The rapid development of large language model (LLM) has established distributed training as the dominant paradigm, yet communication efficiency remains a critical challenge. Existing training cluster(e.g., Rail-Optimized architecture) based on electrical packet switch (EPS), suffering from high costs, excessive power consumption, and inefficient bandwidth utilization. While optical circuit switch (OCS) offers a promising alternative with its high bandwidth, low latency, and power efficiency, its rigid connectivity struggles to accommodate dynamic multi-task workloads, and its per-unit cost remains prohibitive at scale. To address these limitations, we propose ReFly, a reconfigurable architecture for LLM training based on OCS. By modeling GPU communication requirements, we design a Cycle Decomposition (CD) scheme for cluster construction and an Alternating Decomposition (AD) algorithm to dynamically schedule multiple OCS. Experimental results demonstrate that ReFly reduces deployment costs by 77% and power consumption by 98% compared to state-of-the-art Rail-Optimized architecture while achieving comparable performance, and a 229% higher bus bandwidth than Fat-Tree. These advancements position ReFly as an efficient and cost-effective solution for next-generation LLM training clusters.

The explosive growth of artificial intelligent data centers (AIDCs) and AI-driven workloads has led to massive, bursty, and dynamic inter-AIDC traffic, placing unprecedented demands on optical transport networks for capacity, flexibility, and ultra-low latency. To meet these challenges, multi-band wavelength-division multiplexing (MB-WDM) systems extend the spectrum beyond the conventional C-band. However, the interaction between nonlinear effects and dynamic traffic patterns leads to nonlinear bandwidth fragmentation (NBF) and degraded efficiency. Effective defragmentation requires modeling inter-service dependencies, as reallocating one service may impact others through nonlinear coupling. To address this, we propose TransMACB, a Transformer-enhanced multi-agent contextual bandit framework for globally optimized reassignment. Each service is assigned a dedicated agent, and the Multi-Agent Advantage Decomposition Theorem is applied to factorize the joint optimization problem. This enables a Transformer encoder-decoder to capture structured inter-agent dependencies and generate context-aware policies. Simulations demonstrate that TransMACB significantly improves overall network performance, highlighting the necessity of combining multi-agent intelligence with Transformer-based coordination in managing NBF.

Driven by the diverse requirements of industrial internet of things (IIoT) services, passive optical network (PON) has emerged as a promising technology for Industrial Internet, owing to its advantages in multi-service support, immunity to electromagnetic interference (EMI) and cost-effective deployment. However, due to the highly dynamic nature of industrial services and their diverse demands, it poses great challenges to guarantee the deterministic latency demand of time sensitive services while avoiding inevitable conflicts among various services. To address this problem, we propose a flexible-grouping based collaborative dynamic bandwidth allocation (FGC-DBA) scheme. Specifically, FGC-DBA characterizes a cluster-based peak-valley compensation algorithm (CPVC) to achieve deterministic latency of highly dynamic services by fluctuation counterbalance. It aims to find a group of services with complementary demand, thereby dynamic nature only exists within intra-group rather than inter-group and the latency fluctuation is fixed as the size of each group. Furthermore, we present a fine-coarse collaborative transmission window allocation strategy (FCCTA) to accurately allocate transmission window for various services based on the CPVC algorithm. It aims to reduce the conflicts and meet the deterministic latency demand simultaneously when considering the various services’ demand. In the experiment, we demonstrate that compared to existing schemes, FGC-DBA scheme reduces jitter by at least 75% for cyclic time-sensitive (CTS) services and brings queuing time close to zero, while reducing jitter by up to 50% for non-cyclic time-sensitive (TS) services. Notably, all of these improvements are achieved with only a 5% increase in running time.

Multi-band transmission has been considered a competitive solution for expanding optical network capacity in the near term. Research on multi-band transmission has currently been extensive, covering various aspects. Nevertheless, many studies focus solely on end-to-end transmission systems and overlook the dynamic nature of networks, employing simple static methods for all scenarios. To address these challenges, we propose an online transmission optimization method based on multi-stage programmable Raman amplifiers (PRAs). The proposed method includes designing optimal gain profiles based on real-time feedback of channel conditions in dynamic networks and quickly predicting pump settings through an inverse mapping model. Additionally, it dynamically adjusts the pump settings in response to various scenarios, achieving high-precision gain profiles within four iterations at most, thus leading to efficient transmission optimization. Experimental simulation results indicate that, when utilizing 400G QPSK modulation, the optimization method can achieve an excellent average generalized signal-to-noise ratio (GSNR) improvement, reaching up to 2.45 dB compared to traditional methods. We have further validated the benefit of the optimization method under various service scenarios, with an average GSNR improvement exceeding 2.1 dB, demonstrating the adaptation for dynamic service demands.

