ProNC

This is the official Repository for the paper, "Rethinking Continual Learning with Progressive Neural Collapse" (ICLR 2026)

Setup Environment

Create the environment with conda create -n env python=3.10
Install the necessary packages with pip install -r requirements.txt.

Run the Experiments

python utils/main.py --backbone resnet18 --model exorth_ce_svd --dataset seq-cifar100 --n_epochs 50 --lr 0.03 --batch_size 32 --buffer_size 200 --device 0 --num_workers 2 --optim_mom 0. --optim_wd 0.0000 --lr_scheduler multisteplr --lr_milestones 35 45 --main_weight 18 --distill_weight 170 
python utils/main.py --backbone resnet18 --model exorth_ce_svd --dataset seq-tinyimg --n_epochs 100 --lr 0.03 --batch_size 32 --buffer_size 200 --device 0 --num_workers 2 --optim_mom 0. --optim_wd 0.0000 --lr_scheduler multisteplr --lr_milestones 45 75 95 --main_weight 20 --distill_weight 165

Cite Our Work

@inproceedings{wang2026proNC,
  title={Rethinking Continual Learning with Progressive Neural Collapse},
  author={Wang, Zheng and Yu, Wanhao and Yang, Li and Lin, Sen},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}