Time-Series (TS) exhibits pronounced non-stationarity. Consequently, most forecasting methods display compromised robustness to concept drift, despite the prevalent application of instance normalization. We tackle this challenge by first analysing concept drift through a bias-variance lens and proving that weighted ensemble reduces variance without increasing bias. These insights motivate DeepBooTS, a novel end-to-end dual-stream residual-decreasing boosting method that progressively reconstructs the intrinsic signal. In our design, each block of a deep model becomes an ensemble of learners with an auxiliary output branch forming a highway to the final prediction. The block‑wise outputs correct the residuals of previous blocks, leading to a learning‑driven decomposition of both inputs and targets. This method enhances versatility and interpretability while substantially improving robustness to concept drift. Extensive experiments, including those on large-scale datasets, show that the proposed method outperforms existing methods by a large margin, yielding an average performance improvement of 15.8% across various datasets, establishing a new benchmark for TS forecasting.

Large vision and language models deployed at edge encounter continuously evolving input distributions, including not only new tasks but also highly unbalanced long-tail classes. For example, smart-surveillance cameras frequently capture common objects such as pedestrians and cars, while only occasionally observing tail classes like horseback riders or stroller pushers. However, most existing long-tail mitigation techniques are designed for fixed pretraining data. This often leads to poor accuracy of tail classes on evolving data and incurs high computational costs on edge devices. In this article, we propose EdgeTail, a lightweight long-tail mitigation method for edge-side continual learning. EdgeTail’s key design features are: (i) optimal long-tailed mitigation solution search, which adaptively selects the best long-tail learning method for the current distribution/task; and (ii) graph attention classifier and multi-branch adapter, which improves the quality and stability of tail class representations with small overheads. We implement EdgeTail in PyTorch and extensively evaluate it against state-of-the-art methods. The results show that EdgeTail improves the average accuracy by 36.09% under fixed training windows and by 31.12% under different training window sizes.

Artificial intelligence applications are developing rapidly, placing higher demands on the efficient use of computational power. Workflow offloading in wide-area environments can effectively utilize resources to improve task execution efficiency. However, cross-domain computational power scheduling and peak business periods often lead to network congestion and computational node overload, resulting in imbalanced utilization of wide-area computational resources. This makes it difficult to ensure the overall stability of the computational network and the fairness of resource utilization. This paper proposes the workflow scheduling method of computing power network based on load balance of voting mechanism. First, a multi-objective optimization model for workflow task scheduling is constructed. Then, a Voting-mechanism-based Multi-objective Optimization Algorithm is proposed, which combines the Evolutionary Multiobjective Optimization Algorithm based on Unequal Grid Division (IGEA) and the Multi-objective Artificial Rabbit Optimization Algorithm (MOARO). This method adaptively adjusts the weight of each algorithm during the optimization process to optimize the completion time of computationally intensive and data-intensive workflows, as well as the computational network resource utilization. Finally, experiments are conducted on different workflow combination scenarios. The proposed algorithm is compared with the MOARO and IGEA multiobjective optimization algorithms to validate the effectiveness of the method presented in this paper.

This paper proposes a switching and scheduling method for containerized computing environments. The core strength of the method lies in establishing a complete technical framework: standardized image templates and persistent storage ensure data consistency and availability during heterogeneous environment transitions, while the proposed Priority-based Resource Scheduling (PRS) serves as the key to efficient switching and resource allocation. Experimental results demonstrate that PRS achieves significant performance improvements over traditional strategies such as FCFS, SJF, and Random. Specifically, PRS reduces overall task completion time by approximately$5.2\%$, maintains cluster GPU time utilization consistently above$98\%$, and demonstrates superior stability compared to alternative strategies. By leveraging dynamic prediction and priority scheduling to prioritize critical tasks, the proposed method significantly improves overall task completion time and resource utilization in heterogeneous computing environments.

This work proposes a resource-aware adaptive framework for MobileNetV4 that dynamically balances accuracy, latency, and memory under edge device constraints. A Dynamic Universal Inverted Bottleneck (UIB) module architecture is designed with hierarchical decomposition into Perception Block (PB), Semantic Aggregation Block (SAB), and Decision Block (DB). Through dual pruning and reinforcement learning-based variant selection, the framework achieves superior accuracy-efficiency trade-off across diverse hardware. Experiments show up to 30% inference speedup and 67% memory usage compared to the baseline, with$<5 \%$accuracy drop.

To improve the task scheduling efficiency in computing power networks, this paper proposes a global task scheduling method based on network measurement and prediction in computing power networks (GTS-MP), which selects a more suitable computing power cluster and storage platform with a higher network availability score between them to implement tasks. The experiments are conducted on a test bench of urban area networks in 12 cities. The results show that this method greatly reduces the execution time of artificial intelligence model training tasks, with a maximum reduction of 40% compared to other plans.