Search

With the rapid development of Internet technology and the explosive growth of digital text, opinion mining has become one of the important research hotspots in the field of natural language processing (NLP). In recent years, neural network based deep learning algorithms have been applied in the field of opinion mining. Considering the relation between temporal and spatial dimensions of text data and the characteristics of natural language itself, traditional deep learning algorithms cannot be comprehensive in the processing of fully feature extraction. In this paper, we propose a new deep learning framework for opinion mining, which includes a temporal feature extraction layer that consists of two layers of bidirectional simple recurrent unit (Bi-SRU) networks extracting features at the word and grammar levels; a semantic feature extraction layer that mainly contains a multi-head attention module; a spatial feature extraction layer with dilated convolution that is used to extract opinion preference features. The Internet movie database (IMDb) is used to verify the performance of the proposed framework. The experiment results show that the proposed framework can effectively improve the classification accuracy, whose performance is better than that of the compared algorithms. [ABSTRACT FROM AUTHOR]

• This work is the first one to learn discriminative features from two groups of features, and two groups of features have some differences, such as distribution and dimension. • Based on group sparse learning model, we propose a kernelized two-groups sparse learning (KTGSL) model to learn two groups of weights for two groups of features and select more discriminative features according the learned weights. • We propose two learning strategies to learn weights for KTGSL, and by adjusting two penalty coefficients, KTGSL can process two groups of features more flexibly to balance the differences. • Our work automatically determines the effective HSDGs and refines LBP-TOP to replace the previous work that manually determines single optimal HSDG by testing every HSDG in turn. • We eliminate the mechanized concatenation between LBP-TOP and the single HSDG and consider the interaction between two groups of features. As a branch of affective computing and machine learning, recognizing micro-expressions is more difficult than recognizing macro-expressions because micro-expression has a small motion and short duration. A large number of features and methods have been proposed, and feature extraction is a critical focus of research. For improving performance, feature fusion is an effective strategy that involves two groups of features, and two groups of features usually have some differences, such as discriminability, distribution and dimension. In addition, the extracted features usually have redundant or misleading feature information. Thus, before feature fusion, an algorithm that can automatically learn and select discriminative features from two groups of different features is needed. In this paper, we propose a kernelized two-groups sparse learning (KTGSL) model to automatically learn more discriminative features from two groups of features. We propose two learning strategies to learn the weights: one is that the weights of one group of features are fixed and don't be learned; the other one is that both groups of weights are learned and the two are given to different penalty coefficients, which can flexibly adjust the interrelation between the two groups of features by adjusting the two penalty coefficients. This work is the first one to select discriminative features from two groups of features in micro-expression recognition. The experiments are conducted on three datasets (CASME II, SMIC and SAMM). The experimental results show that our method can automatically select discriminative features from two groups of features and achieve state-of-the-art performance. [ABSTRACT FROM AUTHOR]

• Both labeled data and unlabeled images are used to reduce the cost of labeling the flame hole image artificially. Making full use of the efficiency of CNN feature extraction and the comprehensiveness of fusion, as well as the high computation efficiency of ELM training process. • Laplacian regularization is utilized in the paper for local geometry information extraction, the proposed CNN-LapsELM can fully excavate information which contained in all available data. • The fusion of original calculated features for top-level ELM training process can greatly improve the accuracy for SD identification, and the experiment result indicates the SD accuracy is high to 87%. The superheat degree (SD) in industrial aluminum electrolysis cell is a critical index that can maintain the energy balance, improve the current efficiency and improve production. However, the existing SD identification is mainly relying on artificial experience and the accuracy of SD is far from satisfactory. Further, artificial costs and physical equipment are expensive and time-consuming. In this paper, we propose a deep soft sensor method for SD detection. First, CNN is utilized for flame hole image feature extraction. Second, a semi-supervised extreme learning machine (ELM) that integrates Laplacian regularization is further used for SD classification. The main contributions of the paper are: (1) The proposed CNN-LapsELM utilizes the CNN for flame hole image feature extraction and then ELM for further classification, which fully takes advantage of CNN's ability for complex feature extraction, ELM's excellent generalization ability, and high computation efficiency. (2) Both the labeled and unlabeled samples are utilized for the CNN-LapsELM training process. It fully leverages the information contained in unlabeled data. At the same time, Laplacian regularization is utilized for learning the manifold structure of hole image samples, so the performance of the proposed CNN-LapsELM are improved. (3) The proposed CNN-LapsELM algorithm improves the generalization ability and robustness. The comparison result demonstrates that the CNN-LapsELM is superior to the existing SD identification and the accuracy is 87%. [ABSTRACT FROM AUTHOR]

• Method: Pre-trained Model Knowledge Distillation (PMKD) for incremental learning. • Feature representation knowledge transferred via PMKD. • PMKD combined with replay yields competitive performance in incremental learning. The incremental learning paradigm in machine learning has consistently been a focus of academic research. It is similar to the way in which biological systems learn, and reduces energy consumption by avoiding excessive retraining. Existing studies utilize the powerful feature extraction capabilities of pre-trained models to address incremental learning, but there remains a problem of insufficient utilization of neural network feature knowledge. To address this issue, this paper proposes a novel method called Pre-trained Model Knowledge Distillation (PMKD) which combines knowledge distillation of neural network representations and replay. This paper designs a loss function based on centered kernel alignment to transfer neural network representations knowledge from the pre-trained model to the incremental model layer-by-layer. Additionally, the use of memory buffer for Dark Experience Replay helps the model retain past knowledge better. Experiments show that PMKD achieved superior performance on various datasets and different buffer sizes. Compared to other methods, our class incremental learning accuracy reached the best performance. The open-source code is published at https://github.com/TianSongS/PMKD-IL. [ABSTRACT FROM AUTHOR]

• A lightweight and efficient neural network for real-time semantic segmentation. • An efficient split convolution increase the speed of inference and improve the accuracy of feature extraction. • A refined dual attention mechanism can reduce the complexity of the model and improve the accuracy. • Our LRDNet demonstrates a good trade-off in the parameter size, computational cost and accuracy on the Cityscapes dataset. • With a parameter value below 0.66 M, our model can be up to 77 FPS on GTX 1080ti. Most of the current popular semantic segmentation convolutional networks are focus on accuracy and require large amount of computation, which is using complex models. In order to realize real-time performance in practical applications, such as embedded systems and mobile devices, lightweight semantic segmentation has become a new need, where the network model should keep good accuracy in very limited computing budget. In this paper, we propose a lightweight network with the refined dual attention decorder (termed LRDNet) for better balance between computational speed and segmentation accuracy. In the encoding part of LRDNet, we offer an asymmetric module based on the residual network for lightweight and efficiency. In this module, a combination of decomposition convolution and deep convolution is used to improve the efficiency of feature extraction. In the decoding part of LRDNet, we use a refined dual attention mechanism to reduce the complexity of the entire network. Our network attained precise real-time segmentation results on Cityscapes, CamVid datasets. Without additional processing and pretraining, the LRDNet model achieves 70.1 Mean IoU in the Cityscapes test set. With a parameter value below 0.66 M, it can be up to 77 FPS. [ABSTRACT FROM AUTHOR]

