With the rapid evolution of GPU architectures, the heterogeneity of model training infrastructures is steadily increasing. In such environments, effectively utilizing all available heterogeneous accelerators becomes critical for distributed model training. However, existing frameworks, which are primarily designed for homogeneous clusters, often exhibit significant resource underutilization when deployed on heterogeneous accelerators and networks. In this paper, we present Harp, an automated parallel training framework designed specifically for heterogeneous clusters. Harp introduces a fine-grained planner that efficiently searches a wide space for the inter-operator parallel strategy, enabling Harp to alleviate communication overheads while maintaining balanced loads across heterogeneous accelerators. In addition, Harp implements a heterogeneity-aware 1F1B scheduler that adaptively adjusts the execution timing and ordering of microbatches based on network characteristics, maximizing computation-communication overlap under cross-cluster interconnects while incurring only minimal memory overhead. Our evaluation results show that Harp can deliver 1.3×–1.6× higher performance on heterogeneous clusters than state-of-the-art training frameworks.

To enhance the performance of pre-trained foundation models in various natural language processing tasks, the scale of these models has been steadily increasing in recent years, driven by advancements high-performance computing hardware. However, the presence of diverse computing devices in data centers and the varying computational demands of different tasks make it challenging for a single type of high-performance G PU to suffice for the large-scale training of foundation models' and lower-performance GPUs often remain underutilized. Therefore, optimizing resourceallocationand improving training efficiency have become a critical challenge. To address these issues, this study proposes a hybrid parallel training framework for heterogeneous computing devices, named HyperPipe. This framework automatically selects the optimal hybrid parallel strategy based on the performance characteristics of the devices, thereby achieving efficient resource utilization and acceleration training to meet the computational demands of large-scale FMs. Experimental results demonstrate that in a heterogeneous computing environment, training large models like GPT-3-1.3B with HyperPipe achieves significant acceleration, delivering speed-ups ranging from 1.6 x to 2.8 x compared to previous methods.

Sonar imaging, due to its ability to penetrate water and certain obstacles, has become a pivotal technology in marine domains. The complexities of underwater environments, combined with the low resolution and noise interference inherent in sonar images, present significant challenges in sonar image segmentation. To address these issues, we propose a novel approach, GM-SAM, which integrates denoising and multi-scale boundary feature extraction into the Segment Anything Model (SAM). Initially, the Multi-Perspective Feature Module (MPFM) employs parallel attention mechanisms and channel-enhanced convolutions to reduce computational complexity, suppress noise, and capture both local details and global feature. Following this, the Gradient-Enhanced Transformation Module (GETM) leverages gradient-based operations to extract edge features, enhancing the model’s sensitivity to boundaries. Finally, the Sonar-Fusion Adapter module integrates task-specific knowledge from MPFM with general features from SAM. Experimental results demonstrate that GM-SAM significantly outperforms existing methods, achieving superior Dice scores and effectively segmenting valid targets.

The unique nature of map data presents challenges for detecting key error areas in problematic maps, especially in terms of discontinuous boundary feature extraction and neglect of small target information. To address these issues, we propose a lightweight problematic map detection algorithm called YOLO-Map. First, to ensure the integrity of edge extraction and resistance to interference, we designed the Dual-branch Attention Convolution Module (DACM), which utilizes the synergistic effect of two branches to accurately identify national boundary regions. Next, the Multi-Path Feature Aggregation (MPFA) module adopts a bidirectional adaptive fusion strategy, enhancing recursive connections of multi-scale features and improving target localization accuracy. Additionally, we propose the Global Context Fusion Module (GCFM), which strengthens small target feature representation through a multi-branch collaborative attention mechanism. Experimental results show that YOLO-Map achieves an accuracy of 87.3% ([email protected]) on the CME dataset, outperforming many larger models.

With the development of Internet of Things (IoT) technology, user identity recognition technology has become an important line of defense to ensure application access and privacy security. Ultra-Wideband (UWB) bio-radar, with its high range resolution and non-contact penetration detection advantages, can accurately capture the individual-specific and dynamic continuous surface micro-motion characteristics caused by cardiac structural movement through transmitted microwaves, namely Microwave HeartPrint, provides a new technological path for non-contact identity recognition. However, current radar-based non-contact cardiac motion recognition mostly treats the micro-movement of the chest wall as a single scattering point target and primarily relies on single-domain data, thereby limiting the effectiveness of the recognition mechanism. To address the aforementioned issue, this paper first employs UWB bio-radar to continuously monitor the multi-point micro-motions in the range dimension of the cardiac body surface area along the radial direction, and then achieves the selection of effective range unit echoes from the perspective of multiple range tomographic planes. Secondly, based on the echo heartbeat segmentation and combined with short-time fourier transform (STFT) time-frequency transform, the Microwave Heartprint dual-domain data corresponding to the effective range units are obtained. Finally, a multi-view dual-domain feature fusion recognition network is proposed, where the multi-view single-domain feature fusion module (MV-SFFM) is designed to effectively fuse the differential and complementary micro-motion features of effective range units on the cardiac surface. In addition, the cross-attention dual-domain feature fusion module (CA-DFFM) effectively fuses the time-domain and time-frequency domain features to further improve the feature representation capability. Actual experiments show that the recognition accuracy of this method can reach more than 95%. Ablation experiments and visualization analysis strongly demonstrates the rationality and effectiveness of comprehensively utilizing the micro-motion information of multiple effective range units of the heart and dual-domain features.

