Search

Abstract This paper addresses the recognitions of human actions in videos. Human action recognition can be seen as the automatic labeling of a video according to the actions occurring in it. It has become one of the most challenging and attractive problems in the pattern recognition and video classification fields. The problem itself is difficult to solve by traditional video processing methods because of several challenges such as the background noise, sizes of subjects in different videos, and the speed of actions. Derived from the progress of deep learning methods, several directions are developed to recognize a human action from a video, such as the long-short-term memory (LSTM)-based model, two-stream convolutional neural network (CNN) model, and the convolutional 3D model.In this paper, we focus on the two-stream structure. The traditional two-stream CNN network solves the problem that CNNs do not have satisfactory performance on temporal features. By training a temporal stream, which uses the optical flow as the input, a CNN can have the ability to extract temporal features. However, the optical flow only contains limited temporal information because it only records the movements of pixels on the x-axis and the y-axis. Therefore, we attempt to design and implement a new two-stream model by using an LSTM-based model in its spatial stream to extract both spatial and temporal features in RGB frames. In addition, we implement a DenseNet in the temporal stream to improve the recognition accuracy. This is in-contrast to traditional approaches which typically utilize the spatial stream for extracting only spatial features. The quantitative evaluation and experiments are conducted on the UCF-101 dataset, which is a well-developed public video dataset. For the temporal stream, we choose the optical flow of UCF-101. Images in the optical flow are provided by the Graz University of Technology. The experimental result shows that the proposed method outperforms the traditional two-stream CNN method with an accuracy of at least 3%. For both spatial and temporal streams, the proposed model also achieves higher recognition accuracies. In addition, compared with the state of the art methods, the new model can still have the best recognition performance.

Artificial intelligence (AI) systems are becoming wiser, even surpassing human performances in some fields, such as image classification, chess, and Go. However, most high-performance AI systems, such as deep learning models, are black boxes (i.e., only system inputs and outputs are visible, but the internal mechanisms are unknown) and, thus, are notably challenging to understand. Thereby a system with better explainability is needed to help humans understand AI. This paper proposes a dual-track AI approach that uses reinforcement learning to supplement fine-grained deep learning-based sentiment classification. Through lifelong machine learning, the dual-track approach can gradually become wiser and realize high performance (while keeping outstanding explainability). The extensive experimental results show that the proposed dual-track approach can provide reasonable fine-grained sentiment analyses to product reviews and remarkably achieve a 133% promotion of the Macro-F1 score on the Twitter sentiment classification task and a 27.12% promotion of the Macro-F1 score on an Amazon iPhone 11 sentiment classification task, respectively.

Benefiting from the rapid development of big data and high-performance computing, more data is available and more tasks could be solved by machine learning now. Even so, it is still difficult to maximum the power of big data due to each dataset is isolated with others. Although open source datasets are available, algorithms’ performance is asymmetric with the data volume. Hence, the AI community wishes to raise a symmetric continuous learning architecture which can automatically learn and adapt to different tasks. Such a learning architecture also is commonly called as lifelong machine learning (LML). This learning paradigm could manage the learning process and accumulate meta-knowledge by itself during learning different tasks. The meta-knowledge is shared among all tasks symmetrically to help them to improve performance. With the growth of meta-knowledge, the performance of each task is expected to be better and better. In order to demonstrate the application of lifelong machine learning, this paper proposed a novel and symmetric lifelong learning approach for sentiment classification as an example to show how it adapts different domains and keeps efficiency meanwhile.

Machine learning is a very important approach to pattern classification. This paper provides a better insight into Incremental Attribute Learning (IAL) with further analysis as to why it can exhibit better performance than conventional batch training. IAL is a novel supervised machine learning strategy, which gradually trains features in one or more chunks. Previous research showed that IAL can obtain lower classification error rates than a conventional batch training approach. Yet the reason for that is still not very clear. In this study, the feasibility of IAL is verified by mathematical approaches. Moreover, experimental results derived by IAL neural networks on benchmarks also confirm the mathematical validation.

The importance of an interference-less machine learning scheme in time series prediction is crucial, as an oversight can have a negative cumulative effect, especially when predicting many steps ahead of the currently available data. The on-going research on noise elimination in time series forecasting has led to a successful approach of decomposing the data sequence into component trends to identify noise-inducing information. The empirical mode decomposition method separates the time series/signal into a set of intrinsic mode functions ranging from high to low frequencies, which can be summed up to reconstruct the original data. The usual assumption that random noises are only contained in the high-frequency component has been shown not to be the case, as observed in our previous findings. The results from that experiment reveal that noise can be present in a low frequency component, and this motivates the newly-proposed algorithm. Additionally, to prevent the erosion of periodic trends and patterns within the series, we perform the learning of local and global trends separately in a hierarchical manner which succeeds in detecting and eliminating short/long term noise. The algorithm is tested on four datasets from financial market data and physical science data. The simulation results are compared with the conventional and state-of-the-art approaches for time series machine learning, such as the non-linear autoregressive neural network and the long short-term memory recurrent neural network, respectively. Statistically significant performance gains are recorded when the meta-learning algorithm for noise reduction is used in combination with these artificial neural networks. For time series data which cannot be decomposed into meaningful trends, applying the moving average method to create meta-information for guiding the learning process is still better than the traditional approach. Therefore, this new approach is applicable to the forecasting of time series with a low signal to noise ratio, with a potential to scale adequately in a multi-cluster system due to the parallelized nature of the algorithm.

Incremental Attribute Learning (IAL) is a feasible approach for solving high-dimensional pattern recognition problems. It gradually trains features one by one. Previous research indicated that supervised machine learning with input attribute ordering can improve classification results. Moreover, input space partitioning can also effectively reduce the interference among features. This study proposed IAL based on Grouped Feature Ordering, which fused feature partitioning with feature ordering. The experimental results show that this approach is not only applicable for pattern classification improvement, but also efficient to reduce interference among features.

Since conventional methods are incapable of estimating the parameters of Photovoltaic (PV) models with high accuracy, bioinspired algorithms have attracted significant attention in the last decade. Cuckoo Search (CS) is invented based on the inspiration of brood parasitic behavior of some cuckoo species in combination with the Lévy flight behavior. In this paper, a CS-based parameter estimation method is proposed to extract the parameters of single-diode models for commercial PV generators. Simulation results and experimental data show that the CS algorithm is capable of obtaining all the parameters with extremely high accuracy, depicted by a low Root-Mean-Squared-Error (RMSE) value. The proposed method outperforms other algorithms applied in this study.

Precise photovoltaic (PV) behavior models are normally described by nonlinear analytical equations. To solve such equations, it is necessary to use iterative procedures. Aiming to make the computation easier, this paper proposes an approximate single-diode PV model that enables high-speed predictions for the electrical characteristics of commercial PV modules. Based on the experimental data, statistical analysis is conducted to validate the approximate model. Simulation results show that the calculated current-voltage (I-V) characteristics fit the measured data with high accuracy. Furthermore, compared with the existing modeling methods, the proposed model reduces the simulation time by approximately 30% in this work.