Search

Recently, actor-critic structure based neural networks are widely used in many reinforcement learning tasks. It consists of two main parts: (i) an actor module which outputs the probability distribution of action, and (ii) a critic module which outputs the predicted value based on the current environment. Actor-critic structure based networks usually need expert demonstration to provide an appropriate pre-training for the actor module, but the demonstration data is often hard or even impossible to obtain. And most of them, such as those used in the maze and robot control tasks, suffer from a lack of proper pre-training and unstable error propagation from the critic module to the actor module, which would result in poor and unstable performance. Therefore, a specially designed module which is called relatively optimal historical information learning (ROHI) is proposed. The proposed ROHI module can record the historical explored information and obtain the relatively optimal information through a customized merging algorithm. Then, the relatively optimal historical information is used to assist in training the actor module during the main learning process. We introduce two complex experimental environments, including the complex maze problem and flipping game, to evaluate the effectiveness of the proposed module. The experimental results demonstrate that the extended models with ROHI can significantly improve the success rate of the original actor-critic structure based models and slightly decrease the number of iteration required to reach the stable phase of value-based networks. [ABSTRACT FROM AUTHOR]

Learning an expressive representation from multi-view data is a crucial step in various real-world applications. In this paper, we propose a semi-supervised multi-view representation learning approach named Co-Embedding. Unlike conventional multi-view representation learning methods use joint to concatenate different views, which ignores information exchange between views and limits the methods' ability to utilize complementarity of multi-view data, Co-Embedding fulfills mutual help between views by coordinating and exchanging information between views. Specifically, we first build a weighted deep metric learning based multi-view representation learning framework in Co-Embedding. This framework models multi-view information through coordinate alignment, which is customized to exploit the complementary information from well-learned representations to help with modeling the under-learned representations. Then, by exploiting consensus property and neighborhood information, we design a multi-view label propagation algorithm that labels unlabeled data for Co-Embedding. Experimental results on seven benchmark multi-view datasets demonstrate the effectiveness of Co-Embedding. [ABSTRACT FROM AUTHOR]

Recently, person re-identification (PR-ID) has attracted numerous of research interest because of its broad applications. However, most of the existing PR-ID models always follow the supervised framework, which requires substantial labeled data. In fact, it is often very hard to get enough labeled training samples in many practical application scenarios. To overcome this limitation, some semi-supervised PR-ID methods have been presented more recently. Although some of these semi-supervised models achieve satisfied results, there is still much space to improve. In this paper, we propose a novel semi-supervised PR-ID by similarity-embedded cycle GANs (SECGAN). Our SECGAN model can learn cross-view features with limited labeled data by using cycle GANs. Simultaneously, to further enhance the ability of cycle GANs so that it can extract more discriminative and robust features, similarity metric subnet and specific features extracting subnet are embedded into cycle GANs. Extensive experiments have been conducted on three public PR-ID benchmark datasets, and the experimental results show that our proposed SECGAN approach outperforms several typical supervised methods and the existing state-of-the-art semi-supervised methods including traditional and deep learning semi-supervised methods. [ABSTRACT FROM AUTHOR]

Video-based person re-identification (re-id), which aims to match people through videos captured by non-overlapping camera views, has attracted lots of research interest recently. In this paper, we first propose a novel hybrid 2D and 3D convolution-based recurrent neural network (HCRN) for video-based person re-id task. Specifically, the 3D convolutional module can explore the local short-term fast-varying motion information, while the recurrent layer can leverage the global long-term spatial–temporal information. Based on HCRN, we design a scale-fusion framework to make full use of features of different scales to further improve the performance of video-based person re-id. More concretely, the scale-fusion framework preserves a complete subnetwork similar to HCRN for each scale to extract features and exchanges information between all subnetworks at several stages of the framework. Besides, we propose a training method called species invasion to further improve the performance of HCRN and scale-fusion framework by utilizing a large amount of unlabeled data. Experimental results on the publicly available PRID 2011, iLIDS-VID and MARS multi-shot pedestrian re-id datasets demonstrate the effectiveness of the proposed HCRN, scale-fusion framework and species invasion training method. [ABSTRACT FROM AUTHOR]

