Query-based models are extensively used in 3D object detection tasks, with a wide range of pre-trained checkpoints readily available online. However, despite their popularity, these models often require an excessive number of object queries, far surpassing the actual number of objects to detect. The redundant queries result in unnecessary computational and memory costs. In this paper, we find that not all queries contribute equally -- a significant portion of queries have a much smaller impact compared to others. Based on this observation, we propose an embarrassingly simple approach called Gradually Pruning Queries (GPQ), which prunes queries incrementally based on their classification scores. A key advantage of GPQ is that it requires no additional learnable parameters. It is straightforward to implement in any query-based method, as it can be seamlessly integrated as a fine-tuning step using an existing checkpoint after training. With GPQ, users can easily generate multiple models with fewer queries, starting from a checkpoint with an excessive number of queries. Experiments on various advanced 3D detectors show that GPQ effectively reduces redundant queries while maintaining performance. Using our method, model inference on desktop GPUs can be accelerated by up to 1.35x. Moreover, after deployment on edge devices, it achieves up to a 67.86% reduction in FLOPs and a 65.16% decrease in inference time.

Query-based methods with dense features have demonstrated remarkable success in 3D object detection tasks. However, the computational demands of these models, particularly with large image sizes and multiple transformer layers, pose significant challenges for efficient running on edge devices. Existing pruning and distillation methods either need retraining or are designed for ViT models, which are hard to migrate to 3D detectors. To address this issue, we propose a zero-shot runtime pruning method for transformer decoders in 3D object detection models. The method, termed tgGBC (trim keys gradually Guided By Classification scores), systematically trims keys in transformer modules based on their importance. We expand the classification score to multiply it with the attention map to get the importance score of each key and then prune certain keys after each transformer layer according to their importance scores. Our method achieves a 1.99x speedup in the transformer decoder of the latest ToC3D model, with only a minimal performance loss of less than 1%. Interestingly, for certain models, our method even enhances their performance. Moreover, we deploy 3D detectors with tgGBC on an edge device, further validating the effectiveness of our method. The code can be found at https://github.com/iseri27/tg_gbc.

Topic classification is a foundational task in many NLP applications. Traditional topic classifiers often rely on many humandesigned features, while word embedding and convolutional neural network based on deep learning are introduced to realize topic classification in recent years. In this paper, the influence of different word embedding for CNN classifiers is studied, and an improved word embedding named HybridWordVec is proposed, which is a combination of word2vec and topic distribution vector. Experiment on Chinese corpus Fudan set and English corpus 20Newsgroups is conducted. The experiment turns out that CNN with HybridWordVec gains an accuracy of 91.82% for Chinese corpus and 95.67% for English corpus, which suggests HybridWordVec can obviously improve the classification accuracy comparing with other word embedding models like word2vec and GloVe.

In recent years, data stream technology has been widely used in the processing of sensor data. Sonar sensors produce high-speed streaming data with special characteristics of huge volume continuously. Therefore, general data stream management systems with limited memory and processor resources leads to inefficiencies in processing this kind of sonar data stream. However, in distributed data stream management system (DDSMS), continuous query can be reduced into several steps and processed on different processing nodes efficiently and effectively, so DDSMS is often adopted to deal with sonar stream data. Load balancing is a key technology of DDSMS, existing researches mainly focus on using computing load mainly caused by processing queries on the node as the key factor of load balancing algorithm, ignoring the load of transmitting huge volume data mainly caused by moving data among nodes. In this paper, the transmission load of sonar data is computed and used in an improved load balancing algorithm, which is proposed to solve the problem of load scheduling for sonar stream data processing under the distributed environment. The simulation verifies the importance of considering transmission load and shows a better performance of improved load balancing algorithm.

It can improve the performance of mobile data management by caching the frequently accessed and infrequently updated data items. In this paper, a cache replacement policy called Update-based Minimal Access Cost Replacement (UMACR) was proposed to remedy the defects of the existing cache replacement policies such as GDSF and OUR. A variety of factors such as sizes of data items, data access and update information were taken into account in UMACR. To facilitate the replacement policy, two enhanced cache access policies called Update-Server-based Poll-Each-Read (USBPER) and Update-Client-based Call-Back (U2CB) were introduced to guarantee the consistency of data items. USBPER and U2CB remedied the defects of the existing cache access policies such as UPER and UCB by considering updates happening both at the server and clients. Conflict detection and handling was also implemented at the server. At last, we conducted a lot of simulation experiments of mobile data management, and the results demonstrate that the policies we proposed are effective enough, which lay a solid foundation for further research.

As the key point of the ontology research field, concept similarity computation has been widely used in the automatic question-answering systems. This paper firstly analyzes the existing concept similarity computation algorithm which is based on the semantic distance, and then proposes a new computation algorithm formula by modifying and adding some factors which can affect the accuracy of concept similarity computation. Finally, the experimental results show that the proposed algorithm offers better effectiveness for the concept similarity computation.

Skyline query has been gaining much interest in database research communities in recent years. Most existing studies focus mainly on centralized systems, and resolving the problem in a distributed environment such as a peer-to-peer (P2P) network is still an emerging topic. The desiderata of efficient skyline querying in P2P environment include: 1) progressive returning of answers, 2) low processing cost in terms of number of peers accessed and search messages, 3) balanced query loads among the peers. In this paper, we propose a solution that satisfies the three desiderata. Our solution is based on a balanced tree structured P2P network. By partitioning the skyline search space adaptively based on query accessing patterns, we are able to alleviate the problem of "hot" spots present in the skyline query processing. By being able to estimate the peer nodes within the query subspaces, we are able to control the amount of query forwarding, limiting the number of peers involved and the amount of messages transmitted in the network. Load balancing is achieved in query load conscious data space splitting/merging during the joining/departure of nodes and through dynamic load migration. Experiments on real and synthetic datasets confirm the effectiveness and scalability of our algorithm on P2P networks.