Recently, during natural disasters, the use of social media as a source of actionable information has increased significantly. Finding disaster-related tweets and analyzing their textual content and images can help government agencies and rescue organizations make better decisions. The main challenge of disaster-related message detection on social networking sites is how to idenify posts related to emerging disaster events. Since most existing methods extract disaster-specific features that cannot be shared between different disaster events, they are difficult to deal with this challenge. In this paper, we propose a novel Multi-modal Adversarial Neural Network (MANN) to handle the above challenge. MANN consists of three modules: a feature extraction module, a tweet detection module and a disaster discrimination module. The feature extraction module uses BERT and VGG-19 networks to learn textual and visual feature representations from posts. Based on multi-modal features, the tweet detection module is responsible for identifying posts related to disasters. MANN exploits the disaster discrimination module and adversarial training to captures disaster-invariant features for unseen disaster events. Experimental results on real-world datasets show the proposed model outperforms baseline methods in terms of precision, recall and F1-measure.

Artificial neural networks (ANNs) have recently made a significant impact on the field of industrial electronics. Increasing efforts have been focused on implementing ANN using more efficient and compact architectures to achieve improved computational performance and power efficiency. The activation function of the nonlinear layers is crucial for the successful implementation of ANN, as it enables the network to possess generalization ability. In this study, we implemented the hyperbolic tangent function (tanh) using an analog circuit based on phase-change memory (PCM), which is a mature nonvolatile technology. This fully analog PCM-based tanh function is naturally compatible with analog-memory-based ANN architectures, such as the memristive neural networks (MNNs). An analog circuit has been proposed to implement the tanh by transforming the PCM I-V characteristics. The results indicate that the implemented tanh based on PCM presents a low root-mean-square error to the ideal tanh. Furthermore, the performance of the implemented tanh was verified in a classical deep neural network (DNN) with the handwritten digit recognition task. The implemented tanh with realistic characteristics presents a training accuracy (>90%) and good precision for repeatability inference in DNN. This work provides effective means for implementing activation functions in ANN based on analog memory.

Integrating Unmanned Aerial Vehicles (UAVs) into military operations has become increasingly significant in recent years. In practical missions, UAVs are usually operated in unpredictable conditions, requiring robust planning to adapt to real-time changes. Thus, an effective path-planning algorithm is essential for navigating UAVs through complex terrains. However, the commonly used Rapidly-exploring Random Tree (RRT) algorithm for UAV path planning often generates suboptimal paths that require extensive post-processing to improve. While its variants can converge to the optimal solution, they suffer from more memory and computation consumption. On the other hand, purely learning-based methods could result in planning failures or excessively long paths generated for obstacle avoidance. In this work, we introduce a UAV path planning model that integrates diverse sensor data through a neural network architecture featuring CNN, LSTM, and an attention mechanism. Then, the deep learning-enhanced RRT algorithm combines the neural module with the informed sampling methods for adaptive and efficient pathfinding in dynamic environments. After carefully considering the trade-offs among path optimality, computational efficiency, and fewer turning angles, our approach exhibits superior overall performance compared to Informed RRT*, Batch Informed Trees (BIT*) algorithms, and the state-of-the-art Layered RQN.

Scientific information extraction is a crucial step for understanding scientific publications. In this paper, we focus on scientific keyphrase extraction, which aims to identify keyphrases from scientific articles and classify them into predefined categories. We present a neural network based approach for this task, which employs the bidirectional long short-memory (LSTM) to represent the sentences in the article. On top of the bidirectional LSTM layer in our neural model, conditional random field (CRF) is used to predict the label sequence for the whole sentence. Considering the expensive annotated data for supervised learning methods, we introduce self-training method into our neural model to leverage the unlabeled articles. Experimental results on the ScienceIE corpus and ACL keyphrase corpus show that our neural model achieves promising performance without any hand-designed features and external knowledge resources. Furthermore, it efficiently incorporates the unlabeled data and achieve competitive performance compared with previous state-of-the-art systems.

Biomedical named entity recognition (BioNER) is an essential task in biomedical information analysis. Recently, deep neural approaches have become widely utilized for BioNER. Biomedical dictionaries, implemented through a masked manner, are frequently employed in these methods to enhance entity recognition. However, their performance remains limited. In this work, we propose a dictionary-based matching graph network for BioNER. This approach utilizes the matching graph method to project all possible dictionary-based entity combinations in the text onto a directional graph. The network is implemented coherently with a bi-directional graph convolutional network (BiGCN) that incorporates the matching graph information. Our proposed approach fully leverages the dictionary-based matching graph instead of a simple masked manner. We have conducted numerous experiments on five typical Bio-NER datasets. The proposed model shows significant improvements in F1 score compared to the state-of-the-art (SOTA) models: 2.8% on BC2GM, 1.3% on BC4CHEMD, 1.1% on BC5CDR, 1.6% on NCBI-disease, and 0.5% on JNLPBA. The results show that our model, which is superior to other models, can effectively recognize natural biomedical named entities.