Edge–cloud collaboration integrated with network virtualization is indispensable for diversified edge services. Meanwhile, the multilayer elastic optical network (ML-EON) is a promising underlying network for virtual network requests (VNRs) customized for edge–cloud collaborative services. However, the joint allocation of computing resources and high-dimensional ML-EON resources in virtual network embedding (VNE) will pose great computational complexity for online service deployment. In this article, we propose an approximately lossless model compression mechanism to ease the computing burden of the VNE over ML-EON for edge–cloud collaborative services. An integer quadratic constraint programming (IQCP) model is established for the problem. Model compression based on virtual link mapping cost estimation (VLMCE) is investigated to shield the variables and constraints related to ML-EON. In particular, the resource metric and topology metric are introduced into VLMCE to cope with resource contentions among virtual links in the same VNR, and improve estimation accuracy. The model solving relies on Hopfield neural network (HNN) is further studied, where optimizing the compressed model is losslessly converted to minimizing the energy function of HNN. The experimental results reveal that the proposed mechanism guarantees an approximately lossless algorithm performance and a high-time efficiency compared with the original IQCP model. The performances of VNR cost and blocking ratio are also promoted compared with the benchmarks.

With the increasing demand of network slices in vertical industries, slice resource provisioning in transport networks has encountered two challenges, one is efficient slice resource provisioning in the presence of traffic uncertainty of slices, and another is flexible slice resource isolation for customizable isolation needs. In this paper, we propose an innovative flexible hybrid isolation model to support any customized resource isolation from complete isolation to full sharing, and solve the slice resource provisioning problem named Hybrid Slicing Minimum Bandwidth (HSMB) by considering traffic prediction error to mitigate the negative impact of traffic uncertainty in the proposed model. After analyzing the HSMB problem, 1) we first try to solve the problem in steps and decompose the HSMB problem into grouping sub-problem and adjusting sub-problem, 2) we then propose a low-complexity dynamic programming grouping algorithm and a fast iterative adjustment algorithm for the two sub-problems based on probabilistic feature-based analysis, 3) we combine the algorithms of the two sub-problems and further propose a linking algorithm for the potential insufficient resource dilemma and high computational complexity dilemma to improve the efficiency of the solution. The numerical results show that the proposed flexible hybrid isolation model with different factors can facilitate flexible slice isolation with customized isolation demands, while the proposed algorithm can realize efficient slice resource provisioning with a probabilistic guarantee. The comparison result shows the proposed algorithms outperform the other benchmark algorithms.

The commercialization of 5G and the explosive emergence of new applications stimulate the exponential growth of network traffic and diversification of services. It is promising to integrate multi-band elastic optical network (MBEON) and network virtualization for large volume traffic transmission and highly diverse services. However, performing virtual network embedding (VNE) for network virtualization over MBEON faces the challenge of severe inter-channel stimulated Raman scattering (ISRS) effect, which complicates the underlying physical layer effect in the substrate network. In this paper, we investigate the ISRS-aware VNE over MBEON, where a cross-matching mechanism is proposed for virtual node mapping (VNM) and a hypergraph is introduced for parallel virtual link mapping (VLM). A lightpath-level integer linear programming model is first formulated. To integrate the cost and availability of VLM, which significantly affect the performance of the VNE under the ISRS effect, into the VNM process, the “virtual node-substrate node” mapping pairs are specifically evaluated through the cross-matching mechanism. Moreover, to tackle the couplings among multiple lightpaths induced by the wide spectrum ISRS effect, hypergraphs are used to model the ISRS effect-aware quality of transmission (QoT) constraints among multiple lightpaths. A hypergraph maximal weight independent set heuristic is presented for lightpath selection, which guarantees the obedience of basic constraints and generates near-optimal solutions. Experimental results show that the proposed methods decrease blocking ratio by more than 30% compared with the benchmarks with similar computational complexity.

