Reinforced Cross-Domain Knowledge Distillation on Time Series Data

Authors: QING XU, Min Wu, Xiaoli Li, Kezhi Mao, Zhenghua Chen

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical experimental results and analyses on four public time series datasets demonstrate the effectiveness of our proposed method over other state-of-the-art benchmarks.
Researcher Affiliation Academia Qing Xu Institute for Infocomm Research A*STAR, Singapore Nanyang Technological University Xu_Qing@i2r.a-star.edu.sg Min Wu Institute for Infocomm Research A*STAR, Singapore wumin@i2r.a-star.edu.sg Xiaoli Li Institute for Infocomm Research, A*STAR, Singapore A*STAR Centre for Frontier AI Research, Singapore xlli@i2r.a-star.edu.sg Kezhi Mao Nanyang Technological University EKZMao@ntu.edu.sg Zhenghua Chen Institute for Infocomm Research, A*STAR, Singapore A*STAR Centre for Frontier AI Research, Singapore chen0832@e.ntu.edu.sg
Pseudocode Yes Algorithm 1 Proposed RCD-KD
Open Source Code Yes Our source code is available at https://github.com/xuqing88/Reinforced-Cross-Domain-Knowledge-Distillationon-Time-Series-Data.
Open Datasets Yes To evaluate our method, extensive experiments are conducted on four public datasets across three different tasks, namely human activity recognition, rolling bearing fault diagnosis and sleep stage classification. To be specific, the first dataset is called human activity recognition (HAR) [29]... The second evaluation dataset is Heterogeneity human activity recognition (HHAR) [31]... The third dataset is rolling bearing fault diagnosis (FD) [33]... The last evaluation dataset is sleep stage classification (SSC) dataset [35]...
Dataset Splits No Following standard UDA setup, we consider data from two domains: a labeled source domain DL src = {(xi s, yi s)}ns i=1 and an unlabeled target domain DU tgt = {xi t}nt i=1 which shares the same label space as source domain but has different data distributions.
Hardware Specification Yes Specifically, we measure the training time for our proposed method and other benchmarks with a NVIDIA 2080Ti GPU.
Software Dependencies No The paper describes its implementation details, including network architectures and loss functions, but does not specify versions of programming languages or libraries (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes For the proposed RL-based sample selection module, we set γ = 0.9 and δ = 0.999 following [25] in Eq. (5) and (6), respectively. We set N = 10 to calculate teacher s entropy for the reward function and K = 5 for the episodes to generate historical experience... For α1, α2 in Eq. (3) and λ, τ in Eq. (10), we use the grid search and set α1 = 0.2, α2 = 1.8, λ = 1.0, τ = 2 for all experiments.