International Journal on Artificial Intelligence Tools

EEG is a personal privacy information, that is difficult to share data and not conducive to the study of sleep staging. In this paper, the Federated Learning (FL) method is introduced to solve the problem of data island in sleep staging EEG signals. To simulate the distribution of different user data, we introduce the K-means algorithm into EEG dataset partitioning and transform it into a non-independent and identically distributed unbalanced dataset. Then, based on the classic Multilayer Perceptron (MLP) network, we analyse the privacy leakage problem in FL, and Differential Privacy (DP) is used to protect the privacy of user EEG data. Next, we analyse the reason for the accuracy loss between FL and Centralized Learning (CL), and the effect of the imbalance of the data distribution of each client. We propose a self-adaptive batch method to adaptively balance the batch size of each client, and experiments show that the training results of imbalanced datasets can be improved. Finally, when the accuracy of the model does not reach the historical best value in the set epoch interval, the Mandatory Optimization Strategy (MOS) is used to further improve the accuracy and reduce the randomness of the model. The DP-FL based on the self-adaptive batch and MOS strategy is tested on our dataset and the Sleep-EDF Database Expanded dataset, which not only protects user privacy but also achieves the same accuracy as CL. The accuracy is 82.88% on our dataset and 91.81% on the subset of the Sleep-EDF Database Expanded dataset.

Accurate short-term price forecasting is crucial for the power system and electricity market. This paper proposes a hybrid short-term electricity price forecasting model based on empirical mode decomposition (EMD) and deep neural network (DNN). Firstly, EMD is used to denoise the data set. Next, the reconstructed data are input into the DNN composed of a convolutional neural network (CNN) and long-short-term memory (LSTM) neural network to analyze the characteristics and output the prediction results. Finally, the superiority of this model is verified by comparing the electricity price data of the Australian electricity market with a single LSTM network, EMD-CNN model, and CNN-LSTM model.

Distributed optimization algorithms are widely used for solving science and engineering problems in recent years. However, these algorithms become vulnerable in the face of external attacks, so it is urgent to develop novel resilient distributed optimization algorithms against attacks. In this paper, we design a novel median-based resilient distributed optimization algorithm to defend the Byzantine attack. The existence of transition matrix is proved, and the convergence and optimality of the proposed algorithm are analyzed. Simulation experiments are performed to verify the effectiveness of the proposed algorithm and the results show that it is more effective than the compared one and can get the best possible results against the attack.

Learning-based defect inspection systems are becoming more and more influential in the industry. However, these approaches are usually restricted by the insufficiency of training data, especially the defect samples. It is natural to generate “artificial defect samples”. In this paper, a novel approach to generate high-quality industrial defect images from extremely few samples is proposed. We design a generative adversarial network (GAN) for this task. Mode collapse is a common problem of GANs when training data is insufficient, thus, fully convolutional network and Markovian discriminator are applied to eliminate this problem. In order to generate images of high resolution, we also propose a novel sampling strategy to balance memory consumption and training quality. Experiments illustrate that our model is capable of generating high-resolution industrial defect images from extremely few samples of various categories. With the generated “artificial samples”, the classification and detection models outperform superior accuracy compared to the same models with the original rare samples. Thus, this approach can increase the intelligence of defect inspection systems.

It is of great need to accurately locate vessels and oil platforms because huge numbers of these marine man-made targets can cause oil spills or illegal invasion problems. Many different types of marine targets in the complex marine environment make it very difficult to deal with the target detection and classification process accurately. This paper proposes a robust Laplacian of Gaussian operator connected domain controller squeeze excitation residual network (LCSE-ResNet) for marine man-made target classification based on optical remote sensing imagery. Vessel and oil platform candidate regions in remote sensing images are extracted by the Laplacian of Gaussian (LoG) operator and the connected domain controller, which are input into the classification neural networks in the following. It is a whole End-to-End structure from original remote sensing images to the precise target information. That clouds, ripples, and the other man-made target disturbances can be excluded effectively means the better robust character in the actual application. Some shape and size features of vessels and oil platforms are considered in the LCSE-ResNet structure, which improves the explanation of the final results. Experimental results demonstrate that the proposed LCSE-ResNet can effectively detect marine man-made targets and effectively distinguish vessels and oil platforms.

As a prominent semi-supervised learning algorithm, co-training can make full use of a few labeled samples as well as a quantity of unlabeled samples to train robust classifiers, thus it is widely researched in the last decades. However, there are still two severe problems in existing co-training methods: firstly, data used in co-training needs at least two sufficient and redundant views, but few practical datasets can meet it. Secondly, because there are a few labeled samples in the initial stage, the initial classifiers trained by co-training are usually too weak to correctly label unlabeled samples, which may bring in noisy labels for the following training process. To deal with these problems, we propose a co-training algorithm by predicting confidence of unlabeled neighbors (termed LCN-CoTrain). Specifically, LCN-CoTrain first defines an entropy-based view division method to generate redundant views. Meanwhile, LCN-CoTrain introduces a novel labeling confidence prediction strategy, which takes the nearest unlabeled neighbors into consideration when calculating the labeling confidence of a certain sample. To validate the efficiency of our proposed LCN-CoTrain algorithm, we experiment on four UCI datasets and compare LCN-CoTrain with several representative co-training methods. The experimental results indicate that our proposed LCN-CoTrain can learn robust classifiers and outperform most of baseline methods.