Retrospective Adversarial Replay for Continual Learning

Authors: Lilly Kumari, Shengjie Wang, Tianyi Zhou, Jeff A Bilmes

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that this excels on broadly-used benchmarks and outperforms other continual learning baselines especially when only a small buffer is available. We conduct a thorough ablation study over each key component as well as a hyperparameter sensitivity analysis to demonstrate the effectiveness and robustness of RAR.
Researcher Affiliation Collaboration Lilly Kumari University of Washington lkumari@uw.edu Shengjie Wang Byte Dance shengjie.wang@bytedance.com Tianyi Zhou University of Maryland zhou@umiacs.umd.edu Jeff Bilmes University of Washington bilmes@uw.edu
Pseudocode Yes Algorithm 1 RAR Retrospective Adversarial Replay for Continual Learning
Open Source Code Yes Our implementation code is available at https://github.com/lillykumari8/RAR-CL.
Open Datasets Yes Datasets: We evaluate RAR on four supervised image classification benchmarks for task-free CL. (1) Split-MNIST[31]... (2) Split-CIFAR10 [28]... (3) Split-CIFAR100... (4) Split-mini Image Net [14]...
Dataset Splits Yes For hyperparameters tuning on each dataset, we hold-out 5% of the training samples for each task and use it as a validation set.
Hardware Specification Yes All experiments are conducted on NVIDIA A6000 GPUs.
Software Dependencies No Our code is implemented in Python using PyTorch and CUDA for GPU acceleration. The paper mentions software names (Python, PyTorch, CUDA) but does not provide specific version numbers for them.
Experiment Setup Yes Settings and Hyperparameters: We follow the same setup as [2] for deciding model architectures for all four datasets. For Split-MNIST, we use an MLP classifier with two hidden layers... For Split CIFAR10, Split CIFAR-100, and Split mini-Image Net, we use a reduced Res Net-18 classifier [22]. The replay budget k is the same as the mini-batch size (fixed to 10) irrespective of the buffer size m. For hyperparameters tuning on each dataset, we hold-out 5% of the training samples for each task and use it as a validation set. We provide additional details about the implementation settings in Appendix F.