Dynamic Sub-graph Distillation for Robust Semi-supervised Continual Learning

Authors: Yan Fan, Yu Wang, Pengfei Zhu, Qinghua Hu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, experiments conducted on three datasets CIFAR10, CIFAR100, and Image Net-100, with varying supervision ratios, demonstrate the effectiveness of our proposed approach in mitigating the catastrophic forgetting problem in semi-supervised continual learning scenarios.
Researcher Affiliation Academia 1Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University, China 2Haihe Laboratory of Information Technology Application Innovation, China fyan 0411@tju.edu.cn, wangyu @tju.edu.cn, zhupengfei@tju.edu.cn, huqinghua@tju.edu.cn
Pseudocode No The paper describes the algorithm using text and equations, but no formal 'Algorithm' or 'Pseudocode' block is provided.
Open Source Code Yes Our code is available: https://github.com/fanyan0411/DSGD.
Open Datasets Yes We validate our method on the widely used benchmark of class continual learning CIFAR10 (Krizhevsky, Hinton et al. 2009), CIFAR100 (Krizhevsky, Hinton et al. 2009) and Image Net-100 (Deng et al. 2009).
Dataset Splits Yes We use a fixed memory size of 2,000 exemplars, assigning 500 samples to labeled data and the remaining 1,500 samples to unlabeled data under sparse annotations. For the semi-supervised setting, we follow ORDis Co to allocate a small number of labels for each class and adhere to the standard experiment setup for selecting the labeled data (Oliver et al. 2018).
Hardware Specification No The paper does not mention any specific GPU, CPU, or other hardware details used for running the experiments.
Software Dependencies No The paper does not mention specific software names with version numbers for reproducibility (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes We use a fixed memory size of 2,000 exemplars, assigning 500 samples to labeled data and the remaining 1,500 samples to unlabeled data under sparse annotations. For the semi-supervised setting, we follow ORDis Co to allocate a small number of labels for each class and adhere to the standard experiment setup for selecting the labeled data (Oliver et al. 2018). ... In Figure 5(a), we conduct experiments with different hyper-parameters γ {0.9, 0.95, 1, 1.5, 2} in dynamic topology graph construction.