Search

As compared to actions, activities are much more complex, but semantically they are more representative of a human׳s real life. Techniques for action recognition from sensor-generated data are mature. However, few efforts have targeted sensor-based activity recognition. In this paper, we present an efficient algorithm to identify temporal patterns among actions and utilize the identified patterns to represent activities for automated recognition. Experiments on a real-world dataset demonstrated that our approach is able to recognize activities with high accuracy from temporal patterns, and that temporal patterns can be used effectively as a mid-level feature for activity representation. [ABSTRACT FROM AUTHOR]

Conventional models relying on similarity utilizing low-level surface statistical, syntactic and lexical semantic features are suboptimal in contradiction recognition, especially for the large-scale data, such as the sensor and traffic data. To tackle this problem, this work treats the text and hypothesis pair as event graph and proposes a novel model based on event graph to recognize contradiction in big data collection. The proposed model is capable of seamlessly sewing the high-level event semantic features corresponding to the conflicting linguistic phenomena to identify contradictory construction. Experimental results show that our event graph based contradiction recognition model achieves significant improvement as compared to state-of-the-art competitors. [ABSTRACT FROM AUTHOR]

• The novel reconstruction regularization term can preserve the essential information. • Low-rank constraint can well explore the correlation among samples. • An efficient algorithm is presented to optimize the problem with convergence guarantee. • RRLSL can be applied to both supervised and unsupervised situations. • Superior performance is achieved with lowest computational complexity. With the rapid increase of multi-modal data through the internet, cross-modal matching or retrieval has received much attention recently. It aims to use one type of data as query and retrieve results from the database of another type. For this task, the most popular approach is the latent subspace learning, which learns a shared subspace for multi-modal data, so that we can efficiently measure cross-modal similarity. Instead of adopting traditional regularization terms, we hope that the latent representation could recover the multi-modal information, which works as a reconstruction regularization term. Besides, we assume that different view features for samples of the same category share the same representation in the latent space. Since the number of classes is generally smaller than the number of samples and the feature dimension, therefore the latent feature matrix of training instances should be low-rank. We try to learn the optimal latent representation, and propose a reconstruction based term to recover original multi-modal data and a low-rank term to regularize the learning of subspace. Our method can deal with both supervised and unsupervised cross-modal retrieval tasks. For those situations where the semantic labels are not easy to obtain, our proposed method can also work very well. We propose an efficient algorithm to optimize our framework. To evaluate the performance of our method, we conduct extensive experiments on various datasets. The experimental results show that our proposed method is very efficient and outperforms the state-of-the-art subspace learning approaches. [ABSTRACT FROM AUTHOR]

Cross-modal retrieval has recently attracted much interest along with the rapid development of multimodal data, and effectively utilizing the complementary relationship of different modal data and eliminating the heterogeneous gap as much as possible are the two key challenges. In this paper, we present a novel network model termed cross-modal Dual Subspace learning with Adversarial Network (DSAN). The main contributions are as follows: (1) Dual subspaces (visual subspace and textual subspace) are proposed, which can better mine the underlying structure information of different modalities as well as modality-specific information. (2) An improved quadruplet loss is proposed, which takes into account the relative distance and absolute distance between positive and negative samples, together with the introduction of the idea of hard sample mining. (3) Intra-modal constrained loss is proposed to maximize the distance of the most similar cross-modal negative samples and their corresponding cross-modal positive samples. In particular, feature preserving and modality classification act as two antagonists. DSAN tries to narrow the heterogeneous gap between different modalities, and distinguish the original modality of random samples in dual subspaces. Comprehensive experimental results demonstrate that, DSAN significantly outperforms 9 state-of-the-art methods on four cross-modal datasets. • Dual parallel subspaces are proposed, which can better mine the underlying structure information of different modalities as well as modality-specific information. • An improved quadruplet loss is proposed, which integrates relative distance, absolute distance and hard sample mining. On the one hand, it pushes forward the boundaries of positive and negative samples to a certain extent. The introduction of the idea of hard sample mining reduces the complexity of the model and further improves its performance. • An Intra-modal constrained loss is proposed to maximize the distance of the most similar cross-modal negative sample and its corresponding cross-modal positive samples. [ABSTRACT FROM AUTHOR]