Person re-identification (re-id) plays an important role in video surveillance and forensics applications. In many cases, person re-id should be conducted between video clips, i.e., given a query pedestrian video from one camera, the re-id system should retrieve the video clips containing the same person from other cameras. However, person re-id between videos, which we call video-based person re-id, has not been well studied. In this paper, we propose a visual-appearance-level and spatial-temporal-level dictionary learning (VSDL) approach for video-based person re-id. Specifically, we first employ two kinds of models to represent each walking cycle in the video, i.e., visual-appearance features of all frames within the walking cycle, and a spatial-temporal feature vector. By separately learning a visual-appearance-level dictionary and a spatial-temporal-level dictionary from two kinds of representations, each walking cycle can be represented as a coding coefficient. To enhance the discriminative ability of the obtained coding coefficients, we design a representation coefficient discriminant term for VSDL. Experiments on the public iLIDS-VID and PRID 2011 datasets demonstrate the effectiveness of VSDL. [ABSTRACT FROM AUTHOR]

In practice, huge amount of samples is readily available, while labeled samples are often very limited and too expensive to be easily obtained. Multi-view features usually reveal different types of traits of labeled and unlabeled samples. Semi-supervised multi-view learning is a learning paradigm designed to meet the requirement of learning from complementary information of multiple views of labeled and unlabeled samples. In this paper, we propose a semi-supervised multiple kernel intact discriminant space learning (SMKIDSL) method to discover latent intact feature representations for those samples. SMKIDSL employs correlation discriminant analysis and label regression to fully use class label information for enhancing the discriminant power of latent intact feature representations. In SMKIDSL, multi-view collaboration learning mechanism is utilized to efficiently integrate complementary information of multiple views, which enables optimal view being dominant in learning process. Besides, kernel technique is used to tackle nonlinear issue of original multi-view features for exploiting more discriminant information. Comprehensive experiments are conducted on Caltech 101, LFW, MNIST and RGB-D datasets. And the experimental results demonstrate the effectiveness and efficiency of our proposed method. The robustness of our method is also confirmed by those results. [ABSTRACT FROM AUTHOR]

Kernel discriminant subspace learning technique is effective to exploit the structure of image dataset in the high-dimensional nonlinear space. However, for large-scale image recognition applications, this technique usually suffers from large computational burden. Although some kernel accelerating methods have been presented, how to greatly reduce computing time and simultaneously keep favorable recognition accuracy is still challenging. In this paper, we introduce the idea of parallel computing into kernel subspace learning and build a parallel kernel discriminant subspace learning framework. In this framework, we firstly design a random non-overlapping equal data division strategy to divide the whole training set into several subsets and assign each computational node a subset. Then, we separately learn kernel discriminant subspaces from these subsets without mutual communications and finally select the most appropriate subspace to classify test samples. Under the built framework, we propose two novel kernel subspace learning approaches, i.e., parallel kernel discriminant analysis (PKDA) and parallel kernel semi-supervised discriminant analysis (PKSDA). We show the superiority of the proposed approaches in terms of time complexity as compared with related methods, and provide the fundamental supports for our framework. For experiment, we establish a parallel computing environment and employ three public large-scale image databases as experiment data. Experimental results demonstrate the efficiency and effectiveness of the proposed approaches. [ABSTRACT FROM AUTHOR]

Discriminant subspace learning is a widely used feature extraction technique for image recognition, since it can extract effective discriminant features by employing the class information and Fisher criterion. A crucial research topic on this technique is how to rapidly extract sufficient and effective features. Recently, recursive discriminant subspace learning technique has attracted lots of research interest because it can acquire sufficient discriminant features. Generally, it recursively decomposes image samples and extracts features from a number of decomposed sample sets. The major drawback of most recursive discriminant subspace learning methods is that they calculate the projective vectors one by one, such that they suffer from big computational costs. The recursive modified linear discriminant method and the incremental recursive Fisher linear discriminant method employ a simple solution for this problem, which calculates the class number minus one projective vectors in each recursion. However, this solution produces the unfavorable projective vectors with poor discriminant capabilities, and it cannot provide the terminating criterion for recursive computation and make the projective vectors orthogonal. In this paper, we propose a novel recursive learning approach that is group recursive discriminant subspace learning, which can rapidly learn multiple orthogonal subspaces with each spanned by a group of projective vectors. And we present a rule to select favorable projective vectors per recursion and provide a matrix-form-based terminating criterion to determine the number of recursions. Experiments on three widely used databases demonstrate the effectiveness and efficiency of the proposed approach. [ABSTRACT FROM AUTHOR]