In text classification models based on deep learning, feature extraction and feature aggregation are two key steps. As one of the basic feature extraction methods, CNN has certain limitations due to its inability to effectively extract temporal features from text data. Using max-pooling can significantly reduce the amount of calculation while performing feature aggregation, but it will have an adverse effect on the classification results due to the loss of some text features. In this paper, in response to the above two issues, a Bidirectional Gated Temporal Convolutional Attention(BG-TCA) model is proposed. In the feature extraction stage, the BG-TCA model uses the bidirectional TCN to extract the bidirectional temporal features in text data, and a gating mechanism similar to the LSTM is added between the convolution layers. In the feature aggregation stage, the BG-TCA model uses the attention mechanism to replace the max-pooling method, which makes it possible to distinguish the importance of different features while retaining the text features to the maximum. Finally, experimental results on five benchmark datasets show that the classification accuracy of the BG-TCA model has been greatly improved compared to basic models, and is better than several other state-of-the-art models. [ABSTRACT FROM AUTHOR]

In multi-modal learning tasks such as video understanding, the most important operations are feature extraction, feature enhancement for single modality and feature aggregation between modalities. In this paper, we present two attention based algorithms, the Position-embedding Non-local (PE-NL) Network and the Multi-modal Attention (MA) feature aggregation method. Inspired by Non-local Neural Networks and Transformers, our PE-NL is a self-attention liked feature enhancement operation and it can capture long-range dependencies and model relative positions. The MA aggregation method merges visual and audio modals while reduces feature dimension and the number of parameters without losing too much accuracy. Both of PE-NL and MA blocks can be plugged into many multi-modal learning architectures. Our Gated PE-NL-MA network achieves competitive results on Youtube-8M dataset. [ABSTRACT FROM AUTHOR]

In the modern manufacturing industry, remaining useful life (RUL) prediction of the machining tools plays a significant role in promoting machining efficiency, ensuring product quality and reducing production costs. In recent years, many data-driven prognostic approaches have been developed for machining tools, but few of them have considered the operating conditions such as spindle load and rotating speed that may have great impact on the degradation behavior and sensor signals. It may give rise to more uncertainty and lead to an obvious decrease in prediction accuracy when operating condition changes. Besides, feature extraction from the raw signals that are nonstationary, nonlinear, and mixed with abundant noise is essential but quite challenging. To address these issues, this paper proposes a novel prognostic approach for machining tools under dynamic operating condition with varying spindle load. In the proposed approach, an adaptive variational mode decomposition (VMD) is newly developed to adaptively search the optimal parameters for processing the raw vibration data, then several components with good trendability and noise robustness are obtained for feature extraction. Furthermore, a deep learning model combining one-dimensional convolutional long short-term memory (LSTM) with attention mechanism is constructed to perform RUL prediction. Numerical experiments on a real-world case study show the effectiveness and superiority of the proposed approach in comparison with other baseline data-driven approaches. [ABSTRACT FROM AUTHOR]

This paper aims to address the human action recognition issue by using convolutional long short-term memory networks (Conv-LSTM) and fully-connected LSTM (FC-LSTM) with different attentions. To this end, the spatial-temporal dual-attention network (STDAN), which is mainly composed of feature extraction, attention and fusion modules, is designed. Different from the features of high-level fully-connected layer mostly used in previous work, the features of convolution and fully-connected layers of convolutional neural network (CNN) are both extracted in STDAN, which can enrich the initial level of video representation. Besides, the Conv-LSTM and FC-LSTM are employed to handle the long-duration sequential features with different temporal context information. To reinforce the spatial-temporal attention ability, a temporal attention module (TAM) and a joint spatial-temporal attention module (JSTAM) are implemented. Through the principle components analysis (PCA) and features fusion, the potential of STDAN is effectively explored and weighted. Finally, the experimental results show that the proposed STDAN has better recognition performance than existing state-of-the-art methods. [ABSTRACT FROM AUTHOR]

In many machine learning tasks, learning a good representation of the data can be the key to building a well-performant solution. This is because most learning algorithms operate with the features in order to find models for the data. For instance, classification performance can improve if the data is mapped to a space where classes are easily separated, and regression can be facilitated by finding a manifold of data in the feature space. As a general rule, features are transformed by means of statistical methods such as principal component analysis, or manifold learning techniques such as Isomap or locally linear embedding. From a plethora of representation learning methods, one of the most versatile tools is the autoencoder. In this paper we aim to demonstrate how to influence its learned representations to achieve the desired learning behavior. To this end, we present a series of learning tasks: data embedding for visualization, image denoising, semantic hashing, detection of abnormal behaviors and instance generation. We model them from the representation learning perspective, following the state of the art methodologies in each field. A solution is proposed for each task employing autoencoders as the only learning method. The theoretical developments are put into practice using a selection of datasets for the different problems and implementing each solution, followed by a discussion of the results in each case study and a brief explanation of other six learning applications. We also explore the current challenges and approaches to explainability in the context of autoencoders. All of this helps conclude that, thanks to alterations in their structure as well as their objective function, autoencoders may be the core of a possible solution to many problems which can be modeled as a transformation of the feature space. [ABSTRACT FROM AUTHOR]

The rapid increase of social media images has made organizing these resources effectively a huge problem. Labeling unlabeled images becomes the crucial division of social image understanding. However, the enhancement of social image sharpness leads to the increase of surface feature dimension. These multidimensional complex features leads to the curse of dimensionality and the difficulty of feature extraction. In this paper, sparse autoencoder is studied to solve the problem of social image understanding, because sparse autoencoder can make these features represent the original data in a refined way, thus avoiding curse of dimensionality as much as possible and significantly improve the understanding effect. First, we explore the dimensional reduction capability of sparse autoencoder, and use sparse autoencoder to get low-dimensional features. Second, for low-dimensional features, an enhanced multi-label classifier is utilized to assign labels with the help of cosine similarity about tags correlation. The ability of dimensionality reduction of sparse autoencoder is proved by mapping matrix of image-label. Finally, we test our approach on several publicly available social media datasets. The results demonstrate that our proposed method is superior to lots of non-deep learning method among three evaluation indexes of social image understanding. [ABSTRACT FROM AUTHOR]