Abstract As a robust and heuristic technique in machine learning, active learning has been established as an effective method for addressing large volumes of unlabeled data; it interactively queries users (or certain information sources) to obtain desired outputs at new data points. With regard to deep learning techniques (e.g., CNN) and their applications (e.g., image classification), labeling work is of great significance as training processes for obtaining parameters in neural networks which requires abundant labeled samples. Although a few active learning algorithms have been proposed for devising certain straightforward sampling strategies (e.g., density, similarity, uncertainty, and label-based measure) for deep learning algorithms, these employ onefold sampling strategies and do not consider the relationship among multiple sampling strategies. To this end, we devised a novel solution " m ulti- c riteria a ctive l eep l earning"(MCADL) to learn an active learning strategy for deep neural networks in image classification. Our sample selection strategy selects informative samples by considering multiple criteria simultaneously (i.e., density, similarity, uncertainty, and label-based measure). Moreover, our approach is capable of adjusting weights adaptively to fuse the results from multiple criteria effectively by exploring the utilities of the criteria at different training stages. Through extensive experiments on two popular image datasets (i.e., MNIST and CIFAR-10), we demonstrate that our proposed method consistently outperforms highly competitive active learning approaches; thereby, it can be verified that our multi-criteria active learning proposal is rational and our solution is effective. [ABSTRACT FROM AUTHOR]

Recent years have witnessed the fast development of multimedia platforms in China, such as Youku, LeTV and Weibo. Images and videos are usually uploaded with textual descriptions, such as titles and introductions of these media. These texts are the key to multimedia content understanding, and this paper is dedicated to multimedia understanding with visual content entailment via recognizing semantic entailment in these texts. In fact, the natural language processing community has been manifesting increasing interest in semantic entailment recognition in English texts. Yet, so far not much attention has been paid to semantic entailment recognition in Chinese texts. Therefore, this paper investigates on multimedia semantic entailment with Chinese texts. Recognizing semantic entailment in Chinese texts can be cast as a classification problem. In this paper, a classification model is constructed based on support vector machine to detect high-level semantic entailment relations in Chinese text pair, including entailment and non-entailment for the Binary-Class and forward entailment, reverse entailment, bidirectional entailment, contradiction and independence for the Multi-Class. We explore different semantic feature combinations based on three kinds of Chinese textual features, including Chinese surface textual, Chinese lexical semantic and Chinese syntactic features, and utilize them to feed our classification model. The experiment results show that the accuracy of our classification model for semantic entailment recognition with the feature combination using all the three kinds of Chinese textual features achieves a much better performance than the baseline in Multi-Class and slightly better results than the baseline in the Binary-Class. [ABSTRACT FROM AUTHOR]

Music emotion recognition (MER) is an important topic in music understanding, recommendation, retrieval and human computer interaction. Great success has been achieved by machine learning methods in estimating human emotional response to music. However, few of them pay much attention in semantic interpret for emotion response. In our work, we first train an interpretable model between acoustic audio and emotion. Filter, wrapper and shrinkage methods are applied to select important features. We then apply statistical models to build and explain the emotion model. Extensive experimental results reveal that the shrinkage methods outperform the wrapper methods and the filter methods in arousal emotion. In addition, we observed that only a small set of the extracted features have the key effects to arousal. While, most of our extracted features have small contribution to valence music perception. Ultimately, we obtain a higher average accuracy rate in arousal, compared to that in valence. [ABSTRACT FROM AUTHOR]

In this paper, hybrid chromosome genetic algorithm is applied to establishing the real-time classification model for surface defects in a large-scale strip steel image collection. After image preprocessing, four types of visual features, comprising geometric feature, shape feature, texture feature and grayscale feature, are extracted from the defect target image and its corresponding preprocessed image. In order to use genetic algorithm to optimize classification model based on hybrid chromosome, the structure of hybrid chromosome is designed to seamlessly integrate the kernel function, visual features and model parameters. And then the chromosome and the SVM classification model will be evolved and optimized according to the genetic operations and the fitness evaluation. In the end, the final SVM classifier is established using the decoding result of the optimal chromosome. The experimental results show that our method is effective and efficient in classifying the surface defects in a large-scale strip steel image collection. [ABSTRACT FROM AUTHOR]

The Internet has brought about fundamental changes in the way peoples generate and exchange media information. Over the last decade, unsolicited message images (image spams) have become one of the most serious problems for Internet service providers (ISPs), business firms and general end users. In this paper, we report a novel system called RoBoTs ( Ro bust Bo os T rap based s pam detector) to support accurate and robust image spam filtering. The system is developed based on multiple visual properties extracted from different levels of granularity, aiming to capture more discriminative contents for effective spam image identification. In addition, a resampling based learning framework is developed to effectively integrate random forest and linear discriminative analysis (LDA) to generate comprehensive signature of spam images. It can facilitate more accurate and robust spam classification process with very limited amount of initial training examples. Using three public available test collections, the proposed system is empirically compared with the state-of-the-art techniques. Our results demonstrate its significantly higher performance from different perspectives. [ABSTRACT FROM AUTHOR]