Home Archive
Archive
2024 Volume 61 Issue 11  Published: 2024-11-01
    Research Highlights
  • Ennan Zhai , Jiamin Cao , Kun Qian , Yu Guan
    doi: 10.7544/issn1000-1239.202440576

    Large language models (LLMs) with hundreds of billions of parameters have brought significant technological and business transformations to today’s AI and cloud services. However, there exists a fundamental difference in network pattern between LLM training and general cloud computing (e.g., Amazon EC2 Elastic compute service), leading to a variety of new challenges. These challenges mainly include load balancing difficulties due to the traffic pattern difference (Challenge 1), the impact of multi-job communication contention on GPU utilization (Challenge 2), and high sensitivity to network failures (Challenge 3). Therefore, data center network technologies designed for general cloud computing (e.g., network architecture, routing, communication scheduling, and reliability) are no longer suitable for LLM training today. This necessitates the development of new data center networks and accompanying technical solutions specifically for LLM training. We introduce Alibaba Cloud’s high-performance network (HPN) and the multi-job communication scheduling approach Crux, designed to address the aforementioned challenges. HPN introduces a two-layer, dual-plane network architecture, which not only achieves high-speed interconnectivity for 15 000 GPUs within a Pod but also ensures precise routing suitable for LLM training (addressing Challenge 1). Furthermore, HPN proposes a novel dual-top-of-rack (ToR) design, replacing the traditional single ToR switch connection in data center networks and fundamentally avoiding single-point failure reliability risks (partially addressing Challenge 3). To tackle Challenge 2, Crux reduces the NP-complete problem of optimizing GPU utilization by modeling it as a communication scheduling issue related to GPU computational intensity. Crux then proposes an algorithm that prioritizes the flows of job with higher GPU computational intensity, significantly reducing multi-job communication contention and improving GPU utilization. Compared with the state-of-the-art efforts, Crux increases GPU utilization by up to 23%. Both HPN and Crux have been deployed and used in Alibaba Cloud production for over eight months and will continue to evolve and iterate. Building on this, we further envision possible research directions in the field of LLM training and inference, providing guidance for subsequent work.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Qingyong Deng , Qinghua Zuo , Zhetao Li , En Wang , Bin Guo
    doi: 10.7544/issn1000-1239.202440302

    Reputation management is an effective way to improve the quality of data collection in mobile crowdsensing, but existing subjective evaluation-based schemes are difficult to realize the real evaluation of worker reputation under privacy protection. To this end, we propose a bilateral reputation evaluation privacy-preserving (BREPP) scheme based on blockchain. The scheme firstly adopts random interference and Pedersen Commitment to hide the parameters of the workers’ reputation update, and utilizes the compact linkable spontaneous anonymous group (CLSAG) signature to anonymously verify the reputation update process. Besides, it employs a combination of Bulletproofs and Schnnor signature to verify the reputation feedback process of the workers. Then, the sub-address method is adopted to achieve fast location retrieval when the identities of the workers and requesters are anonymized. Finally, a prototype system is deployed on Hyperledger Fabric to evaluate the performance of the reputation update. The simulation results show that compared with existing reputation privacy protection schemes, BREPP scheme reduces the total reputation management delay and reputation evaluation error rate by at least 2.00% and 24.96%, respectively. Compared with schemes without privacy protection, our scheme achieves a better balance between reputation evaluation accuracy and total delay.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Xiang Ji , Changqiao Xu , Hongke Zhang
    doi: 10.7544/issn1000-1239.202440314

    To address the complex challenges posed by node heterogeneity and connectivity changes in integrated stereoscopic heterogeneous networks, we propose a transmission control method with bounded, loop-free, and blocking-free policy update capabilities. This method incorporates network routing algorithms into the transmission control framework through cross-layer cooperative control. The selection of routing update nodes is modeled as a node search problem, ensuring that the resources involved in network transmission are bounded. On this basis, a breadth-first heuristic incremental search algorithm is designed to efficiently update both global and local routing configurations, ensuring that new paths are loop-free and devoid of network black holes. Additionally, to adapt to the volatility of integrated stereoscopic heterogeneous networks, a multi-mode hybrid congestion control mechanism is designed. This mechanism can switch to a more gradual window-increase mode when approaching the network bandwidth threshold, promptly adjusting policies to handle various potential network conditions. Simulation results demonstrate that HWCTC method provides high-quality data transmission services in dynamic and high packet loss rate integrated space-air-ground heterogeneous network environments. Compared with the classical Cubic and Reno algorithms, HWCTC achieves approximately 61.5% improvement in throughput, significantly enhancing data transmission stability and effectively reducing the impact of dynamic node routing changes on transmission performance.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Guopeng Li , Ruiqi Wu , Haisheng Tan , Guoliang Chen
    doi: 10.7544/issn1000-1239.202440380

