Deep Streaming Regularized Discriminant Analysis

A Pytorch implementation of our Deep Streaming Regularized Discriminant Analysis (Deep SRDA) algorithm from our published paper and preprint.

Deep SRDA is a generative classification method that combines Quadratic Discriminant Analysis (QDA) and Linear Discriminant Analysis (LDA) through a regularizing parameter. Combined with a feature extractor, this method can be used as the final layer of CNN to enable Online Continual Learning with a batch size of 1.

Reproducing our results

To reproduce our results on ImageNet ILSVRC-2012 :

• Download the dataset from this link.
• Use our model with a Resenet18 backbone initialized on the first 100 classes following this repository.

Citing

If you use this code please cite us using:

@inproceedings{khawand2023continual,
  title={Continual Learning with Deep Streaming Regularized Discriminant Analysis},
  author={Khawand, Joe and Hanappe, Peter and Colliaux, David},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3455--3462},
  year={2023}
}