Variational Continual Bayesian Meta-Learning

Authors: Qiang Zhang, Jinyuan Fang, Zaiqiao Meng, Shangsong Liang, Emine Yilmaz

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on tasks from non-stationary distributions show that VC-BML is superior in transferring knowledge among diverse tasks and alleviating catastrophic forgetting in an online setting. Finally, extensive experiments show our VC-BML algorithm outperforms seven state-of-the-art baselines on non-stationary task distributions from four benchmark datasets.
Researcher Affiliation Collaboration Qiang Zhang1,2,3 , Jinyuan Fang4 , Zaiqiao Meng5,6, Shangsong Liang4,6 , Emine Yilmaz7 1 Hangzhou Innovation Center, Zhejiang University, China 2 College of Computer Science and Technology, Zhejiang University, China 3 AZFT Knowledge Engine Lab, China; 4 Sun Yat-sen University, China 5 University of Glasgow, United Kingdom 6 Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates 7 University College London, United Kingdom
Pseudocode Yes Pseudo codes for VC-BML are in Section D of Appendix.
Open Source Code Yes Codes and dataset to reproduce the experiments are included in the supplemental material.
Open Datasets Yes Similar to previous works [4, 25], we conduct experiments on four datasets: Omniglot [35], CIFARFS [36], mini Imagenet [37] and VGG-Flowers [38].
Dataset Splits Yes Similar to the conventional meta-learning setting [14], the dataset Dt is split into a support set DS t = {(xi, yi)}NS t i=1 for training and a query set DQ t = {(xi, yi)}NQ t i=1 for validation. We tune the hyperparameters based on the validation sets.
Hardware Specification Yes The amount and the type of computing resource used in our experiments are described in Appendix. We acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU.
Software Dependencies No The paper does not explicitly list specific software dependencies with their version numbers in the main text or the provided sections. While experimental details are mentioned to be in the Appendix, the main paper doesn't specify versioned software.
Experiment Setup Yes We tune the hyperparameters based on the validation sets. More details can be found in Section E of Appendix.