In recent years, Deep CNN (DCNN) models have achieved great success in the field of computer vision. However, such models are still considered to lack interpretability. One of fundamental issues underlying this problem can be noted as follows: The decision-making of a DCNN model is considered as a “black-box” operation. In this study, we propose to use binary tree structure convolution layers (TSCL) to interpret the decision-making mechanism of a DCNN model in the image recognition task. First, we design a TSCL module, in which each parent layer generates two child layers, and then integrate them into a normal DCNN. Second, we design an information coding objective to guide each two child nodes of one parent node to learn the particular information coding that we expected. Through the experiments, we can verify that: 1) the logical process of decision-making made by ResNet models can be explained well based on the "decision information flow path" formed in the TSCL module; 2) the decision-path can reasonably interpret the decision reversal mechanism (Robustness mechanism) of the DCNN model; 3) the credibility of decision-making can be measured by the matching degree between the actual and expected decision-path.
The weighted neighborhood rough set (WNRS) is a novel neighborhood rough set model which doesn’t treat all attributes as equally important. It performs well in attribute reduction on real-valued data. However, there are still so many calculations in the attribute reduction based on WNRS that its efficiency is not high enough. To solve this drawback, we propose a fast attribute reduction algorithm based on WNRS (FWNRS) in this work. Firstly, we propose three fast strategies to reduce the neighborhood search range. The first one is reducing the neighborhood search rang of each testing sample from the whole universe or the boundary region to its corresponding neighborhood. The second one is removing the positive region samples from the corresponding neighborhood. The third one is to set the initial value of the cycle to be greater than the final value and the step size to be negative when the neighborhood radius δ is optimizing. Secondly, we design the fast attribute reduction algorithm based on the fast strategies. Finally, six UCI benchmark datasets are used to verify the efficiency of FWNRS. The experimental results illustrate that FWNRS is more efficient than the state-of-the-art comparison algorithms, which are all fast attribute reduction algorithms.
Because of the effectiveness of federated learning in protecting privacy and breaking the "data silo" phenomenon, it has been widely studied and applied in recent years. Firstly, the definition, classification and structure of federated learning are expounded. Secondly, the principle and development of federated algorithm are summarized. Then, it analyzes the challenges faced by federated learning, such as privacy protection, security aggregation, heterogeneity, and traffic cost, and summarizes the status quo of federated optimization algorithm from these aspects. Finally, it analyzes the development and feasibility of federated learning through the application of various industries, and summarizes the prospect.
In this paper, we propose a new general classification algorithm based on natural neighboring granular spheres, which has good noise immunity and self-adaptability. The old particle sphere generation algorithm is based on purity, and it can have good noise immunity and classification accuracy when the purity threshold is properly selected. However, when the quality of the data is unknown, the selection of the purity threshold will become a difficult problem. The old granular ball generaotr only considers the purity. When the purity reaches the purity threshold, even if the sample points inside the sphere should be divided into two balls, they will not continue to split, which may lead to a large difference between the actual distribution of the generated spheres and the sample points, and cause a negative impact on the next algorithm based on the sphere calculation. To address this situation, this paper proposes to continue the splitting of such balls by applying the natural neighbor-based clustering algorithm: LORE, which does not need parameter k, and thus can provide an effective basis for the granular ball generator.
The main outcome of this work is that we will extend the classical Fourier transform to the fuzzy Fourier transform, which is based on Henstock-Kurzweil integrals on infinite interval. The definition and existence condition of fuzzy Fourier transform are given, then based on the definition and existence theorem of fuzzy Fourier transform, we give an equivalent characterization theorem and its basic properties are also discussed.
