Graph-Based Continual Learning

Authors: Binh Tang, David S. Matteson

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on several benchmark datasets show that our model consistently outperforms recently proposed baselines for task-free continual learning.In this section, we evaluate the proposed GCL model on commonly used continual learning benchmarks.
Researcher Affiliation Academia Binh Tang Department of Statistics and Data Science Cornell University Ithaca, NY 14850 bvt5@cornell.edu David S. Matteson Department of Statistics and Data Science Cornell University Ithaca, NY 14850 matteson@cornell.edu
Pseudocode No The paper describes the GCL algorithm mathematically and conceptually, but it does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper references implementations of baselines (EWC, GEM, MER) from their authors' repositories but does not provide a link or explicit statement about the availability of the source code for their proposed GCL model.
Open Datasets Yes We perform experiments on 6 image classification datasets: PERMUTED MNIST, ROTATED MNIST (Le Cun et al., 1998), SPLIT SVHN (Netzer et al., 2011), SPLIT CIFAR10 (Krizhevsky et al., 2009), SPLIT CIFAR100 (Krizhevsky et al., 2009), and SPLIT MINIIMAGENET (Vinyals et al., 2016).
Dataset Splits No The paper describes an online learning setting with training and testing but does not explicitly specify validation splits, percentages, or a distinct validation methodology for the datasets used.
Hardware Specification No The paper describes model architectures and experimental setups, but it does not specify any details about the hardware (e.g., GPU models, CPU types, memory) used to conduct the experiments.
Software Dependencies No The paper mentions optimizers (Adam, SGD) and model types (MLP, convolutional networks), but it does not specify any programming languages, libraries, or frameworks with their version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We perform experiments on 6 image classification datasets... We consider both single-head and multiple-head settings... We use single-head and one-epoch settings for our model and all baselines... In addition, we also report results for multiple-head and 10-epochs settings on SPLIT CIFAR100 and SPLIT MINIIMAGENET... Appendix G: HYPER-PARAMETERS: optimizer: [Adam (Split SVHN, Split CIFAR10), SGD (Permuted MNIST, Rotated MNIST)] learning rate: [0.0002, 0.001 (Split SVHN, Split CIFAR10, Split CIFAR100, Split Mini Imagenet), 0.01, 0.1 (Permuted MNIST, Rotated MNIST), 0.3, 1.0] graph regularization: [0, 10, 50 (Split SVHN, Split CIFAR10, Split CIFAR100, Split Mini Imagenet), 100, 1000, 5000 (Rotated MNIST)] context temperature: [0.1 (Permuted MNIST, Rotated MNIST), 0.3, 1 (Split SVHN, Split CIFAR10, Split CIFAR100, Split Mini Imagenet), 5, 10] target temperature: [0.1, 0.3, 1, 5 (Permuted MNIST, Rotated MNIST, Split SVHN, Split CIFAR10, Split CIFAR100, Split Mini Imagenet), 10]