Haihang You

Transformer-based policy networks have shown promising prospects in Portfolio Selection (PS). However, constrained by high computational complexity, they primarily focus on short-term price series (e.g., the past 12 h), while overlooking valuable long-term features. Furthermore, the inherent volatility and noise tend to lead the model to overfit on low-level information. To tackle these challenges, we propose an innovative approach. This approach leverages a pre-training model to extract high-level patterns from extended price series (e.g., the past two weeks) and then enhances decision-making in policy networks. The pre-training model is designed based on discrete representations to achieve better generalization and interpretability. Specifically, we tokenize the price series into discrete tokens through Vector-Quantization Variational AutoEncoder (VQ-VAE) and encourage the pre-training model to reconstruct these tokens according to the masked price series. Then we introduce a compact encoder-only policy network named Portfolio Transformer Encoder (PTE). Finally, PTE is provided with high-level patterns from the pre-training model to make more comprehensive decisions. We term the whole approach as Vector-Quantization Portfolio Transformer Encoder (VQ-PTE). VQ-PTE demonstrates superior performance on real-world currency and S&P500 datasets, achieving a minimum improvement of 35% in returns. Additionally, visualization results highlight superior interpretability.

Superconducting circuit development represents a critical area of research in superconductivity, offering significant potential for applications in artificial intelligence and quantum computing. In the design of such circuits, inductance plays a pivotal role, second only to the Josephson junction in importance. As superconducting circuit designs grow increasingly complex, there is a pressing need for efficient and accurate inductance extraction methods. The two primary approaches for this purpose are the volume electric current integral equation and the London equation. While the London equation offers several advantages, its relatively high computational cost and lack of direct compatibility with analog simulation tools have hindered its widespread adoption in superconducting circuit design. To overcome these limitations, we developed a finite element solver for the London equation incorporating both a posteriori and semi a priori error estimators. Our adaptive solver can refine the mesh based on current density, thereby drastically improving computational efficiency. Numerical experiments indicate that the proposed method can reduce the computational time from 39.05% to 84.89% compared to a uniform mesh. Additionally, we introduce an improved inductance extraction method, facilitating seamless integration of the London equation model with analog simulation tools such as JoSIM. The efficacy of our approach is validated through numerical simulations on a testbench comprising multiple advanced ColdFlux circuits, achieving superior accuracy compared to uniform mesh methods. To the best of our knowledge, both our methodology and its application to such sophisticated circuits have not been previously reported. These advancements provide a valuable foundation for next-generation superconducting circuit design and the broader development of superconducting computing technologies.

Serverless computing, characterized by its pay-as-you-go and auto-scaling features, offers a promising alternative for High Performance Computing (HPC) applications, as traditional HPC clusters often face long waiting times and resources over/under-provisioning. However, current serverless platforms struggle to support HPC applications due to restricted inter-function communication and high coupling runtime. To address these issues, we introduce RHINO, which offers end-to-end support for the development and deployment of serverless HPC. Using the Two-Step Adaptive Build strategy, the HPC code is packaged into lightweight, scalable functions. The Rhino Function Execution Model decouples HPC applications from the underlying infrastructures. The Auto-scaling Engine dynamically scales cloud resources and schedules tasks based on performance and cost requirements. We deploy RHINO on AWS Fargate and evaluate it on both benchmarks and real-world workloads. Experimental results show that, when compared to the traditional VM clusters, RHINO can achieve a performance improvement of 10%-30% for small-scale applications and more than 40% cost reduction.

Efficiently processing massive satellite datasets is a critical challenge in the field of remote sensing. The limitations of computational resources, bottlenecks in I/O operations, and pressures from data storage and transmission have long constrained the efficiency of large-scale synthetic aperture radar(SAR) data processing. Based on the concept of cloud-native computing, this study proposes a modular, multi-node parallel framework for time-series InSAR processing. The framework leverages containerization and a microservices-based architecture, incorporating multi-level parallel methods to optimize existing workflows. It achieves efficient resource allocation, alleviates I/O load, and significantly enhances data processing performance. Experimental results demonstrate that, compared to local computing resources of similar scale, this approach improves the efficiency of key processing steps by 33.1% and 16.6%, respectively. When the data volume is increased several-fold, the program's processing efficiency remains stable. The framework exhibits excellent performance under large-scale tasks and diverse computational resource scenarios, with peak CPU utilization reaching nearly 100% and memory utilization stabilizing above 80%. Moreover, it achieves high-efficiency data read/write operations across varying task scales, showcasing outstanding resource scheduling capabilities, elasticity, and scalability. This framework offers an efficient and practical solution for large-scale InSAR data processing and paves the way for broader applications of cloud-native technologies in remote sensing data analysis.

