The Mixture-of-Expert (MoE) architecture constitutes a significant advancement in the scalability of large- language models (LLMs), achieving better performance compared to traditional dense models with the same size of activated parameters. However, the MoE models introduce two main challenges for their edge deployment: (1) increased memory storage requirements due to the sparse parameter composition and (2) aggravated yon Neumann bottleneck caused by the inherent unpredictability of dynamic expert routing. Recent trends in state-of-the-art MoE models suggest that scaling up both activated and the total number of experts leads to better model performance. However, this multi-expert activated LLM architecture further exacerbates the aforementioned issues. To address these challenges, this paper presents a 3D stacked near-DRAM computing architecture based on available hybrid bonding and mini-TSV techniques, thereby increasing both storage capacity with the same die area and memory bandwidth for MoE models. We also employ 6T-SRAM-based Digital Computing-in-Memory (DCIM) macros as the primary computing components and weight buffers to optimize the effective throughput of LLM inference. Furthermore, we propose a hybrid parallel computation scheme with Greedy-Swap Expert Grouping (GSEG) strategy. It balances the weight-parameter placement between tile groups in 3D-MoE architecture to enhance hardware utilization and overall performance. Through comprehensive benchmarking, the proposed accelerator demonstrates 143.10× and 68.89× speedup in token throughput, and 870.98× and 47.11× improvements in energy efficiency compared to CPU and GPU, respectively, confirming its advantages for multi-expert activated LLM inference.

Hydropower is an important and clean renewable energy source . The construction of hydropower projects involves huge investments, and it's important to estimate the investment accurately in the design and planning stage. Machine learning can construct investment estimation models based on a large number of samples. However, in the hydropower industry, given that Engineering Procurement Construction (hereafter EPC) is an emerging investment and construction model, relatively few samples are available for investment estimation of EPC hydropower projects. This study aims to propose a few-shot machine learning method for hydropower project investment estimation. First, a comprehensive similarity index is proposed to select similar projects; after that, a hydropower project investment estimation artificial intelligence model is constructed by integrating an improved support vector machine and swarm intelligence optimization algorithm. In this study, the investment data of 36 EPC hydropower projects are collected. Experiments show that the proposed method can construct a hydropower project investment model using a small number of samples, and the model prediction accuracy is 97.1 %, which is better than the traditional machine learning method . The proposed method helps to provide support for the comparison of investment and construction programs for EPC hydropower projects, thus reducing the investment cost.

Blockchain-based federated learning (BFL) has gained attention for its potential to establish decentralized trust. While existing research primarily focuses on personalized frameworks for various applications, essential aspects including incentive mechanisms—critical for ensuring stable system operation—remain under-explored. To bridge this gap, we propose a game-theoretic incentive mechanism designed to foster active participation in BFL tasks. Specifically, we model a BFL system comprising a model owner (MO), i.e., task publisher, multiple miners, and training terminals, framing their interactions through two-tier Stackelberg games. In the first-tier game, the MO designs reward strategies to incentivize training terminals to contribute more data, enhancing model accuracy. The second-tier game introduces a multi-leader multi-follower Stackelberg game, enabling miners to set model packaging prices based on competitors' strategies and anticipated user behavior. By deriving the Stackelberg equilibrium, we identify optimal strategies for all participants, leading to an incentive mechanism balancing individual interests with overall performance. Compared to its benchmarks, our incentive mechanism offers 5.8% and 53.4% higher utilities in the two games compared to its alternatives, accelerating convergence and improving accuracy.

Modern inertial measurement units (IMUs) with low cost, small size, and low power consumption are the key to improving indoor positioning accuracy. However, the performance of IMU in current mobile phones is spotty. The IMU positioning error of mainstream cell phones is basically greater than 1%, which makes it difficult for indoor fusion positioning technology to achieve accuracy within 3 m on a mobile phone, seriously restricting the development of the industry. Therefore, the existing indoor fusion algorithms need to introduce additional hardware, such as high-density UWB/Wi-Fi/Bluetooth deployment, to compensate for the lack of IMU performance, which significantly increases the deployment cost, difficulty, and reduces environmental applicability. It is found that deep learning can greatly improve the performance of IMU, but traditional deep learning methods have problems, such as privacy protection, difficulty in data collection, and low training efficiency. Therefore, we propose a novel data-driven inertial navigation method based on federated learning, named FedINS, to improve IMU performance and solve the above-mentioned problems. In order to further improve performance and reduce hardware cost, we introduce the concept of segment fusion. FedINS2 (2 means the second S, which is the segment) is formed by combining FedINS with low-cost edge-end ranging, which has terminal computing capabilities and edge-end ranging capabilities. FedINS2 not only greatly improves the performance of the smartphone IMU from 3.6% to 0.8%, but also has the characteristics of privacy protection and efficient data collection. Experimental results demonstrate that the proposed data-driven inertial navigation algorithm is effective.

Not only has information technology evolved rapidly, but the spatial cognition theory for blind and visually impaired (BVI) people has also made great strides, which has opened up a new opportunity for indoor travel assistance systems (ITASs). However, there are still some issues that have not been effectively addressed due to the lack of guidance of the spatial cognition theory. Thus, this article presents a comparative survey among ITASs proposed in the last four years in an effort to inform researchers and developers about system problems and challenges and inform BVI people about the various types and functions of the ITAS. This article will also make researchers and developers aware of the importance of the spatial cognition theory. Furthermore, we give predictions for future trends based on a detailed analysis of 17 ITASs.

Magnetic induction (MI) has been proven to be an efficient wireless communication technique for overcoming the transmission challenges in some very harsh propagation environments, such as underground, underwater, etc. For a random distributed MI ad hoc network in a 3-D space composed of uniform medium, we propose a method for determining the required node density and transmitting power that creates an almost surely fully connected network. For which we involve an MI path-loss model and consider the effect of eddy currents, the effective coverage space and the expected node degree of an MI node are then calculated by a Lambert W-function-based integration. Finally, we propose optimized frequency selection methods for improving the connectivity of MI networks. In addition to an ideal frequency-switching optimization method, we provide for engineering applications a practical frequency-fixed optimization method, which is based on the gradient descent algorithm, where an improved initialization is used to reduce iterations.

Magnetic induction (MI) is a promising technique for communications in the very harsh propagation conditions like underground or underwater, etc. In this paper we investigate the connectivity issue of MI ad hoc networks deployed in a three-dimensional space. The transmitting distance of MI covers a quasi- ellipsoid space other than a standard sphere space in a traditional wireless propagation case. We derive a closed-form expression for the effective coverage space of an MI node to evaluate the expectation of node degree. Moreover, the probability of node isolation is derived, which provides an upper bound of network connectivity probability. With these results, pairs of the critical transmitting magnetic momentum and node density can be obtained to keep the network connected with a probability close to one. Our analytic method provide a guidelines for future studies on MI networking.