In this paper, we propose a data embedding technique for structured data that allows for the direct application of standard vector-based machine learning models without the need for explicit feature extraction. Our approach relies on multiple notions of data proximity, making it suitable for handling mixed data types or incorporating domain knowledge. Our method also reduces the computational costs of pairwise proximity calculations, resulting in improved efficiency and scalability. We demonstrate the effectiveness of our technique on graph and sequence datasets from the biochemical domain. In particular, we show that the method can efficiently be combined with interpretable machine learning approaches like relevance-based learning vector quantization for sophisticated classification learning. • Embedding scheme enabling standard vector-based learning for structured (bio-)data • Savings in computing time, crucial for direct graph or sequence comparison • Allows task-specific mixing of proximity measures, thus avoiding a feature bias • Tailored to shallow machine learning, particularly Learning Vector Quantization • Model interpretability allows evaluation of embedding and proximity mixing schemes [ABSTRACT FROM AUTHOR]

Abstract In this paper, a novel photovoltaic (PV) forecasting model based on multiple reservoirs echo state network (MR-ESN) is proposed to forecast the power output of PV generation system. Firstly, through the unsupervised learning algorithm of restricted Boltzmann machine, the relative feature of input information can be extracted. According to the forecasting performance evaluation criteria of PV forecasting model, principal component analysis is used to extract the main feature, such that the inputs of MR-ESN and the number of reservoirs of MR-ESN can be determined. Secondly, in order to improve the prediction accuracy, an improved parameter optimization method based on Davidon–Fletcher–Powell (DFP) quasi-Newton algorithm is given to optimize the reservoir parameters of MR-ESN. Thirdly, in order to guarantee that the MR-ESN can be stably applied for PV power forecasting, a sufficient condition of the transient stability of PV forecasting model is given. Finally, a PV power generation forecasting example shows that the proposed PV forecasting model can significantly improve the forecasting performance. [ABSTRACT FROM AUTHOR]

Dimensionality reduction as one of most attractive topics in machine learning research area has aroused extensive attentions in recent years. In order to preserve the local structure of data, most of dimensionality reduction methods consider constructing the relationships among each sample and its k nearest neighbors, and they find the neighbors in original space by using Euclidean distance. Since the data in original space contain some noises and redundant features, finding the neighbors in original space is incorrect and may degrade the subsequent performance. Therefore, how to find the optimal k nearest neighbors for each sample is the key point to improve the robustness of model. In this paper, we propose a novel dimensionality reduction method, named Adaptive Neighborhood MinMax Projections (ANMMP) which finds the neighbors in the optimal subspace by solving Trace Ratio problem in which the noises and redundant features have been removed already. Meanwhile, the samples within same class are pulled together while the samples between different classes are pushed far away in such learned subspace. Besides, proposed model is a general approach which can be implemented easily and applied on other methods to improve the robustness. Extensive experiments conducted on several synthetic data and real-world data sets and achieve some encouraging performance with comparison to metric learning and feature extraction methods, which demonstrates the efficiency of our method. [ABSTRACT FROM AUTHOR]

Cross-modal hashing methods map similar data entities from heterogeneous data sources to binary codes with smaller Hamming distance. However, most existing cross-modal hashing methods learn the hash codes with the hand-crafted features which will not generate optimal hash codes and achieve satisfactory performance. Deep cross-modal hashing methods integrate feature learning and hash coding into an end-to-end learning framework which have achieved promising results. However, these deep cross-modal hashing methods do not well preserve the discriminative ability and the global multilevel similarity in hash learning procedure. In this paper, we propose a global and local semantics-preserving based deep hashing method for cross-modal retrieval. More specifically, a large margin is enforced between similar hash codes and dissimilar hash codes from an inter-modal view to learn discriminative hash codes. Therefore the learned hash codes can well preserve local semantic structure. Sequently, the supervised information with the global multilevel similarity is introduced to learn semantics-preserving hash codes for each intra-modal view. As a consequence, the global semantic structure can be preserved into the hash codes. Furthermore, a consistent regularization constraint is added to generate unified hash codes. Ultimately, the feature learning procedure and the hash coding procedure are integrated into an end-to-end learning framework. To verify the effectiveness of the proposed method, extensive experiments are conducted on several datasets, and the experimental results demonstrate that the proposed method achieves superior performance. [ABSTRACT FROM AUTHOR]

In this paper, we focus on unsupervised feature selection. As we have known, the combination of several feature units into a whole feature vector is broadly adopted for effective object representation, which may inevitably includes some irrelevant/redundant feature units or feature dimensions. Most of the traditional feature selection models can only select the feature dimensions without concerning the intrinsic relationship among different feature units. By taking into consideration the group sparsity of feature dimensions and feature units based on an ℓ 2 , 1 minimization, we propose a new unsupervised feature selection model, unsupervised deep sparse feature selection (UDSFS) in this paper. In comparison with the state-of-the-arts, our UDSFS model can not only select the most discriminative feature units but also assign proper weight to the useful feature dimensions concurrently; moreover, the efficiency and robustness of our UDSFS can be also improved without extracting the discarded irrelevant feature units. For model optimization, we introduce an efficient iterative algorithm to solve the non-smooth, convex model and obtain a global optimization with the convergence rate as O ( 1 / K 2 ) ( K is the iteration number). For the experiments, a new medical endoscopic image dataset, Abnormal Endoscopic Image Detection dataset (AEID), is built for evaluation; we also test our model using two public UCI datasets. Various experiments and comparisons with other state-of-the-arts justified the effectiveness and efficiency of our UDSFS model. [ABSTRACT FROM AUTHOR]

Automated seizure detection using EEG has gained increasing attraction in recent years and appeared more and more helpful in both diagnosis and treatment. How to design an appropriate feature extraction method and how to select an efficient classifier are recognized to be crucial in the successful realization. This paper first proposes a new Mahalanobis-similarity-based feature extraction method on the basis of the Mahalanobis distance and discrete wavelet transformation (DWT). Then in order to further improve the performance, this paper designs a fusion feature (MS-SE-FF) in the feature-fusion level, where the Mahalanobis-similarity-based feature characterizing the similarity between signals and the sample-entropy-based feature characterizing the complexity of signals are combined together. Finally, an automated seizure detection method FF-ELM-SD has been built, which is integrated between the novel fusion feature MS-SE-FF and extreme learning machine (ELM). Experimental results demonstrate that the proposed method FF-ELM-SD does a good job in the epileptic seizure detection while preserving the efficiency and simplicity. [ABSTRACT FROM AUTHOR]