    Integrating large language models (LLMs) into personal assistants, like Xiao Ai and Blue Heart V, effectively enhances their ability to interact with humans, solve complex tasks, and manage IoT devices. Such assistants are also termed LLM-driven agents. Upon receiving user requests, the LLM-driven agent generates plans using an LLM, executes these plans through various tools, and then returns the response to the user. During this process, the latency for generating a plan with an LLM can reach tens of seconds, significantly degrading user experience. Real-world dataset analysis shows that about 30% of the requests received by LLM-driven agents are identical or similar, which allows the reuse of previously generated plans to reduce latency. However, it is difficult to accurately define the similarity between the request texts received by the LLM-driven agent through directly evaluating the original request texts. Moreover, the diverse expressions of natural language and the unstructured format of plan texts make implementing plan reuse challenging. To address these issues, we present and implement a plan reuse mechanism for LLM-driven agents called AgentReuse. AgentReuse leverages the similarities and differences among requests’ semantics and uses intent classification to evaluate the similarities between requests and enable the reuse of plans. Experimental results based on a real-world dataset demonstrate that AgentReuse achieves a 93% effective plan reuse rate, an F1 score of 0.9718, and an accuracy of 0.9459 in evaluating request similarities, reducing latency by 93.12% compared with baselines without using the reuse mechanism.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Dongzi Wang , Zhengxin Guo , Linqing Gui , Biyun Sheng , Hui Cai , Fu Xiao
    doi: 10.7544/issn1000-1239.202440386

    Fall detection is crucial in elderly healthcare, traditionally relying on wearable devices for sensing. To avoid the burden and discomfort caused by wearable devices, contact-free methods using radio frequency (RF) signals have emerged as a promising alternative due to their ubiquity and non-invasiveness. Existing contact-free methods predominantly extract physiological motion features, e.g., the speed and acceleration of human falls from RF signals and analyze the motion time series to determine if a fall event has occurred. While current methods achieve high detection accuracy in line-of-sight (LoS) scenarios, they still face limitations in sensing range and non-light-of-sight (NLoS). To address this, we propose a method using the commercial off-the-shelf (COTS) impulse radio-ultra wideband (IR-UWB) devices to detect human falls beyond the sensing range. The key insight of our method is identifying ambient vibration features induced by falls outside the sensing range to determine fall events. By developing the IQ entropy and farthest points-pair distance algorithm, signal features from subtle vibrations caused by falls are extracted. We implement a UWFall prototype system built on COTS IR-UWB devices and conduct extensive experimental evaluations under various environments. The results demonstrate that UWFall system achieves a recognition accuracy over 90% with one single transceiver. Furthermore, the detection accuracy for falls beyond 7 meters exceeds 86%, and this system maintains high robustness in NLoS scenarios.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Siheng Li , Beihong Jin , Fusang Zhang , Zhi Wang , Junqi Ma , Chang Su , Xiaoyong Ren , Haiqin Liu
    doi: 10.7544/issn1000-1239.202440389

    Sleep takes up nearly one-third of a person’s day and is closely related to human health. Since the durations and transitions over different sleep stages during sleep directly affect a person’s sleep quality, identifying sleep stages has become the most basic and important task in sleep monitoring. However, sleep disorders occurring in sleep can lead to complex sleep structures, thereby increasing the difficulty of classifying sleep stages. Most of the existing contactless solutions for sleep stage classification lack a sufficient understanding of the complexity in sleep structure, ignoring the relationship between sleep stage and sleep disorder. Therefore, these solutions fail to achieve great performance in patients with sleep disorders. In this paper, we propose a sleep monitoring system that focuses on predicting sleep stages from ultra-wideband (UWB) signals. We design a sequence prediction model that combines an attention-based sequence encoder and a contrastive learning module to extract the temporal progression of sleep and improve the generalizability of the encoder. Particularly, the key to our approach is a multi-task fine-tuning strategy that incorporates sleep disorder information into sleep staging to reduce the interference of sleep disorders with sleep stage prediction. We conduct extensive experiments on 110 subjects, including healthy individuals and patients with different severities of sleep disorders. The experimental results demonstrate that the performance of our model is superior to the baseline methods.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Xiangyang Li , Fei Shang , Yubo Yan , Shanyue Wang , Feiyu Han , Guoxuan Chi , Zheng Yang , Xiaojiang Chen
    doi: 10.7544/issn1000-1239.202440396

