Feature generation(FG)aims to enhance the prediction potential of original data by constructing high-order feature combinations and removing redundant features.It is a key preprocessing step for tabular scientific data to improve downstream machine-learning model performance.Traditional methods face the following two challenges when dealing with the feature generation of scientific data:First,the effective construction of high-order feature combinations in scientific data necessitates profound and extensive domain-specific expertise.Secondly,as the order of feature combinations increases,the search space expands exponentially,imposing prohibitive human labor consumption.Advancements in the data-centric artificial intelligence(DCAI)paradigm have opened novel avenues for automating feature generation processes.Inspired by that,this paper revisits the conventional feature generation workflow and proposes the multi-agent feature generation(MAFG)framework.Specifically,in the iterative exploration stage,multi-agents will construct mathematical transformation equations collaboratively,synthesize and identify feature combinations exhibiting high information content,and leverage a reinforcement learning mechanism to evolve their strategies.Upon completing the exploration phase,MAFG integrates the large language models(LLMs)to interpretatively evaluate the generated features of each significant model performance breakthrough.Experimental results and case studies consistently demonstrate that MAFG framework effectively automates the feature generation process and significantly enhances various downstream scientific data mining tasks.
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.
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 the 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.
Fault-tolerant deep learning accelerator is the basis for highly reliable deep learning processing and critical to deploy deep learning in safety-critical applications such as avionics and robotics. Since deep learning is known to be computing- and memory-intensive, traditional fault-tolerant approaches based on redundant computing will incur substantial overhead including power consumption and chip area. To this end, we propose to characterize deep learning vulnerability difference across both neurons and bits of each neuron, and leverage the vulnerability difference to enable selective protection of the deep learning processing components from the perspective of architecture layer and circuit layer respectively. At the same time, we observe the correlation between model quantization and bit protection overhead of the underlying processing elements of deep learning accelerators, and propose to reduce the bit protection overhead by adding additional quantization constrain without compromising the model accuracy. Finally, we employ Bayesian optimization strategy to co-optimize the correlated cross-layer design parameters at algorithm layer, architecture layer, and circuit layer to minimize the hardware resource consumption while fulfilling multiple user constraints including reliability, accuracy, and performance of the deep learning processing at the same time.
Today， scientific research has moved from computational science to the era of data science. Discovering laws from massive data and breaking through bottlenecks in scientific development are the main goals of the data science paradigm. At the same time， high-performance computers （HPC） are also paying more and more attention to intelligent computing power. Integrating AI algorithms on the basis of traditional high performance computing methods （HPC+AI） is more conducive to solving practical science problems in the ear of data science， and can give full play to the intelligent computing power of high-performance computers. However， on domestic HPC systems， especially on HPC systems constructed by the new generation of domestic heterogeneous many-core processors， there are many challenges to support HPC+AI programs. In this paper， we propose a data flow computing system for domestic heterogeneous many-core processors， which is called swFLOWpro. The system supports the use of TensorFlow interface to build data flow programs， realizes many-core parallel acceleration transparent to users， and implements two-level parallel strategy based on the whole processor perspective. Tested on sw26010pro processor， swFLOWpro can get up to 545X single core group（CG） many-core speedup ratio for typical OP， 346X for typical deep learning models. Compared with the single CG of sw26010pro， we execute ResNet50 model on all the 6 CGs for one whole processor， and the speedup ration is up to 4.96X， whose parallel efficiency is 82.6%. Experiments show that swFLOWpro can support the efficient execution of data flow programs represented by deep learning on domestic heterogeneous many-core processors.
Computing power network connects computing nodes through the network to break the limitation of single point of computing power, and it is rapidly developed and applied in more and more business fields in recent years.The popular live video broadcasting relies on the transmission and transcoding of a large number of video frames, and it is of great practical importance to explore computing power networks for efficient video distribution. Compared with the traditional large-scale data processing, video applications have higher requirements for transmission delay and bandwidth guarantee. However, the computing power of nodes provided by each cloud service varies, and the state of network links between nodes often varies. Therefore, it is a great challenge to realize low latency and high bandwidth video distribution by selecting nodes with the best combined transmission and transcoding performance.Therefore, we design an efficient video distribution scheme based on heterogeneous computing nodes, including planning video transmission paths and reasonably selecting transcoding nodes through reinforcement learning, using priority queuing scheduling for different video distribution tasks and adaptively adjusting node resources to reduce resource bursty competition, adopting a layered-log-synchronization fault tolerance mechanism to quickly restore data consistency after node failures, and finally deploying multi-cloud service distributed nodes to realize a complete video distribution system. A large number of live ultra-high definition video experiments show that the performance of this scheme is significantly improved compared with existing video distribution methods.