The privacy-preserving data analysis has been gained significant interest across several research communities. The current researches mainly focus on privacy-preserving classification and regression. On the other hand, feature selection is also one of the key problems in data mining and machine learning. However, for privacy-preserving feature selection, the relevant papers are few. In this paper, a local learning-based feature weighting framework is introduced. Moreover, in order to preserve the data privacy during local learning-based feature selection, the objective perturbation and output perturbation strategies are used to produce local learning-based feature selection algorithms with privacy preservation. Meanwhile, we give deep analysis about their privacy preserving property based on the differential privacy model. Some experiments are conducted on benchmark data sets. The experimental results show that our algorithms can preserve the data privacy to some extent and the objective perturbation always obtains higher classification performance than output perturbation when the privacy preserving degree is constant. [ABSTRACT FROM AUTHOR]

The goal of Multiple Kernel Learning (MKL) is to combine kernels derived from multiple sources in a data-driven way with the aim to enhance the accuracy of a target kernel machine. State-of-the-art methods of MKL have the drawback that the time required to solve the associated optimization problem grows (typically more than linearly) with the number of kernels to combine. Moreover, it has been empirically observed that even sophisticated methods often do not significantly outperform the simple average of kernels. In this paper, we propose a time and space efficient MKL algorithm that can easily cope with hundreds of thousands of kernels and more. The proposed method has been compared with other baselines (random, average, etc.) and three state-of-the-art MKL methods showing that our approach is often superior. We show empirically that the advantage of using the method proposed in this paper is even clearer when noise features are added. Finally, we have analyzed how our algorithm changes its performance with respect to the number of examples in the training set and the number of kernels combined. [ABSTRACT FROM AUTHOR]

The unsolicited bulk messages are widespread in the applications of short messages. Although the existing spam filters have satisfying performance, they are facing the challenge of an adversary who misleads the spam filters by manipulating samples. Until now, the vulnerability of spam filtering technique for short messages has not been investigated. Different from the other spam applications, a short message only has a few words and its length usually has an upper limit. The current adversarial learning algorithms may not work efficiently in short message spam filtering. In this paper, we investigate the existing good word attack and its counterattack method, i.e. the feature reweighting, in short message spam filtering in an effort to understand whether, and to what extent, they can work efficiently when the length of a message is limited. This paper proposes a good word attack strategy which maximizes the influence to a classifier with the least number of inserted characters based on the weight values and also the length of words. On the other hand, we also proposes the feature reweighting method with a new rescaling function which minimizes the importance of the feature representing a short word in order to require more inserted characters for a successful evasion. The methods are evaluated experimentally by using the SMS and the comment spam dataset. The results confirm that the length of words is a critical factor of the robustness of short message spam filtering to good word attack. [ABSTRACT FROM AUTHOR]

Research on feature relevance and feature selection problems goes back several decades, but the importance of these areas continues to grow as more and more data becomes available, and machine learning methods are used to gain insight and interpret, rather than solely to solve classification or regression problems. Despite the fact that feature relevance is often discussed, it is frequently poorly defined, and the feature selection problems studied are subtly different. Furthermore, the problem of finding all features relevant for a classification problem has only recently started to gain traction, despite its importance for interpretability and integrating expert knowledge. In this paper, we attempt to unify commonly used concepts and to give an overview of the main questions and results. We formalize two interpretations of the all-relevant problem and propose a polynomial method to approximate one of them for the important hypothesis class of linear classifiers, which also enables a distinction between strongly and weakly relevant features. [ABSTRACT FROM AUTHOR]

Musical tags are used to describe music and are cruxes of music information retrieval. Existing methods for music auto-tagging usually consist of preprocessing phase (feature extraction) and machine learning phase. However, the preprocessing phase of most existing method is suffered either information loss or non-sufficient features, while the machine learning phase depends on heavily the feature extracted in the preprocessing phase, lacking the ability to make use of information. To solve this problem, we propose a content-based automatic tagging algorithm using deep Recurrent Neural Network (RNN) with scattering transformed inputs in this paper. Acting as the first phase, scattering transform extracts features from the raw data, meanwhile retains much more information than traditional methods such as mel-frequency cepstral coefficient (MFCC) and mel-frequency spectrogram. Five-layer RNNs with Gated Recurrent Unit (GRU) and sigmoid output layer are used as the second phase of our algorithm, which are extremely powerful machine learning tools capable of making full use of data fed to them. To evaluate the performance of the architecture, we experiment on Magnatagatune dataset using the measurement of the area under the ROC-curve (AUC-ROC). Experimental results show that the tagging performance can be boosted by the proposed method compared with the state-of-the-art models. Additionally, our architecture results in faster training speed and less memory usage. [ABSTRACT FROM AUTHOR]

Structural damage detection has been an interdisciplinary area of interest for various engineering fields. While the available damage detection methods have been in the process of adapting machine learning concepts, most machine learning based methods extract “hand-crafted” features which are fixed and manually selected in advance. Their performance varies significantly among various patterns of data depending on the particular structure under analysis. Convolutional neural networks (CNNs), on the other hand, can fuse and simultaneously optimize two major sets of an assessment task (feature extraction and classification) into a single learning block during the training phase. This ability not only provides an improved classification performance but also yields a superior computational efficiency. 1D CNNs have recently achieved state-of-the-art performance in vibration-based structural damage detection; however, it has been reported that the training of the CNNs requires significant amount of measurements especially in large structures. In order to overcome this limitation, this paper presents an enhanced CNN-based approach that requires only two measurement sets regardless of the size of the structure. This approach is verified using the experimental data of the Phase II benchmark problem of structural health monitoring which had been introduced by IASC-ASCE Structural Health Monitoring Task Group. As a result, it is shown that the enhanced CNN-based approach successfully estimated the actual amount of damage for the nine damage scenarios of the benchmark study. [ABSTRACT FROM AUTHOR]

Canonical correlation analysis (CCA) is a useful dimensionality reduction method and has been widely used in projecting multi-view data. However, CCA learns from training data consisting of only target classes, ignoring Universum data that belongs to none of the target classes but comes from the same domain as the target classes. Recently, incorporating Universum data in learning has been used to gain more prior knowledge about the application domain and has been shown to achieve favorable improvements. In this paper, we extend CCA with Universum learning for multi-view data and the proposed method is termed as Universum CCA (UCCA). Due to the fact that Universum data in each view does not belong to any target class, correlation between Universum data and target data should be minimized. Consequently, UCCA aims to find basis vectors in multiple views to ensure that correlations between projections of target data are mutually maximized but correlations between projections of Universum data and target data mutually minimized. UCCA can be expressed as a generalized eigenvalue problem and the extracted features express patterns more distinctly. The experimental results on several real-world datasets demonstrate its marked improvements over conventional methods. [ABSTRACT FROM AUTHOR]