    With the deepening integration of human-machine-object fusion, an increasing number of lightweight and large-scale sensing demands are emerging. To meet the deployment needs of multiple scenarios and large scales, low-cost and low-power sensing solutions are becoming increasingly favored. However, there are still some common and specific challenges in the field of low-power sensing that hinder their further development and practical application. Although many excellent reviews have been conducted on a specific sensing modality or application, there is still a lack of work that sorts out the entire field of low-power sensing. In this paper, we summarize recent low-power sensing, introduce three types of sensing modalities including inertial measurement unit (IMU), microphone, and radio frequency signals, summarize their related challenges, and introduce relevant solutions from hardware and software levels. Finally, we introduce the applications of sensing schemes in different scenarios from four aspects: surface sensing, property sensing, physiological sensing, and anti-sensing, along the direction from surface to inside, from object to human body, and from sensing to safety, and summarize several prospects for exploration.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Xiangyang Li , Shanyue Wang , Chi Zhang , Yubo Yan , Haisheng Tan
    doi: 10.7544/issn1000-1239.202440401

    Passive backscatter communication technology, due to its microwatt-level power consumption, makes IoT devices easy to be deployed and maintenance-free. The multi-tag concurrent communication technology enables tags to avoid complex protocols such as collision avoidance, thus reducing device power consumption while increasing the system throughput and scale. However, the limited spectrum resources available to backscatter tags lead to issues of mutual interference and low spectrum utilization in multi-tag concurrency. To this end, we propose CamScatter, a capacity-enhanced multi-tag parallel backscatter communication system. We design strategies for maximizing channel division and efficient allocation with limited spectrum resources, avoiding interference among tags, and significantly enhancing the system capacity. In the preprocessing phase, a bandwidth-maximized channel division algorithm is proposed, providing multiple interference-free channel division schemes for system communication. In the initialization phase, an optimized channel and rate allocation scheme is proposed, assigning the most suitable channels and rates to tags based on the signal-to-noise ratio (SNR) of all tags in the system, thereby mitigating signal interference caused by tag energy differences and improving system throughput. Additionally, during system working, it uses sideband aggregation and matched filtering techniques to enhance signals strength of the tags to be demodulated, reducing interference from other tags. The bandwidth range that this system utilizes is from 142.4 kHz to 773.5 kHz. Due to harmonic interference limitations, the theoretical maximum spectrum utilization rate is 81.1%, and the maximum system capacity without interference is 2925 kbps. Extensive simulation experiments indicate that the algorithm proposed in this paper can achieve a spectrum utilization rate of 78.6%, with a maximum system capacity of up to 1914 kbps. Actual deployment experiments verify that our system CamScatter supports a maximum throughput of 305 kbps and can handle concurrent communication with up to 30 tags, effectively resolving the issue of mutual interference between tags.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Dongzhu Xu , Anfu Zhou , Huadong Ma , Yuan Zhang
    doi: 10.7544/issn1000-1239.202440403

    Efficient scheduling between cloud-network resources and video tasks is crucial for the performance of video Internet of things (VIoT) applications. However, the current scheduling algorithms used in operational VIoT systems are insufficiently adaptable to differentiated task demands and highly dynamic changes in cloud-network resources, resulting in poor performance of VIoT applications. To overcome the aforementioned problem, we propose a continuous learning-based task demand understanding and scheduling method for VIoT (CLTUS) Unlike traditional heuristic or machine learning-driven scheduling algorithms, CLTUS integrates the continuous learning into the matching between cloud-network resources and video task demands. Specifically, it first employs a continuous learning framework to accurately comprehend various video task demands. Subsequently, based on the dependency relationships among video tasks, it achieves an optimal match between tasks and servers, thereby refining the scheduling of cloud-network resources. Finally, the proposed method is deployed on a software-defined VIoT experimental platform. Compared with conventional methods, CLTUS not only improves the average processing efficiency of video tasks by 127.73% but also increases the balanced utilization rate of cloud-network resources to 67.2% on average, effectively improving the performance of VIoT applications.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Bingyi Liu , Dongdong Wang , Haiyong Shi , Enshu Wang , Libing Wu , Jianping Wang
    doi: 10.7544/issn1000-1239.202440404

