Domain generalization (DG) for object detection aims to enhance detectors’ performance in unseen scenarios. This task remains challenging due to complex variations in real-world applications. Recently, diffusion models have demonstrated remarkable capabilities in diverse scene generation, which inspires us to explore their potential for improving DG tasks. Instead of generating images, our method extracts multi-step intermediate features during the diffusion process to obtain domain-invariant features for generalized detection. Furthermore, we propose an efficient knowledge transfer framework that enables detectors to inherit the generalization capabilities of diffusion models through feature and object-level alignment, without increasing inference time. We conduct extensive experiments on six challenging DG benchmarks. The results demonstrate that our method achieves substantial improvements of 14.0% mAP over existing DG approaches across different domains and corruption types. Notably, our method even outperforms most domain adaptation methods without accessing any target domain data. Moreover, the diffusion-guided detectors show consistent improvements of 15.9% mAP on average compared to the baseline. Our work aims to present an effective approach for domain-generalized detection and provide potential insights for robust visual recognition in real-world scenarios. The code is available at Generalized Diffusion Detector

With the continuous development of technologies such as the Internet of Things (IoT), scheduling issues in edge-cloud systems are becoming a research focus. Deep reinforcement learning (DRL) has become an effective way to address scheduling issues in edge-cloud systems due to its ability to interact with the environment and engage in adaptive learning to solve complex decision-making problems. However, due to the increasing scale of edge-cloud systems, traditional DRL still faces challenges in scheduling, such as slow convergence and high computational requirements. To address these challenges, we propose a diffusion model-based deep reinforcement learning algorithm called DMaiS, which utilizes the diffusion model as the policy network in the advantage actor-critic (A2C) to expedite policy learning. Additionally, we develop a distributed service orchestration approach utilizing multi-agent advantage actor-critic (MAA2C) to effectively and flexibly manage extensive and intricate cloud service resources. The experimental results using real-world data demonstrate that compared to the baseline algorithms, DMaiS achieves a higher system throughput rate and a lower scheduling latency when managing scheduling for the edge-cloud system. It also exhibits a faster convergence speed compared to traditional DRL algorithms.

Indoor localization has become an essential demand driven by indoor location-based services (ILBSs) for mobile users. With the rising of Internet of Things (IoT), heterogeneous smartphones and wearables have become ubiquitous. However, the ILBSs for heterogeneous IoT devices confront significant challenges, such as received signal strength (RSS) variances caused by hardware heterogeneity, multipath reflections from complex environments, and localization time restricted by computation resources. This article proposes EdgeLoc, a robust and real-time indoor localization system toward heterogeneous IoT devices to solve the above challenges. In particular, the RSS fingerprinting data of Wi-Fi is employed for localization and tackling the heterogeneity of IoT devices in twofold. First, feature-level and signal-level solutions are presented to address the random RSS variances. At the feature level, this work proposes a novel capsule neural network model to efficiently extract incremental features from RSS fingerprinting data. At the signal level, a multistep dataflow is further devised to process RSS fingerprints into image-like data, which utilizes the feature matrix to reduce absolute sensing errors introduced by hardware heterogeneity. Second, an edge-IoT framework is designed to utilize the edge server to train the deep learning model and further supports real-time localization for heterogeneous IoT devices. Extensive field experiments with over 33 600 data points are conducted to validate the effectiveness of EdgeLoc with a large-scale Wi-Fi fingerprint data set. The results show that EdgeLoc outperforms the state-of-the-art SAE-CNN method in localization accuracy by up to 14.4%, with an average error of 0.68 m and an average positioning time of 2.05 ms.

An increasing number of big data applications in various domains generate datasets continuously, which must be processed for various purposes in a timely manner. As one of the most popular streaming data processing systems, Spark Streaming applies a batch-based mechanism, which receives real-time input data streams and divides the data into multiple batches before passing them to Spark processing engine. As such, inappropriate system configurations including batch interval and executor count may lead to unstable states, hence undermining the capability and efficiency of real-time computing. Hence, determining suitable configurations is crucial to the performance of such systems. Many machine learning- and search-based algorithms have been proposed to provide configuration recommendations for streaming applications where input data streams are fed at a constant speed, which, however, is extremely rare in practice. Most real-life streaming applications process data streams arriving at a time-varying rate and hence require real-time system monitoring and continuous configuration adjustment, which still remains largely unexplored. We propose a novel streaming optimization scheme based on Simultaneous Perturbation Stochastic Approximation (SPSA), referred to as NoStop, which dynamically tunes system configurations to optimize real-time system performance with negligible overhead and proved convergence. The performance superiority of NoStop is illustrated by real-life experiments in comparison with Bayesian Optimization and Spark Back Pressure solutions. Extensive experimental results show that NoStop is able to keep track of the changing pattern of input data in real time and provide optimal configuration settings to achieve the best system performance. This optimization scheme could also be applied to other streaming data processing engines with tunable parameters.

With the unprecedented demand of location-based services in indoor scenarios, wireless indoor localization is emerging as an essential application for mobile users. While the line-of-sight GPS signal is not available at indoor spaces, WiFi fingerprinting using received signal strength (RSS) has become popular with its ubiquitous accessibility. Although the fingerprinting data can be easily collected by portable mobile devices, to achieve robust and efficient indoor localization remains challenging with two constraints. First, the localization accuracy will be degraded by the random fluctuation of signals that caused by multipath effects from RSS signals. Second, indoor localization algorithms are time-consuming due to the handcrafting features and complex filtering on raw dataset. To achieve high localization accuracy with WiFi fingerprinting, in this paper, we propose CapsLoc, a robust indoor localization system by using capsule networks. Specifically, the capsule network model can efficiently extract hierarchical structures from WiFi fingerprint with three main components, including a convolutional layer, a primary capsule layer and a feature capsule layer. We conduct a real-world experimental field test with over 33600 data points. The experimental results show that CapsLoc can achieve accurate indoor localization with an averaged error of 0.68 m, which outperforms conventional machine learning methods (KNN and SVM) and existing deep learning methods (CNN and SAE-CNN).

Many next-generation scientific and business applications feature large-scale data-intensive workflows, which require massive computing resources for execution on high-performance clusters in cloud environments. Such computing resources (e.g., VCores and virtual memory) requested through parameter setting in big data systems, if not fully utilized by workloads, are simply wasted due to the nature of exclusive access made possible by containerization. This necessitates accurate modeling and prediction of workflow performance to make an effective recommendation of appropriate parameter settings to end users. However, it is challenging to determine optimal workflow and system configurations due to the large parameter space and the interaction between various technology layers of big data systems. Towards this goal, we propose a machine learning-based feature selection method to identify influential parameters based on historical performance measurements of Spark-based computing workloads executed in big data systems with YARN. We first identify a comprehensive set of parameters across multiple layers in the big data technology stack including workflow input structure, Spark computing engine, and YARN resource management. We then conduct an in-depth exploratory analysis of their individual and coupled impact on workflow performance, and develop a performance-influence model using random forest for prediction. Experimental results show that the proposed approach identifies important features for performance modeling and achieves high accuracy in performance prediction.