With the rapid development, artificial intelligence (AI) technology may occupy the job positions which require only simple knowledge. Autonomous learning ability has been one of the core abilities that needs to be cultivated in engineering education. However, how to leverage students’ autonomous learning in the practice session of curriculum is still facing several challenges. In view of these challenges, we develop an AI empowered remote experiment platform, which has independent intelligent algorithm center for students to deploy their own algorithms into the real experiment scenarios. Also, in this platform, there is a digital-twins engine to give real-time feedback to the students, helping them to improve their algorithms and even the whole projects. Based on this experiment platform, we propose a teacher-guided autonomous learning practice teaching mode, including autonomous goal setting, autonomous practice process and autonomous feedback optimization, in which the teachers will be a guide to guarantee the learning objective is achieved. A survey was conducted, showing that under this practice teaching mode, students have a deeper understanding of theoretical knowledge, more obvious cultivation of autonomous learning ability, and higher satisfaction with the course.

With the explosive growth of edge applications in the 5G/B5G era, edge-cloud collaboration (ECC) is playing a prominent role in edge service provisioning. For highly diversified edge-cloud collaborative services (ECSs), the joint allocation of heterogeneous computing resources in heteroge-neous servers and multi-dimensional underlying optical network resources should be conducted. In this paper, we investigate the heterogeneous task scheduling for ECSs over multilayer elastic optical network (ML-EON), which involves the joint allocation of heterogeneous computing resources in edge and cloud servers and high-dimensional network resources. We propose a Task-Node Matching Score (TNMS) based method, which evaluates the fitness for each mapping tuple between each task in ECS and each substrate node in ML-EON, and adaptively generates a specific matching score for each task-node pair. Furthermore, TNMS is extended with a pre-allocation mechanism (TNMS-Pre) to estimate the costs of multi-dimensional resources in ML-EON for virtual link (VL) mapping. The estimated VL mapping costs are integrated into the matching scores to guide the task placement to be cost-efficient. To guarantee the feasibility, a maximal weight matching (MWM) based method is presented to determine the task placement schemes. Simulation results demonstrate the effectiveness of the adaptive scoring for heterogeneous task placement and the pre-allocation mechanism for reducing the ML-EON costs.

The intelligent service function chains (SFCs) provisioning is of great significance for agile deployments of 5G vertical applications. However, the heterogeneities of entities in substrate network (SNet) and SFCs hinder deep learning (DL) models to fully integrate the information of SNet and SFCs. Furthermore, the potential infeasibility of output policies and difficulty in training data acquisition also pose challenges to DL methods. To overcome the above limitations, we propose a Differentiated Weight Graph Convolutional Neural Network (DWGCN) model, which configures different weights for different kinds of entities, to predict the optimal virtual network function (VNF) placements. Moreover, the model is integrated with maximal weight matching to enhance the feasibility of VNF placement policies. A transfer learning method is further introduced to reduce the required training data with knowledge transfer. Experimental results demonstrate the effectiveness of the proposed methods in SFC mapping cost, high time efficiency, and knowledge transferability.

Summary As an important method to diagnose gastric cancer, gastric pathological sections images (GPSI) are hard and time‐consuming to be recognized even by an experienced doctor. An efficient method was designed to detect gastric cancer in magnified (20×) GPSI using deep learning technology. A novel DenseNet architecture was applied, modified with a multistage attention module (MSA‐DenseNet). To develop this model focusing on gastric features, a two‐stage‐input attention module was adopted to select more semantic information of cancer. Moreover, the pretraining process was divided into two steps to improve the effect of the attention mechanism. After training, our method achieved a state‐of‐the‐art performance yielding 0.9947 F1 score and 0.9976 ROC AUC on a test dataset. In line with our expectation in clinical practice, a high recall (0.9929) was produced with high sensitivity to the positive samples. These results indicate that this new model performs better than current artificial detection approaches and its effectiveness is therefore validated in cancer pathological diagnoses.

