Online Class-Incremental Continual Learning with Adversarial Shapley Value

Authors: Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, Jongseong Jang9630-9638

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

Reproducibility Variable Result LLM Response
Research Type Experimental To test the efficacy of ASER and its variant ASERµ, we evaluate their performance by comparing them with several state-of-the-art CL baselines. We begin by reviewing the benchmark datasets, baselines we compared against and our experiment setting. We then report and analyze the result to validate our approach.
Researcher Affiliation Collaboration 1 University of Toronto, 2 LG AI Research
Pseudocode Yes Algorithm 1: Generic ER-based method
Open Source Code No The paper does not contain any explicit statement about providing open-source code for their methodology or a link to a code repository.
Open Datasets Yes Split CIFAR-10 splits the CIFAR-10 dataset (Krizhevsky 2009), Split CIFAR-100 is constructed by splitting the CIFAR-100 dataset (Krizhevsky 2009), Split mini Imagenet consists of splitting the mini Image Net dataset (Vinyals et al. 2016)
Dataset Splits Yes The detail of datasets, including the general information of each dataset, class composition and the number of sam-ples in training, validation and test sets of each task is presented in Appendix1 D.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types) used for running the experiments.
Software Dependencies No The paper mentions using an 'SGD optimizer' but does not specify any software names with version numbers for dependencies (e.g., specific deep learning frameworks, Python versions, or libraries).
Experiment Setup Yes We use a reduced Res Net18, similar to (Chaudhry et al. 2019b; Lopez-Paz and Ranzato 2017), as the base model for all datasets, and the network is trained via crossentropy loss with SGD optimizer and mini-batch size of 10. The size of the mini-batch retrieved from memory is also set to 10 irrespective of the size of the memory. More details of the experiment can be found in Appendix1E.