Compact Memory for Continual Logistic Regression

This is an implemenation for "Compact Memory for Continual Logistic Regression", published in NeurIPS 25.

Description for implementation

• example_four_moons.ipynb : results for four-moon task

• setting_dataset.py : task generators

• main_splitcifar100_basereplay_batch.py : baseline experience replay for Split-CIFAR-100

• main_splitcifar100_baselambda_batch.py : baseline K-prior for Split-CIFAR-100

• main_splitcifar100_ourem_batch.py : our method for Split-CIFAR-100

• run_main_splitcifar100.sh : execute experimentsr for Split-CIFAR-100

Once you replace generate_setting_splitcifar100 in each main_**.py using another dataset generator in setting_dataset.py, the code can be run and evaluated on other datasets as well.

Citation

@inproceedings{jung2025compact,
  title={Compact Memory for Continual Logistic Regression},
  author={Jung, Yohan and Lee, Hyungi and Chen, Wenlong and M{\"o}llenhoff, Thomas and Li, Yingzhen and Lee, Juho and Khan, Mohammad Emtiyaz},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2025}
}