The state-of-the-art visual instance retrieval systems are based on learning methods, which require collecting an external dataset and fine-tuning a convolutional neural network for the specific retrieval task. In contrast, non-learning methods just used the pre-trained network on the ImageNet classification task for feature extraction, but have trailed the accuracy of learning based methods thus far. In this paper, we propose a non-learning nonlinear embedding neural codes approach that, for the first time, outperforms the state-of-the-art more complex learning based visual instant retrieval methods. We discover that nonlinear embedding can produce a new feature space, where relevant image features are closer to each other. Compared to previous learning based approaches, we do not need to collect any extra datasets and fine-tune the convolutional neural network, but our method can deliver superior performance in the image retrieval task. In addition, we can also reduce the feature dimensionality during nonlinear embedding, but our method does not reduce the retrieval accuracy with the decrease of feature dimensions. Our experiments on three public available datasets for instance retrieval demonstrate that the proposed method achieved outstanding performance against state-of-the-art methods. Remarkably, our 8D image vector can surpass existing 256D compact representation, which demonstrates that our method can effectively improve the speed and precision of visual instance retrieval systems. [ABSTRACT FROM AUTHOR]

This paper proposes a novel method named Multi-label Double-layer Learning (MDLL) for multi-label cross-modal retrieval task. MDLL includes two stages (layers): L2C (Label to Common) and C2L (Common to Label). In the L2C stage, considering that labels can provide semantic information, we take label information as an auxiliary modality and apply a covariance matrix to represent label similarity in multi-label situation. Thus we can maximize the correlation of different modalities and reduce their semantic gap in the L2C stage. In addition, we find that samples with the same semantic labels may have different contents from users’ view. According to this problem, in the C2L stage, labels are projected to a latent space learned from features of image and text. By this way, the label latent space are more related to the sample’s contents. Then, it is noticed that the samples have same labels but various contents can be decreased. In MDLL, iterative learning of the L2C and C2L stages will improve the discriminative ability greatly and decline the discrepancy between the labels and the contents. To show the effectiveness of MDLL, some experiments are conducted on three multi-label cross-modal retrieval tasks (Pascal Voc 2007, Nus-wide, and LabelMe), on which competitive results are obtained. [ABSTRACT FROM AUTHOR]

Remote sensing and hyperspectral data analysis are areas offering wide range of valuable practical applications. However, they generate massive and complex data that is very difficult to be analyzed by a human being. Therefore, methods for efficient data representation and data mining are of high interest to these fields. In this paper, we introduce a novel pipeline for feature extraction and classification of hyperspectral images. To obtain a compressed representation we propose to extract a set of statistical-based properties from these images. This allows for embedding feature space into fourteen channels, obtaining a significant dimensionality reduction. These features are used as an input for the ensemble learning based on randomized neural networks. We introduce a novel method for forming ensembles of Extreme Learning Machines based on randomized feature subspaces and a trained combiner. It is based on continuous outputs and uses a perceptron-based learning scheme to calculate weights assigned to each classifier and class independently. Extensive experiments carried on a number of benchmarks images prove that using proposed feature extraction and extreme learning ensemble leads to a significant gain in classification accuracy. [ABSTRACT FROM AUTHOR]

Robust texture analysis of multi-modal images is important for practical breast tumor diagnosis applications. Texture features based on ranklet transform were proposed for breast tumor classification of multi-modal images and improved diagnostic performance. However, two limitations still exist in these features. Ranklet transform ignores local characteristics of images which are important for texture feature extraction. In addition, due to application of multi-resolution analysis of ranklet transform, some noises or redundant information may be introduced. These issues may result in performance degradation. To solve these problems, this paper proposes a robust texture feature based on Local Structure Preserving Ranklet (LSP-Ranklet) transform and multi-task learning. First of all, multiple LSP-Ranklet images are generated via LSP-Ranklet transform. In this procedure, the distance-based weighting method is proposed to preserve local structure of images by learning local relevance between pixels. Based on LSP-Ranklet images, texture features based on Gray-Level Co-occurrence Matrix (GLCM) are extracted. To eliminate noises of extracted features, multi-task feature learning is employed to select common feature subsets which are robust for tumor classification of multi-modal images. At last, SVM model is used for tumor classification. Experimental results on our multi-modal breast ultrasound images database demonstrate the effectiveness and robustness of the proposed feature. [ABSTRACT FROM AUTHOR]

Distinct from conventional approaches to generic vehicle type recognition, this paper proposes an approach to fine-grained vehicle type recognition that leverages multiple cues jointly to convolutional neural networks (CNNs). Fine-grained vehicle recognition is inherently difficult with many challenges, including (i) how to incorporate multiple cues through joint optimization instead of separately handling, (ii) how to learn data-driven representation for vehicles instead of hand-crafted features and (iii) how to perceive and localize the vehicle region instead of using the whole image. Our multi-task CNNs localize vehicles in the first stage and recognize subclasses in the second stage, allowing us to handle samples with cluttered backgrounds and those where vehicles do not fill most of the image. Through collaborative feature learning from multiple tasks in each stage, our approach can handle subtle inter-class variations at the subordinate level. We also develop new vehicle-oriented data augmentation strategies in CNN training. To advance research on vehicle-centered tasks, we release a new vehicle dataset that provides both semantic labels and bounding boxes. This dataset can be used in vehicle localization, recognition, viewpoint estimation, and so on. Extensive experiments on two benchmark vehicle datasets demonstrate that our approach outperforms state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Using images that are dispersed in a network can improve image classification performance; however, it is easy to leak the image holders’ privacy. In this paper, we present a novel method that considers both image classification performance and the image holders’ privacy-preservation. Our method contains four stages; the feature extraction stage follows the Bag-of-Features (BoF) framework, but it introduces significantly novel stages in the training, classification and ensemble stages. In the training stage, we split the image classification into many one-vs.-one classification problems; a set of binary classifiers as weak learners are learned on a local machine, to classify each image in an autonomous system. In the classification stage, a local machine sends features of unlabeled images to each image autonomous system, and then, each image autonomous system and local machine finish a prediction task under their own trained model. Afterward, the image autonomous systems send their prediction labels back to the local system. In the ensemble stage, we gather all of the prediction labels from the image autonomous systems and design an ensemble system to ensemble them into final classification results. This ensemble system must consider both the efficiency of the different trained models in each image autonomous system and the efficiency of the trained model on each class in the one-vs.-one classification problems. During the whole process, only the features of unlabeled images and prediction labels are transmitted in the network, which largely reduces the risk of privacy leakage. We conducted experiments on the network environment, to prove that our method has preferable performance. Our first experiment demonstrates that the performance of our method achieves state-of-the-art performance on four public dataset benchmarks, which shows that our method has competitive performance. The second experiment shows that our method achieves state-of-the-art performance with even fewer training samples, which means that there is less communication overhead and better privacy-preservation. [ABSTRACT FROM AUTHOR]

