Fuxian Liu

In the world, most of the successes are results of long-term efforts. The reward of success is extremely high, but before that, a long-term investment process is required. People who are “myopic” only value short-term rewards and are unwilling to make early-stage investments, so they hardly get the ultimate success and the corresponding high rewards. Similarly, for a reinforcement learning (RL) model with long-delay rewards, the discount rate determines the strength of agent's “farsightedness”. In order to enable the trained agent to make a chain of correct choices and succeed finally, the feasible region of the discount rate is obtained through mathematical derivation in this paper firstly. It satisfies the “farsightedness” requirement of agent. Afterwards, in order to avoid the complicated problem of solving implicit equations in the process of choosing feasible solutions, a simple method is explored and verified by theoreti cal demonstration and mathematical experiments. Then, a series of RL experiments are designed and implemented to verify the validity of theory. Finally, the model is extended from the finite process to the infinite process. The validity of the extended model is verified by theories and experiments. The whole research not only reveals the significance of the discount rate, but also provides a theoretical basis as well as a practical method for the choice of discount rate in future researches.

Aerial scene classification is an active and challenging problem in high-resolution remote sensing imagery understanding. Deep learning models, especially convolutional neural networks (CNNs), have achieved prominent performance in this field. The extraction of deep features from the layers of a CNN model is widely used in these CNN-based methods. Although the CNN-based approaches have obtained great success, there is still plenty of room to further increase the classification accuracy. As a matter of fact, the fusion with other features has great potential for leading to the better performance of aerial scene classification. Therefore, we propose two effective architectures based on the idea of feature-level fusion. The first architecture, i.e., texture coded two-stream deep architecture, uses the raw RGB network stream and the mapped local binary patterns (LBP) coded network stream to extract two different sets of features and fuses them using a novel deep feature fusion model. In the second architecture, i.e., saliency coded two-stream deep architecture, we employ the saliency coded network stream as the second stream and fuse it with the raw RGB network stream using the same feature fusion model. For sake of validation and comparison, our proposed architectures are evaluated via comprehensive experiments with three publicly available remote sensing scene datasets. The classification accuracies of saliency coded two-stream architecture with our feature fusion model achieve 97.79%, 98.90%, 94.09%, 95.99%, 85.02%, and 87.01% on the UC-Merced dataset (50% and 80% training samples), the Aerial Image Dataset (AID) (20% and 50% training samples), and the NWPU-RESISC45 dataset (10% and 20% training samples), respectively, overwhelming state-of-the-art methods.

This paper presents a novel neurodynamic approach for solving generalized eigenvalue problems. A series of neurodynamic systems are proposed for finding all eigenvectors to a given pair (A, B) of matrices. Dynamical analysis shows that each system is globally convergent to an exact eigenvector of the pair (A, B) and hence all the eigenvectors can be found inductively by the proposed neurodynamic systems. It also demonstrates that all the neurodynamic systems are primal in the sense that the system's neural trajectories will never escape from the feasible region when starting at it. By constructing a series of energy functions, all the system's stable point sets are guaranteed to be the eigenvector sets of the matrix pair. Compared with the existing neural network models for the generalized eigenvalue problems, the new approach has two robust features: 1) all the eigenvectors of the given pair of matrices (A, B) can be approached by the proposed neurodynamic systems while the existing neural network models can not; 2) the proposed systems can globally converge to the problem's exact eigenvectors while the existing neural network models do not behave proper stabilities