Overcoming Catastrophic Forgetting by Bayesian Generative Regularization

Authors: Pei-Hung Chen, Wei Wei, Cho-Jui Hsieh, Bo Dai

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

Reproducibility Variable Result LLM Response
Research Type Experimental we empirically show that the proposed method outperforms state-of-the-art methods on a variety of tasks, avoiding catastrophic forgetting in continual learning.
Researcher Affiliation Collaboration 1Department of Computer Science, UCLA, California, USA 2Google Cloud, Sunnyvale, California, USA 3Google Brain, Mountain View, California, USA
Pseudocode Yes Algorithm 1 Gibbs-Langevin Dynamic Sampling
Open Source Code No The paper refers to code for baseline methods (GDumb: 'https://github.com/drimpossible/GDumb') and analysis tools ('https://github.com/chihkuanyeh/saliency_evaluation') but does not provide a specific link or explicit statement for the open-source code of its own described methodology.
Open Datasets Yes Permuted-MNIST, Split-MNIST, Fashion-MNIST (Xiao et al., 2017), Caltech-UCSD Birds (CUB) dataset.
Dataset Splits Yes Fashion-MNIST... consists of a training set of 60,000 examples and a test set of 10,000 examples. For this task, we follow the Split-MNIST setup to split the classes into sequence: 0/1 (T-shirt/Trouser), 2/3 (Pullover/Dress), 4/5 (Coat/Sandal), 6/7 (Shirt/Sneaker), and 8/9 (Bag/Ankle boot).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup No The paper states: 'Detailed processing of the dataset, implementation of the baseline methods and hyperparameters of the proposed method are described in the supplementary.' This indicates that the specific experimental setup details are not provided in the main text.