Computing-networking integration takes computing as the center and networking as the foundation,connects heterogeneous computing nodes through the network, and realizes the efficient allocation and scheduling of computing-networking resources. The fairness of resource sharing among competing flows is an important research direction of computing-networking integration. As typical scenarios, video applications are becoming more and more important, but little is known about whether and how much they adhere to the fairness principle. Given the highly diversified network environment and the shortage of automated measurement tools, fairness measurement study entails significant challenges. We investigate this problem by measuring the competing behaviors of typical video application（i.e., Zoom）, and find that, resource competition behaviors are complex and transient, and Zoom has its own selfish behaviors in different operation scenarios. To take a deep dive into these competitive behaviors, we develop automated tools and conduct measurement to understand its QoE（quality of experience）, including end-toend video/voice delay, video frame rate, and video quality. We discover that the strategies of seizing bandwidth are used by Zoom to ensure its own QoE. In the pursuit of better QoE, Zoom tends to selfishly send excessively redundant packets to cope with abnormal network conditions, some of which are not necessary. To this end, it is important to specify a transport algorithm which is able to balance between QoE and fairness goals. We then present the design of QLibra, and demonstrate that it can effectively ensure the QoE and behave harmlessly to competing flows.
生成对抗模仿学习(generative adversarial imitation learning,GAIL)是一种基于生成对抗框架的逆向强化学习(inverse reinforcement learning,IRL)方法,旨在从专家样本中模仿专家策略.在实际任务中,专家样本往往由多模态策略产生.然而,现有的GAIL方法大部分假设专家样本产自于单一模态策略,导致生成对抗模仿学习只能学习到部分模态策略,即出现模式塌缩问题,这极大地限制了模仿学习方法在多模态任务中的应用.针对模式塌缩问题,提出了基于余弦相似度的多模态模仿学习方法(multi-modal imitation learning method with cosine similarity,MCS-GAIL).该方法引入编码器和策略组,通过编码器提取专家样本的模态特征,计算采样样本与专家样本之间特征的余弦相似度,并将其加入策略组的损失函数中,引导策略组学习对应模态的专家策略.此外,MCS-GAIL使用新的极小极大博弈公式指导策略组以互补的方式学习不同模态策略.在假设条件成立的情况下,通过理论分析证明了MCS-GAIL的收敛性.为了验证方法的有效性,将MCS-GAIL用于格子世界和MuJoCo平台上,并与现有模式塌缩方法进行比较.实验结果表明,MCS-GAIL在所有环境中均能有效学习到多个模态策略,且具有较高的准确性和稳定性.
点对学习(pairwise learning)是指损失函数依赖于 2个实例的学习任务.遗憾界对点对学习的泛化分析尤为重要.现有的在线点对学习分析只提供了凸损失函数下的遗憾界.为了弥补非凸损失函数下在线点对学习理论研究的空白,提出了基于稳定性分析的非凸损失函数在线点对学习的遗憾界.首先提出了一个广义的在线点对学习框架,并给出了具有非凸损失函数的在线点对学习的稳定性分析;然后,根据稳定性和遗憾界之间的关系,对非凸损失函数下的遗憾界进行研究;最后证明了当学习者能够获得离线神谕(oracle)时,具有非凸损失函数的广义在线点对学习框架实现了最佳的遗憾界O(T-1/2).
Painting is an important form of culture and art. For thousands of years， a large number of paintings have been produced in ancient China， which contains rich cultural， artistic， scientific， and historical values， But due to natural disasters （earthquake） due to various reasons such as natural weathering and more and more human economic activities， some paintings are more or less damaged or missing in large pieces， which seriously affects the appreciation， cultural creativity， cultural communication and other activities based on these paintings. Compared with natural images， ancient painting images usually have high self-similarity， obvious style characteristics， and rich cultural connotation and delicate texture. Although impressive progress has been made in the inpainting technology of natural images， these methods cannot be directly applied to the inpainting of ancient Chinese paintings. Combined with the characteristics of ancient Chinese paintings， this paper designs the algorithm and model structure and proposes a Chinese ancient painting repair algorithm based on a multi-channel encoder and dual attention module. The goal is to automatically repair the damaged ancient paintings. In order to better repair the ancient paintings from multiple scales， this paper uses a multi-channel encoder to learn the semantic features of ancient paintings at different scales and repairs ancient paintings through the learned macro， meso， and micro semantic features， which solves the difficult problem of rich and delicate texture repair of ancient paintings. In order to better learn the global semantic features of ancient paintings， which is conducive to the harmony and consistency of the repaired ancient paintings， this paper uses the dual attention module to learn the global semantic features of ancient paintings from two aspects: style and content. In order to verify the advanced nature of the algorithm， an ancient painting data set is produced. Experiments on this dataset prove that the algorithm proposed in this paper has better repair quality than the SOTA methods.
