Transfer learning plays a crucial role in low-resource machine translation by addressing the challenge of poor model performance due to limited data in low-resource languages, thereby improving translation accuracy. Current research methods not only utilize pre-trained parent models for parameter initialization and fine-tuning but also use the soft labels output by these parent models to enhance the consistency between parent and child models. However, even if the parent model performs well, there are still instances where certain token predictions are unstable. During training, if the child model incorporates these unstable token predictions, it can hinder its learning effectiveness; the child model might not fully comprehend the parent model’s prediction strategy, potentially affecting overall translation performance. To address this, we propose a training strategy called Token-Level Filter Training, designed to effectively filter out unstable token predictions from the parent model, thereby transferring the parent model’s positive knowledge to the child model. Additionally, we introduce a hierarchical ranking loss method to help the child model better learn the parent model’s prediction strategies and sequence order, thus enhancing translation accuracy and fluency. Experimental results show that our method outperforms baseline methods on the public datasets Global Voices (Id, Ca, Hu, Pl) and WMT17 (Turkish–English), with BLEU score improvements of 1.47, 0.91, 0.50, 0.54, and 0.55, respectively. These results demonstrate the effectiveness and superiority of the proposed method.

In the Xinjiang region of China, water is scarce and unevenly distributed, and due to factors such as global warming, extreme rainfall events occur frequently, posing serious threats to people's lives and property safety. To accurately predict short-term precipitation, we propose a Mixformer model. Specifically, we first use a combination of min-max normalization and reversible instance normalization to reduce the impact of feature numerical ranges on modeling while preserving the distribution of the original data. Next, to enhance the representation of multivariate time series data, we propose a feature mixing module. This module enhances the representation of inter-feature correlation information by calculating the correlation between different sequences, thus improving prediction effectiveness. Finally, to enhance its non-linear characteristics, we introduce a residual prediction method. This method models seasonal and trend components separately and also pays special attention to the residual component. We have validated the proposed method extensively, proving that it outperforms existing state-of-the-art (SOTA) methods in this field.

Due to their inherent characteristics, small objects have weaker feature representation after multiple down-sampling and are even annihilated in the background. FPN’s simple feature concatenation does not fully utilize multi-scale information and introduces irrelevant context into the information transfer, further reducing the detection performance of the small object. To address the above issues, we propose the simple but effective FE-YOLOv5. (1) We designed the feature enhancement module (FEM) to capture more discriminative features of the small object. Global attention and high-level global contextual information are used to guide shallow, high-resolution features. Global attention interacts with cross-dimensional feature interaction and reduces information loss. High-level context complements more detailed semantic information by modeling global relationships through non-local networks. (2) We design the spatially aware module (SAM) to filter spatial information and enhance the robustness of features. Deformable convolution performs sparse sampling and adaptive spatial learning to better focus on foreground objects. According to the experimental results, our proposed FE-YOLOv5 outperforms the other architectures in the VisDrone2019 dataset and Tsinghua-Tencent100K dataset. Compared to YOLOv5, the APS was improved by 2.8% and 2.9%, respectively.

Most person re-identification methods are researched under various assumptions. However, viewpoint variations or occlusions are often encountered in practical scenarios. These are prone to intra-class variance. In this paper, we propose a multiscale global-aware channel attention (MGCA) model to solve this problem. It imitates the process of human visual perception, which tends to observe things from coarse to fine. The core of our approach is a multiscale structure containing two key elements: the global-aware channel attention (GCA) module for capturing the global structural information and the adaptive selection feature fusion (ASFF) module for highlighting discriminative features. Moreover, we introduce a bidirectional guided pairwise metric triplet (BPM) loss to reduce the effect of outliers. Extensive experiments on Market-1501, DukeMTMC-reID, and MSMT17, and achieve the state-of-the-art results on mAP. Especially, our approach exceeds the current best method by 2.0% on the most challenging MSMT17 dataset.

Motor imagery (MI) requires subjects to visualize the requested motor behaviors, which involves a large-scale network that spans multiple brain areas. The corresponding cortical activity reflected on the scalp is characterized by event-related desynchronization (ERD) and then by event-related synchronization (ERS). However, the network mechanisms that account for the dynamic information processing of MI during the ERD and ERS periods remain unknown. Here, we combined ERD/ERS analysis with the dynamic networks in different MI stages (i.e. motor preparation, ERD and ERS) to probe the dynamic processing of MI information. Our results show that specific dynamic network structures correspond to the ERD/ERS evolution patterns. Specifically, ERD mainly shows the contralateral networks, while ERS has the symmetric networks. Moreover, different dynamic network patterns are also revealed between the two types of MIs, in which the left-hand MIs exhibit a relatively less sustained contralateral network, which may be the network mechanism that accounts for the bilateral ERD/ERS observed for the left-hand MIs. Similar to the network topologies, the three MI stages also appear to be characterized by different network properties. The above findings all demonstrate that different MI stages that involve specific brain networks for dynamically processing the MI information.

In device-to-device (D2D) networks, co-channel interference is one of the main reasons that causes high power consumption, which further reduces the life time of mobile devices. In this paper, we adopt the superposition coding between macro and D2D users, which is expected to effectively eliminate the co-channel interference. In particular, we develop two power allocation methods to minimize the power consumption in cooperative and non-cooperative modes, respectively. Then, we use mode selection to obtain the minimum overall power consumption of the whole system. In power allocation, we model the average power consumption as a function of channel gain, power allocation factor, transmission rate, and noise power. Then, we obtain the close-form solution. Our results indicate that the proposed method outperforms the conventional non-cooperative methods in terms of power consumption, outage probability, and energy efficiency.