approach is advantageous on producing bas-relief modellings with both fine details and geometry preservation.

Hindsight Experience Replay (HER) is a multi-goal reinforcement learning algorithm for sparse reward functions. The algorithm treats every failure as a success for an alternative (virtual) goal that has been achieved in the episode. Virtual goals are randomly selected, irrespective of which are most instructive for the agent. In this paper, we present two improvements over the existing HER algorithm. First, we prioritize virtual goals from which the agent will learn more valuable information. We call this property the instructiveness of the virtual goal and define it by a heuristic measure, which expresses how well the agent will be able to generalize from that virtual goal to actual goals. Secondly, we reduce existing bias in HER by the removal of misleading samples. To test our algorithms, we built three challenging environments with sparse reward functions. Our empirical results in both environments show vast improvement in the final success rate and sample efficiency when compared to the original HER algorithm. A video showing experimental results is available at https://youtu.be/xjAiwJiSeLc.

Quantized neural networks (NN) with reduced bit precision are practical solutions to minimize computational and memory resource requirements and play a vital role in machine learning. However, it is still challenging to avoid significant accuracy degradation due to numerical approximation and lower redundancy. In this paper, a novel robustness-aware 2-bit quantization scheme (RAQ) is proposed for NN, based on binary NN and generative adversarial networks (GAN), which improve performance by enriching binary NN information, extracting the structural information and considering the robustness of the quantized NN. Specifically, using a shift-add operation to replace the multiply-accumulate in the quantization process can speed the NN. A structural loss is proposed to represent the difference between the original NN and quantized NN, such that the structural information of data is preserved after quantization. The structural information learned from NN plays an important role in improving the performance and allows for further fine-tuning of the quantized NN by applying the Lipschitz constraint to the structural loss. For the first time, we consider the robustness of the quantized NN and propose a non-sensitive perturbation loss function by introducing an extraneous term of the spectral norm. The experiments were conducted on CIFAR-10, SVHN and ImageNet datasets with popular NN (such as MobileNetV2, ResNet20, etc.). Extensive experiments show that our new 2-bit quantization scheme is more efficient than the state-of-the-art quantization methods. Our scheme effectively reduced the latency by 2 × and the accuracy decline by 1–4%. Meanwhile, the experimental results also demonstrate that the RAQ is robust with adversarial attacks, we not only eliminate the robustness gap between full-precision and quantized models, but also improve the robustness over full-precision ones by 10%.

Dimensionality reduction (DR) is a popular approach to data exploration in which instances in a given dataset are mapped to a lower-dimensional representation or “embedding.” For nonlinear dimensionality reduction (NLDR), the dimensions of the embedding may be difficult to understand. In such cases, it may be useful to learn how the different dimensions relate to a set of external features (i.e., relevant features that were not used for the DR). A variety of methods (e.g., PROFIT and BIR) use external features to explain embeddings generated by NLDR methods with rotation-invariant objective functions, such as multidimensional scaling (MDS). However, these methods are restricted to two-dimensional embeddings. In this paper, we propose BIOT, which makes it possible to explain an MDS embedding with any number of dimensions without requiring visualization.

This paper proposed a novel approach to generate face caricatures automatically from a single portrait image. We decompose the process of 3D face caricatures generation into two independent subtasks: appearance transfer of texture and the geometry transfer of mesh. For the appearance transfer, we design a GAN-based network named CariFaceGAN to learn the style mapping from portrait to caricature, in which facial features are leveraged to preserve identity consistency. For geometry transfer, we first learn the transformation of the landmarks between portraits and caricatures in an embedded space obtained with Locally Linear Embedding method, and then Kriging interpolation is used to manipulate the portrait mesh constructed from a single image. The experimental results show that our proposed CariFaceGAN outperforms the state-of-the-art methods in terms of maintaining identity consistency and providing satisfactory visual effects.

At present, multi-oriented text detection methods based on deep neural network have achieved promising performances on various benchmarks. Nevertheless, there are still some difficulties for arbitrary shape text detection, especially for a simple and proper representation of arbitrary shape text instances. In this paper, a pixel-based text detector is proposed to facilitate the representation and prediction of text instances with arbitrary shapes in a simple manner. Firstly, to alleviate the influence of the target vertex sorting and achieve the direct regression of arbitrary shape text instances, the starting-point-independent coordinates regression loss is proposed. Furthermore, to predict more accurate text instances, the text instance accuracy loss is proposed as an assistant task to refine the predicted coordinates under the guidance of IoU. To evaluate the effectiveness of our detector, extensive experiments have been carried on public benchmarks which contain arbitrary shape text instances and multi-oriented text instances. We obtain 84.8% of F-measure on Total-Text benchmark. The results show that our method can reach state-of-the-art performance.