边缘计算通过在靠近用户的网络边缘侧部署计算和存储资源,使用户可将高延迟、高耗能应用程序卸载到网络边缘侧执行,从而降低应用延迟和本地能耗.已有的卸载研究通常假设卸载的任务之间相互独立,且边缘服务器缓存有执行任务所需的所有服务.然而,在真实场景中,任务之间往往存在依赖关系,且边缘服务器因其有限的存储资源只能缓存有限的服务.为此,提出一种在边缘服务器计算资源和服务缓存有限的约束下,权衡时延和能耗(即成本)的依赖性任务卸载方法.首先,松弛研究问题中的约束将其转换为凸优化问题;采用凸优化工具求最优解,并用解计算卸载任务的优先级.然后,按照优先级将任务卸载到成本最小的边缘服务器,若多个依赖任务卸载到不同的边缘服务器,为了使总成本最小,则采用改进粒子群算法求解边缘服务器的最佳传输功率.最后,为了验证所提方法的有效性,基于真实数据集进行了充分的实验.实验结果表明,所提方法与其他方法相比能够降低总成本8%～23%.
Community detection aims to uncover the community structure embedded in complex networks， which is one of the important tasks in complex network analysis. However， most of the existing community detection methods are aimed at single-layer network data， and less research has been done on the widespread multi-layer networks in the real world. In order to solve the problem of community detection in multi-layer networks， this paper proposes a two-stage ensemble-based community detection algorithm. The algorithm can improve the accuracy and interpretability of community discovery results. First， after the base communities are obtained at each layer， the local ensemble is performed based on the base community structure information of each layer and the optimal base community structure of each other layers. Secondly， the stability of each community in the local community division of each layer is measured based on information entropy， and the accuracy of each local community division is evaluated through the results of other layer community division. Finally， a global weighted ensemble is carried out based on the importance of each community and the community structure to obtain the final community structure. A comparative analysis is carried out on artificial multi-layer networks and real multi-layer networks with existing multi-layer network community discovery algorithms. The experimental results show that the proposed algorithm in this paper is superior to the existing algorithms in terms of multi-layer modularity and normalized mutual information.
The deep neural network has been widely used in natural language processing. In text generation tasks with multi-domain data， there is often a discrepancy of data in different domains. And the introduction of new domains can simultaneously bring about the problem of data deficiency. The supervised methods require a large amount of data containing ground-truth in the domain of the task to train a deep neural network text generation model， and the trained model cannot achieve good generalization in a new domain. To address the problems of data distribution differences and data deficiency in multi-domain tasks， a comprehensive transfer text generation method inspired by transfer learning methods is designed to reduce the data distribution differences in text data between different domains while leveraging the semantic correlation on text data between source and target domains to help deep neural network text generation models generalize over new domains. The effectiveness of the proposed method for domain transfer is verified through experiments on a publicly available dataset， and the transfer deep neural network text generation model has a better performance in text generation on new domains. Also， it improves in all text generation evaluation metrics compared with other existing transfer text generation methods.
A controller area network （CAN） bus protocol is widely used in the vehicular system and an efficient standard bus enabling communication between all electronic control units （ECUs）. However， the CAN bus is easy to be attacked because of a lack of security defense features. This paper proposes self-attention mechanism （SAM） enhanced grid long short-term memory （Grid LSTM） for vehicular intrusion detection， namely SALVID. The SAM can enhance the characteristics of CAN bus-oriented attack behavior， and the Grid LSTM can effectively extract the depth features of time series data. This paper generates five attack datasets by extracting benign CAN data from the actual car， including denial of service （DoS） ， fuzzy， spoofing， replay， and delete attacks. This paper compares the performance of various models with different model depths， and the results demonstrate that SALVID has the best performance in detecting the attacks on CAN Bus. The model can identify attacks with small-batch features according to an overall detection accuracy of 98.98%， which is hard to do in previous studies. This paper also designs and implements SALVID based on field programmable gate array （FPGA） embedded platform and uses parallel optimization and quantification to accelerate the model based on previous experiments. Even with a certain degree of quantification， SALVID still displays high detection accuracy of 98.81% and a latency of 1.88ms. The investigation provides a new idea for designing high-performance and real-time vehicular intrusion detection systems.