    Cellular vehicle-to-everything (C-V2X) communication technology is a crucial component of future intelligent transportation systems (ITS). Millimeter wave (mmWave), as one of the primary carriers for C-V2X technology, offers high bandwidth to users. However, due to limited propagation distance and sensitivity to obstructions, mmWave base stations must be densely deployed to maintain reliable communication. This requirement causes intelligent connected vehicle (ICV) to frequently switch communications during travel, easily leading to local resource shortages, thus degrading service quality and user experience. To address these challenges, we treat each ICV as an agent and model the ICV communication switching issue as a cooperative multi-agent game problem. To solve this problem, we propose a cooperative reinforcement learning framework based on a teammate model. Specifically, we design a teammate model to quantify the interdependencies among agents in complex dynamic environments. Furthermore, we propose a dynamic weight allocation scheme that generates weighted mutual information among teammates for the input of the mixing network, aiming to assist teammates in switching to base stations that provide satisfactory QoS and QoE, thereby achieving high throughput and low communication switching frequency. During the algorithm training process, we design an incentive-compatible training algorithm aimed at aligning the individual goals of the agents with collective goals, enhancing communication throughput. Experimental results demonstrate that this algorithm achieves a 13.8% to 38.2% increase in throughput compared with existing communication benchmark algorithms.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Yong Wang , Yi Xiong , Tianyu Yang , Yiran Shen
    doi: 10.7544/issn1000-1239.202440415

    Earable devices are used as typical AIoT edge sensing devices. Protecting the privacy of legitimate users and preventing illegal use has become extremely important. In response to the current user authentication methods for earable devices, which are limited by input interfaces, sensor costs, and device power consumption, resulting in insufficient security, low universality, and poor user experience, a user authentication model based on the built-in inertial measurement unit (IMU) of earable devices is proposed. This model extracts user-specific information by collecting vibration signals generated by users performing facial interaction gestures, and achieves diversified implicit continuous user authentication based on intelligent analysis of the above information. To extract accurate and reliable user-specific information, a deep neural network feature encoder based on a Siamese network is proposed, which maps gesture samples of the same user closer in the feature space and enlarges the distance between gesture samples of different users, achieving effective encoding of user-specific information. For continuous user authentication based on user-specific information, a weighted voting strategy based on the distance of the one-class support vector machine hyperplane is proposed, which can adaptively optimize the discrimination boundary to better capture the contained features and structures. The confidence level of the sample is determined based on the distance of the sample points inside and outside the hyperplane, and a weighted voting is designed for authentication. Experimental results show that the method in this paper achieves an authentication accuracy of 97.33% in a single vote, and achieves an authentication accuracy of 99.993% after seven rounds of continuous authentication, which is better than all the methods compared in this paper. It provides a smoother user experience and a higher level of security without the need for passwords, and has a high practical application value.

  • Special Issue on Advanced Technologies of Artificial Intelligence of Things
  • Maosheng Zhu , Baohan Wang , Mancong Kang , Wei Yu , Lichao Yang
    doi: 10.7544/issn1000-1239.202440416

    The collection, storage and sharing technology of massive perceptual data has promoted the rise of the Internet of things. Its large-scale application has put forward an urgent requirement for database to have strong data consistency and high resource efficiency. However, the existing database system architecture and management and control methods mostly reduce communication costs by reducing the amount of data transmitted and increasing node storage replicas, and lack of integrated awareness and optimization of system network resources, resulting in low database efficiency. For this reason, we construct an integrated computing network database system to drive the integrated perception representation of computing resources and networks, and enables the database with intelligent Internet of things technology (AIoT) to realize the joint intelligent scheduling of computing resources and networks, so as to reduce the overall cost and improve the efficiency of computing networks. First, we construct a distributed database network integrated with computing and network, and analyze its architecture characteristics. Then, in order to realize the unified representation of computing and communication optimization variables, an integrated cost perception model is constructed. On this basis, an intelligent one-stop resource optimization algorithm is proposed to maximize the overall computing efficiency of the database system. Finally, simulation experiments verify the superiority of the proposed architecture and algorithm in network performance, algorithm convergence, integration cost and resource efficiency.