Multi-class weather classification is a fundamental and significant technique which has many potential applications, such as video surveillance and intelligent transportation. However, it is a challenging task due to the diversity of weather and lack of discriminate feature. Most existing weather classification methods only consider two-class weather conditions such as sunny-rainy or sunny-cloudy weather. Moreover, they predominantly focus on a fixed scene such as popular tourism and traffic scenario. In this paper, we propose a novel method for scene-free multi-class weather classification from single images based on multiple category-specific dictionary learning and multiple kernel learning. To improve the discrimination of image representation and enhance the performance of multiple weather classification, our approach extracts multiple weather features and learns dictionaries based on these features. To select a good subset of features, we utilize multiple kernel learning algorithm to learn an optimal linear combination of feature kernels. In addition, to evaluate the proposed approach, we collect an outdoor image set that contains 20 K images, called MWI (Multi-class Weather Image) set. Experimental results show the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Object retrieval has attracted much research attention in recent years. Confronting object retrieval, how to estimate the relevance among objects is a challenging task. In this paper, we focus on view-based object retrieval and propose a multi-scale object retrieval algorithm via learning on graph from multimodal data. In our work, shape features are extracted from each view of objects. The relevance among objects is formulated in a hypergraph structure, where the distance of different views in the feature space is employed to generate the connection in the hypergraph. To achieve better representation performance, we propose a multi-scale hypergraph structure to model object correlations. The learning on graph is conducted to estimate the optimal relevance among these objects, which are used for object retrieval. To evaluate the performance of the proposed method, we conduct experiments on the National Taiwan University dataset and the ETH dataset. Experimental results and comparisons with the state-of-the-art methods demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

The quality of point clouds obtained by RGB-D camera-based indoor mobile mapping can be limited by local degradation because of complex scenarios such as sensor characteristics, partial occlusions, cluttered backgrounds, and complex illumination conditions. This paper presents a machine learning framework to assess the local quality of indoor mobile mapping point cloud data. In our proposed framework, a point cloud dataset with multiple kinds of quality problems is first created by manual annotation and degradation simulation. Then, feature extraction methods based on 3D patches are treated as operating units to conduct quality assessment in local regions. Also, a feature selection algorithm is deployed to obtain the essential components of feature sets that are used to effectively represent local degradation. Finally, a semi-supervised method is introduced to classify quality types of point clouds. Comparative experiments demonstrate that the proposed framework obtained promising quality assessment results with limited labeled data and a large amount of unlabeled data. [ABSTRACT FROM AUTHOR]

Globally, the rate of preterm births are increasing, thus resulting in significant health, development and economic problems. Current methods for the early detection of such births are inadequate. Nevertheless, there has been some evidence that the analysis of uterine electrical signals, collected from the abdominal surface, could provide an independent and easier way to diagnose true labour and detect the onset of preterm delivery. Using advanced machine learning algorithms, in conjunction with Electrohysterography signal processing, numerous studies have focused on detecting true labour several days prior to the event. However, in this paper, the Electrohysterography signals have been used to detect preterm births. This has been achieved using an open dataset, which contains 262 records for women who delivered at term and 38 who delivered prematurely. Several new features from Electromyography studies have been utilised, as well as feature-ranking techniques to determine their discriminative capabilities in detecting term and preterm records. Seven different artificial neural networks were then used to identify these records. The results illustrate that the combination of the Levenberg–Marquardt trained Feed-Forward Neural Network, Radial Basis Function Neural Network and the Random Neural Network classifiers performed the best, with 91% for sensitivity, 84% for specificity, 94% for the area under the curve and 12% for the mean error rate. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel local feature, called Local Orientation Adaptive Descriptor (LOAD), to capture regional texture in an image. In LOAD, we proposed to define point description on an Adaptive Coordinate System (ACS), adopt a binary sequence descriptor to capture relationships between one point and its neighbors and use multi-scale strategy to enhance the discriminative power of the descriptor. The proposed LOAD enjoys not only discriminative power to capture the texture information, but also has strong robustness to illumination variation and image rotation. Extensive experiments on benchmark data sets of texture classification and real-world material recognition show that the LOAD yields the state-of-the-art performance. It is worth to mention that we achieve a superior classification accuracy on Flickr Material Database by using a single feature. Moreover, by combining LOAD with Convolutional Neural Networks (CNN), we obtain significantly better performance than both the LOAD and CNN. This result confirms that the LOAD is complementary to the learning-based features. [ABSTRACT FROM AUTHOR]

3D shape features play a crucial role in graphics applications, such as 3D shape matching, recognition, and retrieval. Various 3D shape descriptors have been developed over the last two decades; however, existing descriptors are handcrafted features that are labor-intensively designed and cannot extract discriminative information for a large set of data. In this paper, we propose a rapid 3D feature learning method, namely, a convolutional auto-encoder extreme learning machine (CAE-ELM) that combines the advantages of the convolutional neuron network, auto-encoder, and extreme learning machine (ELM). This method performs better and faster than other methods. In addition, we define a novel architecture based on CAE-ELM. The architecture accepts two types of 3D shape representation, namely, voxel data and signed distance field data (SDF), as inputs to extract the global and local features of 3D shapes. Voxel data describe structural information, whereas SDF data contain details on 3D shapes. Moreover, the proposed CAE-ELM can be used in practical graphics applications, such as 3D shape completion. Experiments show that the features extracted by CAE-ELM are superior to existing hand-crafted features and other deep learning methods or ELM models. Moreover, the classification accuracy of the proposed architecture is superior to that of other methods on ModelNet10 (91.4%) and ModelNet40 (84.35%). The training process also runs faster than existing deep learning methods by approximately two orders of magnitude. [ABSTRACT FROM AUTHOR]

