In the rapid development of artificial intelligence, multimodal large models (VLM) have led a new wave of technological progress with their revolutionary breakthroughs in the field of natural language processing (NLP). In the wave of artificial intelligence, on-device multimodal large models (On-Device VLM) are becoming the new favourites of technological innovation with their rapid development speed and broad application prospects, and the demand for on-device inference is growing. This study conducts in-depth adaptation and optimization of multimodal large models on the Neural Network Processing Unit (NPU) based on the Qualcomm platform. By adopting the QNN (Qualcomm Neural Network) framework and model compression techniques such as quantization, pruning, knowledge distillation, low-rank factorization, and Lookahead decoding, efficient inference acceleration on Qualcomm NPU is achieved, significantly improving the model’s response speed and decoding efficiency. Experimental results show that the optimized model has significant improvements in first response time and decoding speed, providing a new solution for on-device AI applications.

A pivotal problem in the Internet of Things (IoT) is resource allocation, where the goal is to optimize allocation strategies of IoT resources. In general, resource allocation problems are formulated as constrained optimization problems, which can be effectively solved by the zeroth-order Frank-Wolfe algorithm. However, the existing zeroth-order Frank-Wolfe algorithms suffer from slow convergence since they scale the zeroth-order gradient in all directions. For this reason, we propose a faster zeroth-order Frank-Wolfe algorithm, referred to as ZO-AdaSFW, which incorporates adaptive gradient methods and the variance-reduced technique (SPIDER) into the zeroth-order Frank-Wolfe algorithm. Moreover, ZO-AdaSFW can achieve the convergence rate of$O(T^{-1}) $in the convex setting, where T is the time horizon. Meanwhile, we also prove that ZO-AdaSFW has the best-known convergence rate$O(T^{-1/2})$among zeroth-order algorithms in the nonconvex setting. In addition, the experimental results show that the performance of ZO-AdaSFW outperforms state-of-the-art zeroth-order Frank-Wolfe algorithms on different applications.

Multiple unmanned aerial vehicles (UAVs) can function as aerial base stations to provide flexible and reliable communication services for massive ground devices (GDs). It is quite a challenging task to analyze such multi-UAV networks when considering practical mutually exclusive relationships among UAVs. Based on the tools of stochastic geometry, we develop in this paper a theoretical framework for modeling and analyzing the coverage probability and average rate in a 3D air-ground network with UAVs following Matérn Hard-Core Point Process (MHCPP) distribution. As the tractable probability generating functional (PGFL) for such repulsive point processes is unavailable, we employ the approximate signal to interference ratio (SIR) analysis based on the Poisson point process (ASPPP) to obtain the Laplace transform approximation expression of an arbitrary GD’s cumulative interference in the network by considering both line-of-sight (LoS) and none-line-of-sight (NLoS) communications. Finally, extensive simulation results are presented to validate the effective and accuracy of our proposed framework.

Multiple unmanned aerial vehicles (UAVs) can function as aerial base stations to provide flexible and reliable communication services for massive ground devices (GDs). It is quite a challenging task to analyze such multi-UAV networks when considering practical mutually exclusive relationships among UAVs. Based on the tools of stochastic geometry, we in this paper develop a theoretical framework for modeling and analyzing aerial networks with UAVs following Matérn hard-core point process (MHCPP). As the tractable probability generating functional (PGFL) of repulsive point processes is unavailable, we employ an approximate approach based on the Poisson point process to analyze the cumulative interference and the signal-to-interference ratio (SIR) of a typical GD. By considering both line-of-sight (LOS) and none-line-of-sight (NLOS) communications, we obtain the approximation expressions of the network coverage probability and average rate. Finally, extensive simulation results are presented to validate the efficiency and accuracy of our proposed framework.

Along with the fast growing web-based applications, the recommender system is now attracting much attention due to its core function that matches the target users’ interest with the potential resources from the massive online information. Since the recommender system is a user centric application, in this work, we propose a recommendation framework based on user interaction, so as to explore the user’s real-time interest from the instant feedback. Naturally, we utilize the tag information assigned to different resources as the medium for user interaction. During the interaction, the most effective tags will be provided for users to choose, and the chosen tag words will be considered as the personalized preference and utilized to dynamically adjust the recommendation list during the process. However, the interaction procedure may cause the problem of potential false dismissal during the candidate filtering. In this work, we propose to analyze the association between different tags, and utilize the tag co-occurrence to refine the recommendation candidate, so as to avoid false dismissal. To generate the recommendation list from the filtered candidates, we design the representation of user and resource characteristics based on tag information and user historical behavior. We distinguish the significance of each tag word for the corresponding resource item, so as to precisely describe the item feature. Probabilistic matrix factorization is employed in our work to overcome the rating sparsity, which is enhanced by embedding the similar user and resource information. The experiments on real-world datasets demonstrate that the proposed algorithm can achieve more accurate predictions and higher recommendation efficiency.

Automobile source is a significant contributor to ambient particulate matter (PM), especially for PM2.5, which has become of growing public concern due to its urgency and sensitivity related to human health. A set of hourly PM2.5data from 35 monitoring stations published by Beijing Environmental Protection Bureau (BJ-EPB), covering 2 months during autumn 2013 was collected. The association between PM2.5concentrations at midnight and morning rush hours has been analyzed to estimate PM2.5concentration from automobile source. Spatial interpolation method was conducted in order to form a PM2.5distribution map of automobile source. Integrated with population data, we calculated the population exposure to automobile source PM2.5in each geographic census unit. Computational and analytical results show that the hot spots are mainly concentrated in the south and northwest of Beijing, especially near the main ring roads.

Domestic heating is a significant contributor to haze pollution in winter. Its role was estimated in a series of severe haze pollutions based on satellite and surface observations. The variations of haze clouds observed from satellite had different trends with the surface observations, indicating that anthropogenic emissions near surface had no direct relationship with haze clouds. Statistical results showed that domestic heating contributed to the increase in most of the monitored air pollutants, especially for SO2. Spatial interpolation and overlay analysis were conducted to get its spatial features in order to clarify its sources. Under the influence of transport from industrial regions of Hebei and Tianjin, the spatial distribution of particulate matter (PM2.5) induced from domestic heating were mainly concentrated in the south of Beijing. However, due to the scattered coal-fired heating in rural areas, domestic heating-induced gaseous pollutant (SO2), were mainly concentrated in suburban regions.