Background Channel selection is a key topic in brain-computer interface (BCI). Task-irrelevant and redundant channels used in BCI may lead to low classification accuracy, high computational complexity, and inconvenience for application. By selecting optimal channels, the performance of BCI could enhance significantly. Method In this paper, a new binary harmony search (BHS) is proposed to select the optimal channel sets and optimize the system accuracy. The BHS is implemented on the training data sets to select the optimal channels and the test data sets are used to evaluate the classification performance on the selected channels. The sparse representation-based classification, linear discriminant analysis, and support vector machine are performed on the common spatial pattern (CSP) features for motor imagery (MI) classification. Results Two public EEG datasets are employed to validate the proposed BHS method. The paired t-test is conducted on the test classification performance between the BHS and traditional CSP with all channels. The results reveal that the proposed BHS method significantly improved classification accuracy as compared to the conventional CSP method (p Conclusion This study proposed the BHS method to select the optimal channels in MI -based BCI. On the one hand, the results confirm the validity of the BHS algorithm as a channel selection method for motor imagery data. On the other hand, the BHS method with costing shorter computation time relatively yields a better average test accuracy than the steady-state genetic algorithms. The proposed method could significantly improve the practicability and convenience of the BCI system.

To interact with humans in collaborative environments, machines need to be able to predict (i.e., anticipate) future events, and execute actions in a timely manner. However, the observation of the human limb movements may not be sufficient to anticipate their actions unambiguously. In this work, we consider two additional sources of information (i.e., context) over time, gaze, movement and object information, and study how these additional contextual cues improve the action anticipation performance. We address action anticipation as a classification task, where the model takes the available information as the input and predicts the most likely action. We propose to use the uncertainty about each prediction as an online decision-making criterion for action anticipation. Uncertainty is modeled as a stochastic process applied to a time-based neural network architecture, which improves the conventional class-likelihood (i.e., deterministic) criterion. The main contributions of this paper are four-fold: (i) We propose a novel and effective decision-making criterion that can be used to anticipate actions even in situations of high ambiguity; (ii) we propose a deep architecture that outperforms previous results in the action anticipation task when using the Acticipate collaborative dataset; (iii) we show that contextual information is important to disambiguate the interpretation of similar actions; and (iv) we also provide a formal description of three existing performance metrics that can be easily used to evaluate action anticipation models.Our results on the Acticipate dataset showed the importance of contextual information and the uncertainty criterion for action anticipation. We achieve an average accuracy of 98.75% in the anticipation task using only an average of 25% of observations., 27 pages, 16 figures, Neurocomputing

Face super-resolution (SR) has become an indispensable function in security solutions such as video surveillance and identification system, but the distortion in facial components is a great challenge in it. Most state-of-the-art methods have utilized facial priors with deep neural networks. These methods require extra labels, longer training time, and larger computation memory. In this paper, we propose a novel Edge and Identity Preserving Network for Face SR Network, named as EIPNet, to minimize the distortion by utilizing a lightweight edge block and identity information. We present an edge block to extract perceptual edge information, and concatenate it to the original feature maps in multiple scales. This structure progressively provides edge information in reconstruction to aggregate local and global structural information. Moreover, we define an identity loss function to preserve identification of SR images. The identity loss function compares feature distributions between SR images and their ground truth to recover identities in SR images. In addition, we provide a luminance-chrominance error (LCE) to separately infer brightness and color information in SR images. The LCE method not only reduces the dependency of color information by dividing brightness and color components but also enables our network to reflect differences between SR images and their ground truth in two color spaces of RGB and YUV. The proposed method facilitates the proposed SR network to elaborately restore facial components and generate high quality 8x scaled SR images with a lightweight network structure. Furthermore, our network is able to reconstruct an 128x128 SR image with 215 fps on a GTX 1080Ti GPU. Extensive experiments demonstrate that our network qualitatively and quantitatively outperforms state-of-the-art methods on two challenging datasets: CelebA and VGGFace2., Neurocomputing'2021

State-of-the-art deep neural networks (DNNs) typically consist of several layers of features representations, and especially rely on skip connections to avoid the difficulty of model optimization. Despite the proliferation of different DNN models that employ various forms of skip connections to achieve remarkable results on benchmarking datasets, a concrete explanation for the successful operation and improved generalization capability of these DNNs is surprisingly still lacking. In this paper, we focus on investigating the role of skip connections for training very deep DNNs. Our exposition directly provides interesting insights and new interpretations to the following important questions (i) why model optimization is easier (ii) why model generalization is better. Theoretical results reveal that skip connections allow DNNs to circumnavigate the singularity of latent representations that translate to optimization and generalization problems, which plague models without skip connections referred to as PlainNets. For substantiating our analysis, our investigation puts into context some of the most successful skip-connection based DNNs, which include residual networks (ResNets) and residual network with aggregated features (ResNeXt) in relation to PlainNets. Experimental evaluations of these models support the theoretical analysis.

Facing the challenge of distributed computing on processing large-scale data, this paper proposes a consensus-based decentralized training method with communication compression. First, the decentralized training method is designed based on the decentralized topology to reduce the communication burden on the busiest agent and avoid any agent revealing its locally stored data. The convergence of the decentralized training algorithm is then analyzed, which demonstrates that the decentralized trained model can reach the minimal empirical risk on the whole dataset, without the sharing of data samples. Furthermore, model compression combined with the error-compensated method is considered to reduce communication costs during the decentralized training process. At last, the simulation study shows that the proposed decentralized training with error-compensated communication compression is applicable for both IID and non-IID datasets, and exhibits much better performance than the local training method. Besides, the proposed algorithm with an appropriate compression rate shows comparable performance with decentralized training and centralized training, while saving a lot of communication costs.