Search

Your search keyword '"Zhao, Xiaoxue"' showing total 22 results
Start Over Author "Zhao, Xiaoxue" Publisher ieee
22 results on '"Zhao, Xiaoxue"'

3. A Novel Two-Stage Knowledge Distillation Framework for Skeleton-Based Action Prediction.

Author

Liu, Cuiwei, Zhao, Xiaoxue, Li, Zhaokui, Yan, Zhuo, and Du, Chong

Subjects
MACHINE learning, PREDICTION models, FORECASTING, DISTRIBUTION (Probability theory), KNOWLEDGE transfer
Abstract

This letter addresses the challenging problem of action prediction with partially observed sequences of skeletons. Towards this goal, we propose a novel two-stage knowledge distillation framework, which transfers prior knowledge to assist the early prediction of ongoing actions. In the first stage, the action prediction model (also referred to as the student) learns from a couple of teachers to adaptively distill action knowledge at different progress levels for partial sequences. Then the learned student acts as a teacher in the next stage, with the objective of optimizing a better action prediction model in a self-training manner. We design an adaptive self-training strategy from the perspective of undermining the supervision from the annotated labels, since this hard supervision is actually too strict for partial sequences without enough discriminative information. Finally, the action prediction models trained in the two stages jointly constitute a two-stream architecture for action prediction. Extensive experiments on the large-scale NTU RGB+D dataset validate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Published
2022
Full Text
View/download PDF

21. A Minimum Resource Neural Network Framework for Solving Multiconstraint Shortest Path Problems.

Author

Zhang, Junying, Zhao, Xiaoxue, and He, Xiaotao

Subjects
*EIGENVECTORS, *NEURAL circuitry, *ARTIFICIAL neural networks, *MULTIPLE correspondence analysis (Statistics), *GAUSSIAN beams
Abstract

Characterized by using minimum hard (structural) and soft (computational) resources, a novel parameter-free minimal resource neural network (MRNN) framework is proposed for solving a wide range of single-source shortest path (SP) problems for various graph types. The problems are the k-shortest time path problems with any combination of three constraints: time, hop, and label constraints, and the graphs can be directed, undirected, or bidirected with symmetric and/or asymmetric traversal time, which can be real and time dependent. Isomorphic to the graph where the SP is to be sought, the network is activated by generating autowave at source neuron and the autowave travels automatically along the paths with the speed of a hop in an iteration. Properties of the network are studied, algorithms are presented, and computation complexity is analyzed. The framework guarantees globally optimal solutions of a series of problems during the iteration process of the network, which provides insight into why even the SP is still too long to be satisfied. The network facilitates very large scale integrated circuit implementation and adapt to very large scale problems due to its massively parallel processing and minimum resource utilization. When implemented in a sequentially processing computer, experiments on synthetic graphs, road maps of cities of the USA, and vehicle routing with time windows indicate that the MRNN is especially efficient for large scale sparse graphs and even dense graphs with some constraints, e.g., the CPU time taken and the iteration number used for the road maps of cities of the USA is even less than \sim2\% and 0.5% that of the Dijkstra's algorithm. [ABSTRACT FROM AUTHOR]

Published
2014
Full Text
View/download PDF

Catalog

Books, media, physical & digital resources
See catalog results