查看更多>>摘要：With the advancement of accelerated hardware in recent years, there has been a surge in the development and application of intelligent systems. Deep learning systems, in particular, have shown exciting results in a wide range of tasks: classification, detection, and recognition. Despite these remarkable achievements, there remains an active area critical for the safety of those systems. Deep learning algorithms have proven to be brittle against adversarial attacks. That is, carefully crafted adversarial inputs can consistently trigger an erroneous classification output from a network model. Hence, the motivation of this paper, we survey four different attacks, two adversarial defense methods on three benchmark datasets to gain a better understanding of how to protect those systems. We motivate our findings by achieving state-of-the-art accuracy and collecting empirical evidence of attack effectiveness against deep neural networks. Additionally, we leverage network explainability methods to investigate an alternative approach to defend deep neural networks.
查看更多>>摘要：Surface cracks need to be detected regularly to ensure the safety of concrete buildings. For the sake of efficiency and accuracy, concrete surface cracks are detected by machine vision technology. This paper briefly introduced the convolutional neural network (CNN) algorithm used for identifying concrete surface cracks. Then, the traditional CNN algorithm was improved by the particle swarm optimization (PSO) algorithm, and it was compared with the support vector machine (SVM) algorithm and the traditional CNN algorithm in the simulation experiment. The results showed that the improved CNN algorithm effectively identified the concrete surface cracks with different cracking degrees; moreover, the precision ratio, recall ratio, and F value of the improved CNN algorithm were superior to those of SVM and traditional CNN algorithms in recognizing cracks on the concrete surface, and the training and testing time was shorter than that of SVM and CNN algorithms.
查看更多>>摘要：Facial Expression Recognition (FER) is a task usually framed as predicting an emotional state given a facial image. FER has received numerous attentions from both business and academia as it could serve as a crucial component in various applications, e.g., automatic evaluation of customer satisfaction, sign language recognition, and human-computer interaction. Despite active research on the subject, facial expression recognition remains greatly challenging due to the diversity of individual expressions, shapes, and sizes of face, eyes, mouth, and other facial features, as well as orientation, alignment, and lighting. In this paper, we aim to improve the fast and robust online-learning FER with unseen data identification. We compare two widely used feature extraction methods for FER, namely Curvelet Transform (CT) and Local Curvelet Transform (LCT). Furthermore, we explore factors underlying several online extreme learning approaches for unseen data identification. Our experimental results demonstrate that (1) CT is suitable for a cleaner and well-prepared dataset, while LCT seems to work well on a dataset with diverse quality and on level of consistency. (2) The Identity Structural Tolerance Sequential Circular Extreme Learning Machine outperforms other Extreme Learning algorithms employed in FER. (3) LC can provide unseen identification capability to Extreme Learning algorithms. These findings emphasize the common underlying foundation between the extreme learning approach and other traditional learning approaches.
查看更多>>摘要：One of the main challenges in sentiment analysis is the polarity shift. Studies have shown that the detection of polarity shifts is very effective to improve the accuracy of sentiment analysis. However, the problem of polarity shift detection has not been well studied, and most studies have only focused on detecting negations, one kind of polarity shifts. This paper aims to provide a semantic method based on domain knowledge for the detection of polarity shifts. In the proposed method, a polarity shift-tagged corpus is created using the idea of distant supervision. Thereafter, the polarity shifts are detected by training the machine learning classifiers on the resulting corpus, based on the semantic features extracted from the domain knowledge. The experimental results reveal that the SVM classifier with training on the constructed corpus is capable of detecting the polarity shifts with 79.33% accuracy and 81.21% F-measure, which are 24.6% and 17.5% more accurate than the best-performing existing method, respectively. Also, the results show that with the use of the polarity shift tag as a feature, SVM classifier F-measure for sentiment analysis has been improved up to 1.2%.
查看更多>>摘要：The pH level of oceans has been largely monitored and studied to make sure that aquatic ecosystems are thriving. However, the pH level of other large bodies of waters, such as rivers, has largely been glanced over. Many rivers contain very sensitive underwater ecosystems, and as a result even small pH changes can largely impact the relative biodiversity. With the addition of increased carbon emissions and pollution, large bodies of water are absorbing more carbon and consequently the pH levels of rivers can rapidly change. This paper studies different deep learning approaches to analyze and forecast pH levels, which include a long short-term memory (LSTM), gated recurrent unit (GRU), recurrent neural network (RNN), and a Temporal Fusion Transformer (TFT) model to determine which algorithm provides the best pH predictive forecast. We demonstrate that the TFT outperforms other deep learning methods through various metrics. In addition, we clarify the importance of temperature as a feature in pH prediction. Lastly, we use the TFT to predict pH anomalies and discover the significance of the predicted data. We found that nine out of the ten predicted data sets have a significant difference compared to the original data.
NSTL
Springer Nature