A5047932415- Neural Networks and Applications
- Industrial Vision Systems and Defect Detection
- Graph Theory and Algorithms
- Mobile Crowdsensing and Crowdsourcing
- Face and Expression Recognition
- Human Pose and Action Recognition
- Advanced Graph Neural Networks
- Complex Network Analysis Techniques
- Network Packet Processing and Optimization
- Wireless Body Area Networks
- Robotics and Sensor-Based Localization
- Data Management and Algorithms
- Software Reliability and Analysis Research
- Distributed and Parallel Computing Systems
- Welding Techniques and Residual Stresses
- Multimodal Machine Learning Applications
- Energy Efficient Wireless Sensor Networks
- Data Visualization and Analytics
- Brain Tumor Detection and Classification
- Experimental Behavioral Economics Studies
- Autonomous Vehicle Technology and Safety
- Auction Theory and Applications
- Forecasting Techniques and Applications
- Advanced Neural Network Applications
- Imbalanced Data Classification Techniques
Massachusetts Institute of Technology
2023-2024
Nanjing University
2021-2023
North China Electric Power University
2020-2021
East China University of Science and Technology
2013
Sparse convolution plays a pivotal role in emerging workloads, including point cloud processing AR/VR, autonomous driving, and graph understanding recommendation systems. Since the computation pattern is sparse irregular, specialized high-performance kernels are required. Existing GPU libraries offer two dataflow types for convolution. The gather-GEMM-scatter easy to implement but not optimal performance, while dataflows with overlapped memory access (e.g. implicit GEMM) highly performant...
Graph summarization has a wide range of applications that involve many large-scale graph data processing problems. However, most existing approaches for neglect attributes and relationships residing in nodes edges graphs. In this paper, an aggregation-based attributed approach is proposed by fully taking into account during summarization. We present new to quantitatively calculate node merge error, heuristic measure dynamically determine the threshold error per iteration. implemented our...
Intelligent procedure expert system was developed to select appropriate GTAW in this paper. First, the function design and implementation methods of welding were introduced. The can present card, multimedia display process, output makes data sharing more convenient. Then, database based on C/S mode presented where knowledge stored. At last, neural network model established realize selection learning ability case from database. With BPNN model, parameters be obtained input conditions.
Abstrac t—In this paper, an output partitioning algorithm is proposed to improve the performance of neural network (NN) learning. It assumed that negative interaction among attributes may lower training accuracy when we have only one single produce all outputs. Our partitions space into multiple groups according correlation, with strong correlation within each group. After partitioning, group employs a learner train itself. The results from are integrated final result. According our...
To improve the performance of neural network (NN), a new approach based on input space partitioning is introduced, i.e. according to correlation between attributes.As result, effect weak and non-correlation excluded from crucial stage training.After partitioning, CBP introduced train different sub-groups.The results networks are then integrated.According experimental results, improved attained.
This paper presents a new output partitioning approach with the advantages of constructive learning and parallelism.Classification error is used to guide process so that several smaller sub-dimensional data sets are divided from original set.When training each set in parallel, constructively trained sub-network uses whole input vector produces portion final where class represented by one unit.Three classification test validity this algorithm, while results show method feasible.