公钥可搜索加密(public key encryption with keyword search,PEKS)技术使用户能够搜索存储在不可信云服务器上的加密数据,这对于数据隐私保护具有重要意义,也因此受到了广泛关注.公钥认证可搜索加密要求数据发送方使用接收方的公钥对关键词进行加密,同时还使用其自身私钥对关键词进行认证,使得敌手无法构造关键词密文,从而抵抗公钥可搜索加密面临的关键词猜测攻击(keyword guessing attack,KGA).提出了一个可证明安全的基于软件防护扩展(software guard extensions,SGX)的公钥认证可搜索加密(public key authenticated encryption with keyword search,PAEKS)方案,通过在云服务器上建立一个可信区并运行一个执行关键词匹配的飞地程序来完成对密文数据的搜索.正式的安全性证明显示方案具备密文不可区分性和陷门不可区分性,即可抵抗关键词猜测攻击.进一步地,给出搜索模式隐私性的定义,确保敌手无法仅通过陷门来判断 2次搜索是否针对同一关键词,从而避免向外部敌手泄露部分隐私.此外,所提方案具有易扩展的优势,很容易被扩展为支持复杂搜索功能或者具备其他增强隐私保护性质的方案,如前向安全.作为示例,给出了多关键词搜索、搜索能力分享这 2个功能扩展方案以及具备前向安全性的扩展方案的简单介绍.真实环境中的实验表明,与其他对比方案相比,所提方案在效率上同样具有出色的表现.
Given a collection of sets， the set similarity self-join （SSSJ） finds all pairs of sets whose similarity is higher than a given threshold， which has been widely used in various applications. The SSSJ adopting the filter-and-verification framework and the parallel and distributed computing framework MapReduce is an active research topic. But existing algorithms produce large candidate sets when the given threshold is low and leads to poor performance. To address this problem， this paper proposes to compute the SSSJ with frequent pattern tree structures and their derivants （FP-tree* for short）， which are used to compress data in memory for computation， targeting at reducing the size of candidate sets. First， based on an investigation on existing FP-tree* and the characteristics of SSSJ computation， we propose a more traversal-efficient linear prefix tree structure， i.e.， TELP-tree， and an original SSSJ algorithm accordingly， i.e.， TELP-SJ， which includes a two-phase filtering strategy: tree-construction oriented filtering algorithm and tree-traversal oriented filtering strategy， which will reduce the size of TELP-trees and tree-traversals. Third， we design the parallel and distributed SSSJ computation algorithm FastTELP-SJ with MapReduce. Finally， 4 groups of empirical comparison studies have been carried out on 4 sets of real application datasets: 1) The comparison between the TELP-SJ algorithm and another 3 FP-tree* based algorithms; 2) The comparison between the FastTELP-SJ algorithm and the TELP-SJ algorithm; 3) The comparison between the FastTELP-SJ algorithm and another 3 existing MapReduce based algorithms. The experimental results demonstrate that FastTELP-SJ achieves better performance in term of the execution time， memory usage， disk usage and scalability for similarity joins over large scale collections of sets with high dimension.
Cross-domain named entity recognition aims to alleviate the problem of insufficient annotation data in the target domain. Most existing methods， which exploit the feature representation or model parameter sharing to achieve cross-domain transfer of entity recognition capabilities and can only partially utilize structured knowledge entailed in text sequences. To address this， we propose a multi-level structured semantic knowledge enhanced cross-domain model （MSKE-CDNER）， which could facilitate the transfer of entity recognition capabilities by aligning the structured knowledge representations embedded in the source and target domains from multiple levels. First， the MSKE-CDNER uses the structural feature representation layer to achieve structured semantic knowledge representations of texts from different fields by structured alignment. And then， these structured semantic representations are aligned at the corresponding layers by a latent alignment module to obtain cross-domain invariant knowledge. Finally， this cross-domain consistent， structured knowledge is fused with domain-specific knowledge to enhance the model’s generalization capability. Experiments on five datasets and a specific cross-domain named entity recognition dataset have shown that the average performance of MSKE-CDNER improved by 0.43% and 1.47% compared with the current models. All of these indicate that exploiting text sequences’ structured semantic knowledge representation could effectively enhance entity recognition in the target domain.