  • Graphics and Image Processing
  • Xinkai Yan , Yuchi Huo , Hujun Bao
    doi: 10.7544/issn1000-1239.202330483

    Neural rendering is a new image and video generation method based on deep learning. It combines the deep learning model with the physical knowledge of computer graphics to obtain a controllable and realistic scene model, and realize the control of scene attributes such as lighting, camera parameters, posture and so on. On the one hand, neural rendering can not only make full use of the advantages of deep learning to accelerate the traditional forward rendering process, but also provide new solutions for specific tasks such as inverse rendering and 3D reconstruction. On the other hand, the design of innovative hardware structures that adapt to the neural rendering pipeline breaks through the parallel computing and power consumption bottleneck of existing graphics processors, which is expected to provide important support for future key areas such as virtual and augmented reality, film and television creation and digital entertainment, artificial intelligence and metaverse. In this paper, we review the technical connotation, main challenges, and research progress of neural rendering. On this basis, we analyze the common requirements of neural rendering pipeline for hardware acceleration and the characteristics of the current hardware acceleration architecture, and then discuss the design challenges of neural rendering processor architecture. Finally, the future development trend of neural rendering processor architecture is prospected.

  • Graphics and Image Processing
  • Xun Liang , Zhiying Li , Hongxun Jiang
    doi: 10.7544/issn1000-1239.202330077

    The processing, transmission, and semantic segmentation of point clouds are important analytical tasks in the field of 3D computer vision. Nowadays, the effectiveness of graph neural networks and graph structures in point cloud research has been confirmed, and graph-based point cloud (GPC) research continues to emerge. Therefore, a unified research perspective, framework and methodology need to be formed. This paper systematically sorts out various application scenarios of GPC research, including registration, denoising, compression, representation learning, classification, segmentation, detection and other tasks, summarizes the general framework of GPC research, and proposes a technical route covering the current GPC research. Specifically, this paper gives the hierarchical concept category of GPC research, including low-level data processing, intermediate representation learning, and high-level recognition tasks. GPC models or algorithms in various fields are reviewed, including static and dynamic point cloud processing algorithms, supervised and unsupervised representation learning models, and traditional or machine learning GPC recognition algorithms. The representative achievements and their core ideas are summarized, such as dynamically updating the nearest neighbor graph at each layer of feature space, hierarchical and parameter sharing dynamic point aggregation module, and segmentation accuracy improvement employing graph partitioning as well as graph convolution. The model performances are compared, including OA (overall accuracy), mAcc (mean accuracy) and mIoU (mean intersection over union). Based on the analysis and comparison of existing models and methods, the main challenges faced by GPC are summarized, the corresponding research issues are put forward, and the future research directions are explored. The GPC research framework established in this paper is general and comprehensive, which provides field positioning, technical summary and macro perspective for subsequent researchers to engage in this new cross-field research of GPC. The emergence of point cloud research is the result of the rapid progress of detector hardware technology. The current state of point cloud research indicates some challenges between theory and practical applications, and there are still some key issues to be addressed. However, the development of point cloud research is expected to propel artificial intelligence into a new era.

  • Graphics and Image Processing
  • Zetao Jiang , Wencai Zhu , Xin Jin , Peiqi Liao , Jingfan Huang
    doi: 10.7544/issn1000-1239.202330523

