The existing joint embedding Visual Question Answering models use different combinations of image characterization, text characterization and feature fusion method, but all the existing models use static word vectors for text characterization. However, in the real language environment, the same word may represent different meanings in different contexts, and may also be used as different grammatical components. These differences cannot be effectively expressed by static word vectors, so there may be semantic and grammatical deviations. In order to solve this problem, our article constructs a joint embedding model based on dynamic word vector—none KB-Specific network (N-KBSN) model which is different from commonly used Visual Question Answering models based on static word vectors. The N-KBSN model consists of three main parts: question text and image feature extraction module, self attention and guided attention module, feature fusion and classifier module. Among them, the key parts of N-KBSN model are: image characterization based on Faster R-CNN, text characterization based on ELMo and feature enhancement based on multi-head attention mechanism. The experimental results show that the N-KBSN constructed in our experiment is better than the other 2017—winner (glove) model and 2019—winner (glove) model. The introduction of dynamic word vector improves the accuracy of the overall results.

In neural machine translation (NMT), the sophistication of word embeddings plays a pivotal role in the model’s ability to render accurate and contextually relevant translations. However, conventional models with single granularity of word segmentation cannot fully embed complex languages like Chinese, where the granularity of segmentation significantly impacts understanding and translation fidelity. Addressing these challenges, our study introduces the Dynamic Multi-Granularity Translation System (DMGTS), an innovative approach that enhances the Transformer model by incorporating multi-granularity position encoding and multi-granularity self-attention mechanisms. Leveraging a Directed Acyclic Graph (DAG), the DMGTS utilizes four levels of word segmentation for multi-granularity position encoding. Dynamic word embeddings are also introduced to enhance the lexical representation by incorporating multi-granularity features. Multi-granularity self-attention mechanisms are applied to replace the conventional self-attention layers. We evaluate the DMGTS on multiple datasets, where our system demonstrates marked improvements. Notably, it achieves significant enhancements in translation quality, evidenced by increases of 1.16 and 1.55 in Bilingual Evaluation Understudy (BLEU) scores over traditional static embedding methods. These results underscore the efficacy of the DMGTS in refining NMT performance.

Object detection algorithms for optical remote sensing images often face challenges in computational efficiency, particularly when detecting small and densely packed targets. This paper introduces CGBi_YOLO, a novel lightweight land target detection network designed to optimize computational resource utilization while maintaining detection capabilities for small-scale targets. Our approach incorporates an innovative lightweight optimization strategy featuring a new lightweight backbone feature extraction network: CSPGhostNet. This model significantly enhances the detection ability of small objects within optical remote sensing images without increasing computational demands. The efficacy of the proposed model is validated through rigorous experimentation on the DOTA dataset. Compared to the baseline model, CGBi_YOLO achieves a 30% reduction in parameters and a 36% increase in inference speed. The model demonstrates exceptional performance in handling small and densely packed targets within optical remote sensing images, showcasing its potential for real-world applications in fields such as environmental monitoring, urban planning, and disaster management.

Deep learning techniques can make use of a large amount of time-series data from wearable devices and greatly benefit the development of sensor-based human activity recognition (HAR). However, representation learning in a supervised manner requires massive labeled sensory data that are time-consuming to obtain and hindered by privacy concerns. To address these issues, we utilize the plentiful unlabeled sensory data and propose a novel self-supervised learning framework, namely Temporal Contrastive Learning in Human Activity Recognition (TCLHAR), which learns the meaningful feature representations for time-series data without labels. Our TCLHAR framework utilizes the temporal co-occurrence relationship among time windows as the supervisory signals to construct positive pairs in the encoder pretraining stage. The encoder is designed for cross-modality fusion, which leverages the local interactions of each sensor modality and the global fusion of features from different sensors. The proposed framework is extensively evaluated on public HAR datasets in supervised, self-supervised, and semi-supervised settings. Our method outperforms several self-supervised learning benchmark models, achieving comparable results with fully labeled data training. When labeled data is scarce, our method can boost the F1 score by up to 65% over traditional supervised training, which demonstrates the effectiveness of our feature representations.