The increasing number of demanding consumer digital multimedia applications has boosted interest in no-reference (NR) image quality assessment (IQA). In this paper, we propose a perceptual NR blur evaluation method using a new machine learning technique, i.e., extreme learning machine (ELM). The proposed metric, Blind Image Blur quality Evaluator (BIBE), exploits scene statistics of gradient magnitudes to model the properties of blurred images, and then the underlying blur features are derived by fitting gradient magnitudes distribution. The resultant feature is finally mapped into an associated quality score using ELM. As subjective evaluation scores by human beings are integrated into training, machine learning techniques can predict image quality more accurately than those traditional methods. Compared with other learning techniques such as support vector machine (SVM), ELM has better learning performance and faster learning speed. Experimental results on public databases show that the proposed BIBE correlates well with human perceived blurriness, and outperforms the state-of-the-art specific NR blur evaluation metrics as well as generic NR IQA methods. Moreover, the application of automatic focusing system for digital cameras further confirms the capability of BIBE. [ABSTRACT FROM AUTHOR]

The random-hidden-node extreme learning machine (ELM) is a much more generalized cluster of single-hidden-layer feed-forward neural networks (SLFNs) which has three parts: random projection, non-linear transformation, and ridge regression (RR) model. Networks with deep architectures have demonstrated state-of-the-art performance in a variety of settings, especially with computer vision tasks. Deep learning algorithms such as stacked autoencoder (SAE) and deep belief network (DBN) are built on learning several levels of representation of the input. Beyond simply learning features by stacking autoencoders (AE), there is a need for increasing its robustness to noise and reinforcing the sparsity of weights to make it easier to discover interesting and prominent features. The sparse AE and denoising AE was hence developed for this purpose. This paper proposes an approach: SSDAE-RR (stacked sparse denoising autoencoder – ridge regression) that effectively integrates the advantages in SAE, sparse AE, denoising AE, and the RR implementation in ELM algorithm. We conducted experimental study on real-world classification (binary and multiclass) and regression problems with different scales among several relevant approaches: SSDAE-RR, ELM, DBN, neural network (NN), and SAE. The performance analysis shows that the SSDAE-RR tends to achieve a better generalization ability on relatively large datasets (large sample size and high dimension) that were not pre-processed for feature abstraction. For 16 out of 18 tested datasets, the performance of SSDAE-RR is more stable than other tested approaches. We also note that the sparsity regularization and denoising mechanism seem to be mandatory for constructing interpretable feature representations. The fact that a SSDAE-RR approach often has a comparable training time to ELM makes it useful in some real applications. [ABSTRACT FROM AUTHOR]

Sparse representation, a locality-based data representation method, leads to promising results in many scientific and engineering fields. Meanwhile in the study of feature selection, locality preserving is widely recognized as an effective measurement criterion. In this paper, we introduce l 1 -norm driven sparse representation into feature selection, and propose a novel joint feature weights learning algorithm, named sparse discriminative feature weights (SDFW). SDFW assigns the highest score to the feature that has the smallest difference between within-class reconstruction residual and between-class reconstruction residual in the space of selected features. It possesses the following advantages: (1) compared with feature selection methods based on k nearest neighbors, SDFW automatically (vs. manually) determines neighborhood for individual sample; (2) compared with conventional heuristic feature search which selects features individually, SDFW selects feature subset in batch mode. Extensive experiments on different data types demonstrate the effectiveness of SDFW. [ABSTRACT FROM AUTHOR]

Recently, hand-designed local descriptors like spatiotemporal Gabor filters and VLBP have been successfully applied in video-based facial expression recognition. One major drawback of these methods is that they are hard to generalize to different problems. In this paper, we propose a new video-based facial expression recognition method by automatically learning features from video data. Specifically, we use sparse coding algorithm to learn spatiotemporal features from unlabeled facial expression videos. For modeling spatiotemporal layout information embedded in facial expressions to improve recognition performance, we extend the idea of spatial pyramid matching (SPM) into video case, and perform spatiotemporal pyramid feature pooling following sparse coding feature extraction. Experimental results on widely used Cohn–Kanade database show that the classification performance can be improved effectively by considering spatiotemporal layout of facial expressions, and our method outperforms popular methods using hand-designed features. [ABSTRACT FROM AUTHOR]

As an important learning scheme for Multi-Instance Learning (MIL), the Instance Prototype (IP) selection-based MIL algorithms transform bags into a new instance feature space and achieve impressed classification performance. However, the number of IPs in the existing algorithms linearly increases with the scale of the training data. The performance and efficiencies of these algorithms are easily limited by the high dimension and noise when facing a large scale of training data. This paper proposes a discriminative subspaces-based instance prototype selection method that is suitable for reducing the computation complexity for large scale training data. In the proposed algorithm, we introduce the low-rank matrix recovery technique to find two discriminative and clean subspaces with less noise; then present a ℓ 2 , 1 norm-based self-expressive sparse coding model to select the most representative instances in each subspace. Experimental results on several data sets show that our algorithm achieves superior and stable performance but with lower dimension compared with other IP selection strategies. [ABSTRACT FROM AUTHOR]

This paper presents a novel computer-aided diagnosis (CAD) system for the diagnosis of breast cancer based on extreme learning machine (ELM). In view of a mammographic image, it is first eliminated interference in the preprocessing stages. Then, the preprocessed images are segmented by the level set model we proposed. Subsequently, a model of multidimensional feature vectors is built. Since not every feature vector contributes to the improvement of performance, feature selection is done by the combination of support vector machine (SVM) and extreme learning machine (ELM). Finally, an optimal subset of feature vectors is inputted into the classifiers for distinguishing malignant masses from benign ones. We also compare our breast mass classification approach based on ELM with several state-of-the-art classification models, and the results show that the proposed CAD system not only has good performance in terms of specificity, sensitivity and accuracy, but also achieves a significant reduction in training time compared with SVM and particle swarm optimization-support vector machine (PSO-SVM). Ultimately, our system achieves the better performance with average accuracy of 96.02% which indicates that the proposed segmentation model, the utilization of selected feature vectors and the effective classifier ELM provide satisfactory system. [ABSTRACT FROM AUTHOR]

