Deep predictive learning based on video data (hereinafter referred to as "deep predictive learning") is a research direction of deep learning, being interacted with computer vision and reinforcement learning. It is a key part of intelligent prediction and decision-making systems in weather forecasting, autonomous driving, robotics, and other scenarios, and has become a hot research field of machine learning in recent years. Deep predictive learning follows the self-supervised learning paradigm, using internal constraints from unlabeled video data to learn the underlying spatiotemporal patterns. In this paper, we review the existing deep learning techniques for predictive learning in detail. First, we summarize the research scope and application fields of deep predictive learning. Second, we present the datasets and evaluation metrics commonly used in this research field. Third, we summarize current mainstream deep prediction learning models from three perspectives: predictive models based on observation space, predictive models based on state space, and visual planning methods based on the predictive models. Finally, we discuss the hot issues and future research directions in the field of deep predictive learning.

Due to the domain shift, ship detection in multi-source data suffers from image variations caused by different source sensors. In addition, training a specific model for a particular data source consumes high computational cost, which severely limits its practical application in military and civilian fields. Therefore, designing a universal network to effectively detect ship objects from multi-source remote sensing images has become a research hotspot. To this end, the paper proposes a universal ship detection algorithm based on invariant features, which realizes a universal remote sensing object detection network by fully utilizing the shared knowledge among multi-source data. Our method mainly consists of two parts, i.e., an image-level style transfer network and a feature-level domain adaptive network. Specifically, the former employs style transfer network to generate pseudo-multi-source images that are close to the real distribution, narrow the distribution between multi-source data, and extract the invariant features of multi-source data at the image level; To extract invariant features at the feature level, the latter decouples the multi-source features through adaptive network, and realizes feature reorganization through adaptive weight allocation of domain attention network. We evaluate the proposed method using multiple datasets including NWPU VHR-10, SSDD, HRSC and SAR-Ship-Dataset. Experimental results show that the proposed method alleviates the problem of domain shift by complementing the information between invariant features, and can effectively detect multi-source remote sensing data. The average mAP of our method on the above-mentioned multi-source datasets is 90.8%, which exceeds 1.4%-10.6% compared with the existing mainstream ship object detection methods.

Anomaly detection has an important impact on the development of the electric power industry, and how to detect anomalies based on large-scale power data is a research hotspot. At present, most researches use clustering or neural network to detect anomalies. But these methods ignore the potential relationship between the data and miss some specific important information, and do not fully exploit the potential value of the data. Therefore, an abnormal detection model based on graph attention and transformer is proposed. The model first constructs a heterogeneous information network based on the power data (mainly including user ID, meter ID, user type, electrical current, voltage, power, etc.) collected in the data center; then, in order to reduce the model parameters and avoid the phenomenon of overfitting, on the basis of the graph convolutional network (GCN) model, a non-negative matrix factorization (NNMF) method is introduced to perform similarity learning; finally, a graph attention network (GAT) and Transformer are jointly used to capture the correlation relationships between data, thus improving the detection accuracy. The validation analysis is carried out based on the power data of a region in China. The experimental results show that the proposed method can effectively perform anomaly detection.

Hashing, as an effective data representation technology, has played an important role in dealing with the explosive growth of multimedia data. Due to the advantages of its low storage and high efficiency, it has received more and more attention in the field of multimedia retrieval. At present, multi-modal hashing methods have been well researched and developed in multimedia retrieval tasks. However, most of these methods usually use the inner product of hashing features to reconstruct larger pairwise similarity, aiming to preserve the structural information of the original data, which will bring more complex optimization problems. Besides, some models lack discriminant ability, which leads to limitations in the improvement of retrieval performance. In order to overcome the above-mentioned problems, this paper proposes a new multi-modal fusion hashing method. Under the supervision of category information, Hadamard matrix is ??used to generate target codes for data, and the margin between categories is increased by relaxing strict binary constraints. At the same time, the graph embedding approach is used to promote compactness within the class. The proposed method in this paper not only ensures the strong discriminative ability of the model, but also simplifies the optimization process. The experimental results on three public datasets show that the method proposed in this paper is very effective in multimedia data retrieval, and the average performance is 8.47% higher than that of the optimal comparison method.