Atmospheric visibility is an indicator of atmospheric transparency and its range directly reflects the quality of the atmospheric environment. With the acceleration of industrialization and urbanization, the natural environment has suffered some damages. In recent decades, the level of atmospheric visibility shows an overall downward trend. A decrease in atmospheric visibility will lead to a higher frequency of haze, which will seriously affect people's normal life, and also have a significant negative economic impact. The causal relationship mining of atmospheric visibility can reveal the potential relation between visibility and other influencing factors, which is very important in environmental management, air pollution control and haze control. However, causality mining based on statistical methods and traditional machine learning techniques usually achieve qualitative results that are hard to measure the degree of causality accurately. This article proposed the seq2seq-LSTM Granger causality analysis method for mining the causality relationship between atmospheric visibility and its influencing factors. In the experimental part, by comparing with methods such as linear regression, random forest, gradient boosting decision tree, light gradient boosting machine, and extreme gradient boosting, it turns out that the visibility prediction accuracy based on the seq2seq-LSTM model is about 10% higher than traditional machine learning methods. Therefore, the causal relationship mining based on this method can deeply reveal the implicit relationship between them and provide theoretical support for air pollution control.

Summary The motif discovery of multi‐dimensional time series datasets can reveal the underlying behavior of the data‐generating mechanism and reflect the relationship between time series in different dimensions. The study of motif discovery is of important significance in environmental management, financial analysis, healthcare, and other fields. With the growth of various information acquisition devices, the number of multi‐dimensional time series datasets is rapidly increasing. However, it is difficult to apply traditional multi‐dimensional motif discovery methods to large‐scale datasets. This paper proposes a novel method for motif discovery and analysis in large‐scale multi‐dimensional time series. It can effectively find multi‐dimensional motifs and the correlation among the motifs. The experimental results show that the proposed method achieves better performance than the related arts on synthetic and real datasets. It is further validated on practical air quality data and provides theoretical support for real air pollution control in places such as Beijing.

Identifying if a subsurface target is salt or not automatically and accurately is of vital importance to oil drilling. But unfortunately, obtaining the precise position of large salt deposits is very difficult. Professional seismic imaging still requires the interpretation of salt bodies by experts. This leads to very subjective, highly variable renderings. More alarmingly, it leads to potentially dangerous situations for drillers in oil and gas companies. In this paper, a Squeeze-Extraction Feature Pyramid Networks (referred to as Se-FPN) was proposed to tackle the task of image segmentation of salt deposits. Specifically, we utilized SeNet as backbone so as to implicitly learn to suppress irrelevant regions in an input image while highlighting salient features useful for the task. Considering the importance of multi-scales information, we proposed an improved FPN to integrate information of different scales. In order to further fuse the information from multiple scales, the Hypercolumns module was inserted at the end of the network. The proposed Se-FPN has been applied to the TGS Salt Identification Challenge and achieved high quality segmentation effect. The Mean Intersection over Union value can reach 0.86.

Air pollution is threatening human's health since the industrial revolution, but there are not efficient ways to solve air pollution, so forecasting air quality has become an efficient measure to prevent citizens from hurting of heavy air pollution. In this paper, we proposed an advanced Seq2Seq (Sequence to Sequence) model called attention-based air quality forecasting model (ABAFM) whose RNN encoder is replaced by pure attention mechanism with position embedding. This improvement not only reduces the training time of Seq2Seq model with attention but also enhances the robustness of Seq2Seq models. We implemented ABAFM in Olympic center and Dongsi monitoring stations in Beijing to forecast PM2.5 in future 24 hours. The experimental results showed that the proposed model outperformed the related arts, especially in sudden changes.

The diagnosis of breast cancer in the middle and early period is conducive to later treatment, but the current diagnosis rate is not very desirable. Using machine learning to predict the benign and malignant of breast cancer can provide some assist to doctors' treatment in clinical practice. In this paper, we have collected data from digitized images of a fine needle aspirate (FNA) of a breast mass. They describe characteristics of the cell nuclei presented in the image. This work adopts several feature selection methods to select the most related features for breast cancer diagnosis. Based on the selected features, four machine learning models, Support Vector Machine (SVM), Decision Tree (DT), AdaBoost and Random Forest (RF) are built and their performance are evaluated. The experimental results show that the accuracy of RF is higher than the other three methods.

