Continual Learning with Scaled Gradient Projection

Authors: Gobinda Saha, Kaushik Roy

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments ranging from continual image classification to reinforcement learning tasks and report better performance with less training overhead than the state-of-the-art approaches.
Researcher Affiliation Academia Elmore Family School of Electrical and Computer Engineering Purdue University, West Lafayette, Indiana, USA gsaha@purdue.edu, kaushik@purdue.edu
Pseudocode Yes The pseudocode of the algorithm is given in Algorithm 1 in the Appendix B.
Open Source Code Yes Code available at https://github.com/sahagobinda/sgp.
Open Datasets Yes Split CIFAR-100, CIFAR-100 Superclass and Split mini Image Net datasets. ... Krizhevsky 2009 ... Vinyals et al. 2016
Dataset Splits No The paper mentions training for a certain number of epochs with early stopping criteria, implying a validation set is used, but does not explicitly provide the split percentages or counts for a validation set.
Hardware Specification No The paper does not provide specific details on the hardware used, such as GPU or CPU models.
Software Dependencies No The paper mentions algorithms and optimizers like PPO and Adam but does not specify any software dependencies with version numbers.
Experiment Setup Yes each task in Split CIFAR-100 and CIFAR-100 Superclass is trained for 200 and 50 epochs respectively with early stopping criteria with batch size of 64, whereas each Split mini Image Net task is trained for 10 epochs with batch size of 10. We use the same threshold values (ϵth) as GPM and use scale coefficient (α) of 1, 3 and 10 for Split mini Image Net, CIFAR-100 Superclass and Split CIFAR-100 datasets respectively.