Aspect-term extraction and aspect-level sentiment classification extract aspect-sentiment pairs in the sentence. They help social media platforms such as Twitter and Facebook to mine users’ sentiments of different aspects， which is of great significance to personalized recommendation. In the field of multimodality， the existing methods use two independent models to complete two subtasks respectively. The former identifies goods， important people and other entities or entities’ aspects in the sentence， and the latter predicts the user’s sentiment orientation according to the given aspect terms. There are two problems in the above method: first， using two independent models loses the continuity of the underlying features between the two tasks， and cannot model the potential semantic association of sentences; second， aspect-level sentiment classification can only predict the sentiment of one aspect at a time， which does not match the throughput of aspect-term extraction model that extracts multiple aspects simultaneously， and the serial execution of the two models makes the efficiency of extracting aspect-sentiment pairs low. To solve the above problems， a unified framework for multimodal aspect-term extraction and aspect-level sentiment classification is proposed in this paper. Firstly， the shared feature module is built to realize the latent semantic association modeling between tasks， and it makes the two tasks only need to care about their upper network， which reduces the complexity of the model. Secondly， multiple aspects and their corresponding sentiment categories in the sentence are output at the same time by using sequence tagging， which improves the extraction efficiency of aspect-sentiment pairs. In addition， we introduce part of speech in two tasks at the same time: using the grammatical information to improve the performance of aspect-term extraction， and the information of opinion words is obtained through part of speech to improve the performance of aspect-level sentiment classification. The experimental results show that the unified model has superior performance compared with multiple baseline models on the two benchmark datasets of Twitter2015 and Restaurant2014.
MIBS is a lightweight block cipher which was proposed by Izadi et al. at CANS 2009. Its overall encryption structure uses the typical Feistel network， and the round function adopts the SP network. It supports both MIBS-64 and MIBS-80 versions， that is， it has 64-bit and 80-bit two key lengths with a 64-bit block size， and is suitable for strictly resource-constrained devices， such as low-cost RFID （radio frequency identification） tags. This paper researches the integral attack on the block cipher MIBS. Firstly， we observe the key schedules of MIBS-64 and MIBS-80， and find some properties between their round keys by using the automatic search algorithm for key-bridging technique， respectively. Secondly， using the bit-based division property and the automatic modeling search method based on MILP （mixed integer linear programming）， we find some 8-round and 9-round integral distinguishers of MIBS. Finally， based on the 8-round integral distinguisher， we launch a 12-round key recovery attack for MIBS-64 with the data complexity2 60 ， and the time complexity2 63.42 . Based on the 9-round integral distinguisher， we launch a 14-round key recovery attack for MIBS-80 with the data complexity2 63 ， and the time complexity2 66 . These two key recoveries are the current best integral attacks on the block cipher MIBS-64 and MIBS-80.
In the attribute-based signature （ABS） scheme， the secret key of the signer is generated by attribute authority with different attributes， and the signature can be generated successfully only when the attributes meet the given signing policy. The verifier does not need to know the identity of the signer to determine whether the signature is valid. As a result， ABS has attracted wide attention due to its anonymity and fine-grained access control. In ABS scheme， once the key leakage occurs， the attacker can use the leaked key to generate a valid signature. The original message often contains some sensitive information. For example， in e-health or electronic finance scenarios， personal privacy information is contained in personal medical records or transaction records. If the original message is not desensitized， sensitive personal information will be leaked. In order to solve the problems of key leakage and sensitive information leakage， an efficient and forward-secure attribute-based sanitizable signature （FABSS） scheme is proposed. The security of FABSS is reduced to the η-DHE （η- Diffie-Hellman Exponent assumption） problem under the standard model. The proposed scheme not only protects signer privacy and supports fine-grained access control， but also has the ability to hide sensitive information and resists key leakage. In addition， the length of signature is constant， and only a constant number of pairing operations need to be calculated in the verification stage. Experimental analysis shows that the performance of the proposed scheme is efficient.