With the development of perceptual computing and sensor integration technology, hand gesture motion data collected by various sensor devices provides a new data-driven way for human-computer interaction, and widely used in smart home, telemedicine, virtual reality and other fields. Due to hand gestures have temporality and spatial connectivity, it is necessary to consider spatial connection and long-distance dependence of hand gesture in gesture recognition. However, existing hand gesture recognition models ignore the aforementioned two problems, resulting in low recognition accuracy. Therefore, we propose a spatial-temporal graph neural network based hand gesture recognition model(STGNN-HGR). From the perspective of spatial distribution of sensors, based on the spatial location information of sensors, the model represents spatial correlation of hand gesture data with the help of graph neural networks(GNN), and introduces gated recurrent unit(GRU) to solve temporality and long-distance dependence in dynamic hand gestures, so as to enhance the performance of gesture recognition. The experimental results on a variety of datasets show that our model is feasible and effective.

Cloud detection is a key step in the preprocessing of satellite remote sensing data. This paper proposes a cloud detection method by combining a multilayer perceptron with a radiative transfer model. The method is to identify cloud from moderate resolution satellite image using visible and near-infrared band reflectance information. In this method, firstly, the santa barbara DISORT atmospheric radiative transfer model(SBDART) is used to simulate and obtain datasets of reflectance values for a variety of complex terrestrial surfaces, which provides training samples for the multilayer perceptron. Secondly, the trained network model is used to distinguish cloud pixels from total pixels of the advanced medium Resolution Spectral Imager(MERSI II) image in the FengYun3D satellite MERSI II image, and then verified using vertical feature mask(VFM) product of the cloud-aerosol LIDAR infrared pathfinder satellite observations satellite(CALIPSO) and compared horizontally with the cloud mask product(MYD35) of the moderate resolution imaging spectroradiometer(MODIS). The results show that the accuracy of cloud detection for the multilayer perceptron is 76.25%, and especially this method works best in summer and low latitudes, achieves an accuracy of 91.74% for surface identification near the equator. In this paper, the method is more effective in detecting clouds under complex surface type conditions such as urban, farmland and bare soil, with accuracies of 83.37%, 84.52% and 73.11% respectively, which are higher than the 83.25%, 83.31% and 72.66% of the MYD35 product respectively. To further validate the effectiveness of the multilayer perceptron combined with the radiative transfer model, the training samples obtained from the radiative transfer model simulations are used in the k-nearest neighbors, Naive Bayesian, and Random Forest algorithms, respectively, and compared with the multilayer perceptron algorithm in this paper. The results show that the combination of the multilayer perceptron and the radiative transfer model has a higher accuracy.

Medication recommendation aims to make effective prescriptions based on electronic healthcare records (EHRs) of patients, and assists caregivers in clinical decision making. Obtaining temporal patterns of patient conditions as well as contextual information contained in EHRs are the key issues for the success of recommendation. Existing methods do not take the difference in the amount of medical records of different patients into account, and fails to change the focus or number of iterations during information extraction according to personalized patient conditions. To address these problems, the medication recommendation model adaptive multi-hop reading with selective coverage mechanism (AMHSC) is proposed. The model stores encoded temporal patterns with memory neural networks (MemNN), and applies the selective coverage mechanism to balance attention weights over selected information during the attentive multi-hop reading on MemNN. Meanwhile, AMHSC adaptively determines the number of reading hops on MemNN according to personalized patient conditions. Experiments on real-world clinical dataset demonstrate that AMHSC successfully derives important information from EHRs to build informative patient representations for medication recommendation.

Dynamic reconfiguration is an efficient fault-tolerant approach for integrated modular avionics(IMA) systems. The reconfiguration blueprint defines the application migration and resource reconfiguration scheme in the system failure environment, which is the key to reconfiguring and recovering the system function with minimum cost. How to generate effective reconfiguration blueprints rapidly and automatically in complex multi-level associated failure modes is the difficulty. This paper proposes an IMA system reconfiguration method based on sequential game multi-agent reinforcement learning to solve the problem. The sequential game model is introduced in this method. We define the application software needs to be migrated as the agent in the game. The sequence of sequential game is determined according to the priority of the application software. Aiming at the problem of competition and cooperation among multiple agents in the process of sequential game, the algorithm introduces policy gradient of reinforcement learning and optimizes the reconfiguration effect by controlling the action selection probability in interaction with the environment. The policy gradient Monte Carlo tree search algorithm based on biased estimation is applied to update game strategy, which solves the problems of oscillation, difficulty in convergence, long calculation time of the traditional policy gradient algorithm. Experimental results indicate that compared with differential evolution and Q-learning methods, the proposed algorithm has significant advantages in convergence and efficiency.