When handling pattern classification problem such as face recognition and digital handwriting identification, image data is always represented to high dimensional vectors, from which discriminant features are extracted using dimensionality reduction methods. So in this paper, we present a supervised manifold learning based dimensionality reduction method named constrained discirminant neighborhood embedding (CDNE). In the proposed CDNE, on one hand, the class information of samples is taken into account to construct both an inter-class graph and an intra-class graph, where neighborhood points in the intra-class graph are selected from those with the same class and any point in the inter-class graph should sample those labeled different classes as its neighborhood points. On the other hand, locally least linear reconstruction technique is also introduced to model an objective function under the local uncorrelation constraint to explore a discriminant subspace. Compared to some related and state-of-the-art dimensionality reduction methods such as discriminant neighborhood embedding (DNE), supervised locality discriminant manifold learning (SLDML), discriminant sparse neighborhood preserving embedding (DSNPE), local graph embedding based on maximum margin criterion (LGE/MMC), uncorrelated discriminant locality preserving projection (UDLPP) and locally uncorrelated discriminant projection (LUDP), the proposed CDNE has been validated to be efficient and feasible by experimental results on some benchmark face data sets including CMU PIE, ORL and FERET. [ABSTRACT FROM AUTHOR]

Manifold learning and Sparse Representation Classifier are two popular techniques for face recognition. Because manifold learning can find low-dimensional representations for high-dimensional data, it is widely applied in computer vision and pattern recognition. Most of the manifold learning algorithms can be unified in the graph embedding framework, where the first step is to determine the adjacent graphs. Traditional methods use k nearest neighbor or the ε -ball schemes. However, they are parametric and sensitive to noises. Moreover, it is hard to determine the size of appropriate neighborhoods. To deal with these problems, in this paper, Graph Regularized Sparsity Discriminant Analysis, GRSDA, for short, is proposed. Based on graph embedding and sparsity preserving projection, the weight matrices for intrinsic and penalty graphs are obtained through sparse representation. GRSDA seeks a subspace in which samples in intra-classes are as compact as possible while samples in inter-classes are as separable as possible. Specifically, samples in the low-dimensional space can preserve the sparse locality relationship in the same class, while enhancing the separability for samples in different classes. Hence, GRSDA can achieve better performance. Extensive experiments were carried out on ORL, YALE-B and AR face databases, and the results confirmed that the proposed algorithm outperformed LPP, UDP, SPP and DSNPE. [ABSTRACT FROM AUTHOR]

In this paper, we present an efficient concept detection system based on a novel bag of words extraction method and sparse ensemble learning. The presented system can efficiently build the concept detectors upon large scale image dataset, and achieve real-time concept detection on unseen images with the state-of-the-arts accuracy. To do so, we first develop an efficient bag of visual words (BoW) construction method based on sparse non-negative matrix factorization (NMF) and GPU enabled SIFT feature extraction. We then develop a sparse ensemble learning method to build the detection model, which drastically reduces learning time in order of magnitude over traditional methods like Support Vector Machine. To overcome the difficulty of manual annotation of training dataset, we construct a large training set with both pseudo relevance feedback of negative samples and interactive feedback of positive samples. Experiments on TRECVID 2012 dataset and MIRFlickr-1M dataset show both efficiency and effectiveness of our system. [ABSTRACT FROM AUTHOR]

This paper introduces an approach that combines machine learning and adaptive hardware to improve the efficiency of ultra-low-power sensor interfaces. Adaptive feature extraction circuits are assisted by hardware embedded training to dynamically activate only the most relevant features. This selection is done in a context- and power cost-aware manner, through modification of the C4.5 algorithm. As proof-of-principle, a Voice Activity Detector illustrates the context-dependent relevance of features, demonstrating average circuit power savings of 70%, without accuracy loss. The RECAS database developed for experimenting with this context- and dynamic resource-cost-aware training is presented and made open-source for the research community. [ABSTRACT FROM AUTHOR]

Feature combination is an effective way for image classification. Most of the work in this line mainly considers feature combination based on different low-level image descriptors, while ignoring the complementary property of different higher-level image features derived from the same type of low-level descriptor. In this paper, we explore the complementary property of different image features generated from one single type of low-level descriptor for image classification. Specifically, we propose a soft salient coding (SSaC) method, which overcomes the information suppression problem in the original salient coding (SaC) method. We analyse the physical meaning of the SSaC feature and the other two types of image features in the framework of Spatial Pyramid Matching (SPM), and propose using multiple kernel learning (MKL) to combine these features for classification tasks. Experiments on three image databases (Caltech-101, UIUC 8-Sports and 15-Scenes) not only verify the effectiveness of the proposed MKL combination method, but also reveal that collaboration is more important than selection for classification when limited types of image features are employed. [ABSTRACT FROM AUTHOR]

In this paper, we propose multi-layer non-negative matrix factorization (NMF) network for classification task, which provides intuitively understandable hierarchical feature learning process. The layer-by-layer learning strategy was adopted through stacked NMF layers, which enforced non-negativity of both features and their coefficients. With the non-negativity constraint, the learning process revealed latent feature hierarchies in the complex data in intuitively understandable manner. The multi-layer NMF networks was investigated for classification task by studying various network architectures and nonlinear functions. The proposed multilayer NMF network was applied to document classification task, and demonstrated that our proposed multi-layer NMF network resulted in much better classification performance compared to single-layered network, even with the small number of features. Also, through intuitive learning process, the underlying structure of feature hierarchies was revealed for the complex document data. [ABSTRACT FROM AUTHOR]

Accurately identifying Internet traffic at the early stage is very important for the applications of traffic identification. Recent years, more and more research works have tried to build effective machine learning models to identify an Internet flow with the few packets at its early stage. However, a basic and important problem still needs to be studied in depth, that is how many packets are most effective in early stage Internet traffic identification. In this paper, we try to resolve this problem. Three Internet traffic data sets are applied. And the sizes of the first 10 packets are extracted for study. We firstly apply mutual information to analyze the information that the first n packets provide to the flow type. Then correlation analysis of each pair of adjacent packets is carried out to find out the feature redundancies. And then we execute a number of crossover identification experiments with different numbers of packets using 11 well-known supervised learning algorithms. Finally, statistical tests are applied for the experimental results to find out which number is the best performed one. Our experimental results show that 5–7 are the best packet numbers for early stage traffic identification. [ABSTRACT FROM AUTHOR]

In this paper, a novel framework is proposed for dynamic textures (DTs) recognition by learning a high level feature using deep neural network (DNN). The insight behind the method is that a DT appearing in different videos should share similar features, which can be learned for better recognition performance. Unlike many prior works only focus on low level or middle level features, we propose a novel high level feature learning method using DNN. Our goal is to construct a compact and discriminative semantic feature. The conventional bag of features approach using k-means is not semantically meaningful since the clustering criterion is based on appearance similarity. The proposed framework can effectively overcome the problem by capturing the semantic relations of the middle level by DNN. Extensive experiments with qualitative and quantitative results demonstrate the efficacy of our approach. [ABSTRACT FROM AUTHOR]