Current artificial intelligence faces challenges in improving computational efficiency due to increasing scale and complexity. Superconducting circuit, as one of the most promising technologies in the post-Moore era, offers ultrahigh-speed computation and ultralow power consumption. Superconducting circuits are driven by pulses, which enables direct execution of pulse-based neuromorphic computing. Consequently, superconducting circuits hold the potential to facilitate higher efficiency and larger scale neuromorphic chips. However, existing efforts neglect the limitations and constraints of superconducting circuits, such as the extra overhead of pulse-based logic, the lack of superconducting memory, and low integration. Hence, their work cannot be utilized in fabricating real superconducting neuromorphic chips. This article introduces superconducting spiking neural network (SSNN), which aims to enable full neuromorphic computing on superconducting circuits. The design of SSNN addresses key issues including a superconducting circuit-based neuron model, weight processing methods suitable for superconducting pulses, and superconducting neuromorphic on-chip networks. SSNN enables complete neuromorphic computing on superconducting circuits. We validate the feasibility and accuracy of SSNN using a standard cell library of superconducting circuits and successfully fabricate the world's first superconducting neuromorphic chip. Our evaluation demonstrates a remarkable $50\times$ increase in power efficiency compared to state-of-the-art semiconductor designs.

Current approaches for cardiac amyloidosis (CA) identification are time-consuming, labor-intensive, and present challenges in sensitivity and accuracy, leading to limited treatment efficacy and poor prognosis for patients. In this retrospective study, we aimed to leverage machine learning (ML) to create a diagnostic model for CA using data from routine blood tests. Our dataset included 6,563 patients with left ventricular hypertrophy, 261 of whom had been diagnosed with CA. We divided the dataset into training and testing cohorts, applying ML algorithms such as logistic regression, random forest, and XGBoost for automated learning and prediction. Our model’s diagnostic accuracy was then evaluated against CA biomarkers, specifically serum-free light chains (FLCs). The model’s interpretability was elucidated by visualizing the feature importance through the gain map. XGBoost outperformed both random forest and logistic regression in internal validation on the testing cohort, achieving an area under the curve (AUC) of 0.95 (95%CI: 0.92–0.97), sensitivity of 0.92 (95%CI: 0.86–0.98), specificity of 0.95 (95%CI: 0.94–0.97), and an F1 score of 0.89 (95%CI: 0.85–0.92). Its performance was also superior to the serum FLC-kappa and FLC-lambda combination (AUC of 0.88). Furthermore, XGBoost identified unique biomarker signatures indicative of multisystem dysfunction in CA patients, with significant changes in eGFR, FT3, cTnI, ANC, and NT-proBNP. This study develops a highly sensitive and accurate ML model for CA detection using routine clinical laboratory data, effectively streamlining diagnostic procedures, and providing valuable clinical insights and guiding future research into disease mechanisms.

Graph neural network training is mainly categorized into mini-batch and full-batch training methods. The mini-batch training method samples subgraphs from the original graph in each iteration. This sampling operation introduces extra computation overhead and reduces the training accuracy. Meanwhile, the full-batch training method calculates the features and corresponding gradients of all vertices in each iteration, and therefore has higher convergence accuracy. However, in the distributed cluster, frequent remote accesses of vertex features and gradients lead to huge communication overhead, thus restricting the overall training efficiency. In this paper, we introduce the cached-based distributed full-batch graph neural network training framework (CDFGNN). We propose the adaptive cache mechanism to reduce the remote vertex access by caching the historical features and gradients of neighbor vertices. Besides, we further optimize the communication overhead by quantifying the messages and designing the graph partition algorithm for the hierarchical communication architecture. Experiments show that the adaptive cache mechanism reduces remote vertex accesses by 63.14% on average. Combined with communication quantization and hierarchical GP algorithm, CDFGNN outperforms the state-of-the-art distributed full-batch training frameworks by 30.39% in our experiments. Our results indicate that CDFGNN has great potential in accelerating distributed full-batch GNN training tasks.

Subgraph matching is a challenging NP-complete problem that involves finding identical subgraphs of a query graph $q$ in a larger data graph $G$ . It has numerous applications in diverse fields, including social and biological networks. However, existing subgraph matching algorithms assume that the graph structure is fixed, which limits their performance in solving more difficult matching cases. To address this issue, we propose a novel approach called Dynamic Graph Editing (DGE), which dynamically edits the query graph to optimize the subgraph matching algorithm. Based on this approach, we introduce an efficient enumeration method called Dynamic Graph Editing Enumeration, which significantly improves the performance of the algorithm. Our experimental results show that DGE outperforms current state-of-the-art algorithms in terms of computational efficiency and ability to solve more complex subgraph matching cases.