At present, high road surface temperature (RST) is threatening the safety of expressway transportation. It can lead to accidents and damages to road, accordingly, people have paid more attention to RST forecasting. Numerical methods on RST prediction are often hard to obtain precise parameters, whereas statistical methods cannot achieve desired accuracy. To address these problems, this paper proposes GBELM-RBF method that utilizes gradient boosting to ensemble Radial Basis Function Extreme Learning Machine. To evaluate the performance of the proposed method, GBELM-RBF is compared with other ELM algorithms on the datasets of airport expressway and Badaling expressway during November 2012 and September 2014. The root mean squared error (RMSE), accuracy and Pearson Correlation Coefficient (PCC) of these methods are analyzed. The experimental results show that GBELM-RBF has the best performance. For airport expressway dataset, the RMSE is less than 3, the accuracy is 78.8% and PCC is 0.94. For Badaling expressway dataset, the RMSE is less than 3, the accuracy is 81.2% and PCC is 0.921.

The scale of the modern city has been expanding, which leads to a lot of serious environmental pollution problems. Among them, the air pollution problem is the most prominent. In order to control the air pollution in urban cities, the government has deployed a lot of air pollutant monitoring equipment which produce massive multi-dimensional time series data. Through the motif discovery and analysis of these multi-dimensional time series, we can find the relationships and the rules between air pollutants to provide support and suggestions for controlling the air pollution. In this paper, a novel method on motif discovery and analysis for large-scale multi-dimensional time series data is proposed. The new method can effectively find multi-dimensional motifs and the correlation between them as much as possible, which reveals the underlying rule of different air pollutants. It is validated on practical historical data of air pollutants in Beijing. The experimental results show that the proposed method could obtain better performance than the related work.

Admittedly, the broadband access network has been improved largely with the developing technologies, it is still facing challenges on managing and maintaining existed resources efficiently. In order to build up an intelligent and open network architecture, and solve the problem of heterogeneous networks consisting of devices from different vendors, we have worked out a web-based managing system implementing the concept of Software-Defined Network (SDN) and Network Functions Virtualization. The controlling plane is centered into the Controller layer and decoupled from the forwarding layer. The frame we proposed is also applicable for old routers, which do not support SDN, with an Agent on it to translate the OpenFlow messages. For a more intelligent routing schema, the controller is able to calculate with a fine-tuned ant colony optimization algorithm. At the top of the controller, the web-based managing system is accessible for operators, and they can manage the resource they possessed. With the above framework, we achieve the goal of an intelligent and open network architecture and verify it.

In this paper we present a novel and robust framework that decomposes continuous people detection into three parts, including off-line detection, tracking and learning. We introduce temporal coherence and spatial constraints into off-line detection phase by collecting a dynamical model from tracker which is estimated and updated by the learning algorithm. This method makes the detector aim at regions where a potential target will appear in the next frame, capable of handling pedestrians with occlusions and variety of scales, which as result greatly improves performance of pedestrian detection. We carry out a quantitative and qualitative evaluation on the public datasets.

Exemplar-based methods have been widely used in human action recognition. To analyze human action in monocular video has always been a challenging problem, due to depth information loss and ambiguities. In this paper we presented a method applying skin color detection and then calculating relative positions of face and hands to solve self-occlusions and to eliminate ambiguities. Then we applied 2D shape analysis to classify basic human actions. Several low level features were used to describe shapes, which needs less computation and can improve recognition speed to real-time level. We testified our method on a public action database and got satisfying results.

This paper considers the comparison between optical and electrical interconnect system from a parallel computing performance view. The motivation of the work is based on increasing demand on parallel computing tasks and the rapid development of optical chip-to-chip interconnects techniques. To meet the increasing demand of large-scale parallel or multiprocessor computing tasks, an analytic method to evaluate the computing performance of interconnect systems is proposed in this paper. The bandwidth-limit model and full-bandwidth model are both under our investigation. The paper characterizes the influence of unit processing time, communication overhead, processor number and the transmission time. The speedup and efficiency are selected to represent the parallel performance of an interconnect system. The analytic expressions of these two indexes are derived in the paper. Deploying the proposed models, we depict the performance gap between the optical and electrical interconnect systems. Results show that the large communication bandwidth optical chip-to-chip system has an obvious speedup gain, which is up to 93%. However, the existence of efficiency peak point indicates that the immoderate pursuing of high-bandwidth have no use for system efficiency improvement.