The human brain connectome has long been recognized as a crucial component for various cognitive functions. While connectome-based predictive modeling (CPM) has been extensively explored for predicting behavior outcomes at the individual-level, its application to electroencephalogram (EEG) remains limited due to the inherent diversity and complexity of EEG frequency information. In the present work, we aim to address this issue by developing a filter bank CPM (FBCPM) framework that leverages narrowband EEG functional connectivity (FC) for individual prediction. Four independent datasets comprising 280 healthy subjects with 392 EEG recordings during the psychomotor vigilance test (PVT), were adopted here. Using the discovery dataset (i.e., Dataset 1) with 137 recordings, the feasibility of FBCPM was evaluated via predicting mean reaction time (RT) measures within a 15-min PVT task. The results showed that FBCPM framework achieved notable prediction accuracy and outperformed four benchmark approaches. Subsequent comprehensive internal and external validation analyses further affirmed its robustness across various hyper-parameters and generalizability to another three independent datasets (i.e., Dataset 2 to Dataset 4) with divergent recording or preprocessing settings. Moreover, the FBCPM framework exhibited satisfactory performance when generalized to time-on-task (TOT) effect measures (i.e., $\mathit {\Delta RT}$ and $\mathit {TOT_{slope}}$). Further investigation of contributing features to mean RT prediction indicated the remarkable predictive ability of negative features, manifesting as a pattern of low-frequency (below 8 Hz) predominance and complex topological distributions. Overall, these findings indicated that FBCPM provided a significant methodological advance in EEG-based individual prediction approaches, moving a step forward towards practical application in cognitive neuroscience.

Underwater Image Restoration Technology (UIRT) constitutes a pivotal base for subsequent tasks yet confronts obstacles like data scarcity and image distortion in practical deployments. To optimize the use of scarce annotated and copious unlabeled data, we introduce an avant-garde semi-supervised method for underwater image restoration, named STUI-Net. Concurrently, we develop a novel teacher-student architecture utilizing Transformer technology, named GFT-Net. Initially, GFT-Net employs a tripartite branch network—comprising Gradient, Feature Extraction, and Transformer modules to thoroughly extract features from underwater images. This method then amalgamates multi-source features, encompassing edge, gradient, local, and global data, addressing underwater image deterioration in multifaceted environments. Additionally, we engineer a Supervisor role to rectify potential misdirection by the teacher network in the semi-supervised model, thereby enhancing training stability on unlabeled datasets. Empirical studies affirm the robust generalization competence of STUI-Net in authentic underwater environments.

The fidelity of underwater visual data is significantly affected by various factors, encompassing light refraction and color distortion in water. These factors often give rise to noise and distortion within underwater images. In response to the widespread nature of this issue, we propose a novel approach to restore authentic underwater images through a comprehensive perspective. To address the inherent scarcity of real underwater image datasets, we introduce an innovative underwater image generation model that leverages Generative Adversarial Networks (GAN) with integrated physical constraints, aiming to produce realistic and stable underwater image datasets. Within the underwater image restoration framework, we incorporate a multi-channel feature extractor (MCFE) module, which is designed to enhance the model’s capability in image feature extraction. Additionally, we introduce object edge information as a novel loss function to perform different repairs on the target and background. Experimental evaluations demonstrate that our method exhibits superior performance in both qualitative and quantitative assessments compared to state-of-the-art approaches. Restoration results on real underwater images further showcase its exceptional performance in practical applications. The source code and sample dataset are publicly available at here.

In recent years, with the continuous development of artificial intelligence, deep learning algorithms are becoming more and more complex, and the scale of model training is also growing. The artificial intelligence platform also involves large-scale model training in our computing network operating system project. However, with the increasing size of data sets and models, the traditional single-card training makes the training speed very slow, and the training accuracy needs to converge, which has yet to meet people's computational needs. This has led to the development of GPipe, PipeDream, and other famous pipelines. In this paper, an efficient pipeline parallel training optimization method is proposed. In our approach, multiple computing nodes process small batches of data in parallel in a pipeline manner. We have mainly done the following two aspects of work: First, we designed a weight buffer strategy to limit the number of weight versions generated and ensure the model's accuracy. And we also developed a tensor compression mechanism to improve the transmission rate. Secondly, we propose a prefix sum partition algorithm to ensure that the pipeline can achieve balanced partitioning and save the memory of computing resources. Compared with several popular pipeline parallel frameworks, the proposed method can achieve about twice the training acceleration and save about 30% - 40% of the memory usage.

In recent years, accurate and real-time long-range small target detection has become a popular and challenging task, particularly in time-sensitive scenarios such as unmanned aerial vehicle (UAV) scene analysis and military reconnaissance. Most existing solutions rely on deep CNNs to learn strong feature representations of objects isolated from the background to detect small objects with minimal visual features in images. However, this approach incurs significant computational overhead. In this paper, we propose DualYOLO, a fast and accurate long-range small object detection method that combines multi-level multi-scale feature fusion (MLMFF) and concat channel attention (CatCA). Specifically, in order to prevent small targets from becoming more and more blurred after multilayer convolution operations, DualYOLO fuses the features of different layers in the backbone network to obtain small target features with strong semantics and high detail. Furthermore, we use a new loss function to address the sensitivity of IoU to small object position deviations, thereby improving detection accuracy. In terms of data preprocessing, we utilize an image slicing strategy to process the dataset. The experimental results show that DualYOLO achieves 82.2% accuracy (in terms of [email protected]) on the VEDAI dataset processed using slices, with a performance more than 2% higher than that of large models (e.g., YOLOv5x,YOLOR and YOLOv7).