MiN: Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025)

September 29, 2025 ยท View on GitHub

Overall

How to reproduce our method?

Data

Six datasets are included in our experiments, i.e., CIFAR-100, CUB-200, ImageNet-A, ImageNet-R, FOOD-101 and Omnibenchmark.

  1. CIFAR-100: it will be automatically downloaded by the code.
  2. ImageNet-A: official: link
  3. Omnibenchmark: Onedrive: link
  4. CUB-200: AWS: link
  5. FOOD-101: AWS: link
  6. ImageNet-R: official: link

Then, revise the data path in data_process/data.py.

Install

The core dependencies are as follows:

  1. torch==2.0.0+cu118 or torch==2.8.0+cu12.8
  2. torchvision==0.15.1+cu118 or torchvision==0.23.0+cu118
  3. timm
  4. scikit-learn
  5. scipy

How to run?

Use the script "run.sh" to reproduce the restlts of 10 steps setting:

"sh run.sh"

Log Files

We hasv tested our code before uploading. The log files can be found in MiN/log. Due to the equipment change, we have to use 5090 to reproduce our results with torch==2.8.0+cu12.8. The result is basically the same as that stated in the paper.

Main results

CIFAR ViT-B/16-IN21K:

CIFAR

CUB ViT-B/16-IN21K:

CUB

IN-A ViT-B/16-IN21K:

IN-A

IN-R ViT-B/16-IN21K:

IN-R

FOOD ViT-B/16-IN21K:

IN-R

Omni. ViT-B/16-IN21K:

IN-R

If you use any content of this repo for your work, please cite the following bib entries:

@inproceedings{MiN,
  title={Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning}, 
  author={Kai Jiang and Zhengyan Shi and Dell Zhang and Hongyuan Zhang and Xuelong Li},
  booktitle={The Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
  year={2025},
  pages={},
}