Online Coreset Selection for Rehearsal-based Continual Learning

Authors: Jaehong Yoon, Divyam Madaan, Eunho Yang, Sung Ju Hwang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of our coreset selection mechanism over various standard, imbalanced, and noisy datasets against strong continual learning baselines, demonstrating that it improves task adaptation and prevents catastrophic forgetting in a sample-efficient manner. 5 EXPERIMENTS
Researcher Affiliation Collaboration Jaehong Yoon1 Divyam Madaan2 Eunho Yang1,3 Sung Ju Hwang1,3 KAIST1 New York University2 AITRICS3
Pseudocode Yes Algorithm 1 Online Coreset Selection (OCS)
Open Source Code No The paper does not explicitly state that the source code for the described methodology is publicly available, nor does it provide a link to a code repository.
Open Datasets Yes Datasets. We validate OCS on domain-incremental CL for Balanced and Imbalanced Rotated MNIST using a single-head two-layer MLP... We use the Long-Tailed CIFAR-100 (Cui et al., 2019)... This dataset contains a sequence of five benchmark datasets: MNIST (Le Cun et al., 1998), fashion-MNIST (Xiao et al., 2017), Not MNIST (Bulatov, 2011), Traffic Sign (Stallkamp et al., 2011), and SVHN (Netzer et al., 2011)...
Dataset Splits No The paper describes how data is used within tasks (e.g., 'standard single epoch training'), and mentions 'test accuracy', but does not explicitly provide percentages or counts for distinct training, validation, and test dataset splits in the general sense for all experiments, nor does it explicitly mention a separate validation set for hyperparameter tuning.
Hardware Specification Yes The running time reported in Table 5 was measured on a single NVIDIA TITAN Xp.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes Table A.7 shows the initial learning rate, learning rate decay, and batch size for each dataset that are shared among all the methods.