    As visual features closer to the text domain, the grid features extracted by the CLIP (contrastive language-image pre-training) image encoder are easy to convert into the corresponding semantic natural language, which can alleviate the semantic gap problem, so it may become an important source of visual features in the image captioning in the future. However, this method does not consider that the division of image content may cause a complete object to be divided into several grids. The segmentation of the objects will inevitably lead to the lack of a complete expression of the object information in the feature extraction results, and further lead to the lack of an accurate expression of the object and the relationship between the objects in the generated sentence. Aiming at the phenomenon of grid features extracted by CLIP image encoder, we propose dual semantic collaborative network (DSC-Net) for image captioning. Specifically, dual semantic collaborative self-attention (DSCS) module is first proposed to enhance the expression of object information by CLIP grid features. Then dual semantic collaborative cross-attention (DSCC) module is proposed to integrate semantic information between grid and object to generate visual features, and to be used to predict sentences. Finally, dual semantic fusion (DSF) module is proposed to provide region-oriented fusion features for the above two semantic cooperation modules, and to solve the problem of correlation conflicts that may arise in the process of semantic cooperation. After a large number of experiments on the COCO dataset, the proposed model achieves a CIDEr score of 138.5% on the offline test set divided by Karpathy et al., and a CIDEr score of 137.6% in the official online test. Compared with the current mainstream image captioning methods, this result has obvious advantages.

  • Artificial Intelligence
  • Chen Li , Yidong Chen , Zhonghua Lu , Xueying Yang , Zitian Wang , Xuebin Chi
    doi: 10.7544/issn1000-1239.202330093

    Multi-objective optimization problems are common in real-world applications. However, the high computational complexity severely hinders their solution. Currently, multi-objective evolutionary algorithms are mostly used for their solving. Nevertheless, these methods usually contain random operations in the process of population initialization and evolution to maintain diversity, which leads to non-reproducibility and lacking the theoretical guarantee of global convergence. Therefore, a novel multi-objective Dividing Rectangles(DIRECT) algorithm is introduced based on decomposition strategy. In this algorithm, the original problem is decomposed into a series of subproblems by a normalized decomposition method that can better capture the complex frontier to reduce the computational complexity of the problem. Meanwhile, the Dividing Rectangles algorithm is used to simultaneously optimize the decomposed subproblems, and the generated candidate solutions are assigned to the corresponding subproblems based on the global association mechanism in the optimization process to better retain the good candidate solutions and improve the algorithm search efficiency. Finally, the convergence of the algorithm is demonstrated. To further improve the computational efficiency, a multi-level and multi-granularity parallel scheme based on the adaptive associative migration strategy is proposed, based on which the proposed algorithm is parallelized. Furthermore, the proposed algorithm and its parallel version are applied to several benchmark optimization problems, and the experimental results show that the proposed algorithm can obtain a Pareto set with better convergence and diversity than NSGA-II, and the parallel version can further improve the accuracy of Pareto frontier approximation while massively reducing the problem-solving time.

  • Artificial Intelligence
  • Xiao Kang , Xingbo Liu , Pengyu Lu , Zhijie Zhao , Xiushan Nie , Shaohua Wang , Yilong Yin
    doi: 10.7544/issn1000-1239.202330433

    Online cross-modal hashing has received increasing attention due to its efficiency and effectiveness in handling cross-modal streaming data retrieval. Despite promising progress, most existing methods rely on accurate and clear supervised information. There are only a few online cross-modal hashing studies concentrating on unsupervised learning mode, and numerous challenges still need to be tackled. For example, streaming data usually suffer from the unbalanced distribution problem due to the limited volume of data in each chunk. Most existing methods neglect this problem, resulting in heightened sensitivity to outlier samples and compromised robustness. Moreover, existing models typically exploit global data distribution, while ignoring local neighborhood information that can promote hash learning. To solve these problems, we propose an unsupervised online cross-modal hashing method with double structure-preserving, called SPOCH(structure preserving online cross-modal hashing). It simultaneously explores the global structure and local structure to generate the corresponding common representation; thereafter, the learned common representation can be used to guide the hash learning process. In terms of global structure-preserving, we design the loss function based on $ {L}_{2,1} $ norm, which can alleviate the sensitivity of outlier samples. In terms of local structure-preserving, we reconstruct sample representation based on neighbor relations that integrates the multi-modality information. In addition, to alleviate the forgetting problem, we propose joint optimization on streaming data, and design the corresponding update strategy to improve the training efficiency. We conduct experiments on two widely-used cross-modal retrieval datasets. Compared with the existing state-of-the-art unsupervised online cross-modal hashing methods, SPOCH achieves superior retrieval accuracy within a comparable or even shorter training time, validating the effectiveness of the proposed approach.