EigenGame: PCA as a Nash Equilibrium #9
Description
1. Overview (basic ideas)
- PCA can be seen as a differentiable game.
2. Novelty
- Formulation of PCA as finding the Nash equilibrium of a suitable game
- Decentralized version that is competitive with modern streaming k~PCA
- Demonstration by computing top-32 PCs of the matrix of ResNet200 activations with imagenet.
3. Method (Technical details)
Todo
4. Results
5. links to papers, codes, etc.
- Arxiv : https://arxiv.org/abs/2010.00554
- OpenReiview: https://openreview.net/forum?id=NzTU59SYbNq
- Deepmind: https://deepmind.com/research/publications/EigenGame-PCA-as-a-Nash-Equilibrium
- GoodNotes:
6. Thoughts, Comments
I just thought the title was cool. But it turns out that this formulation could allow PCA to work faster on GPUs/TPUs and so it might be useful for subspace computation.
7. bibtex
@misc{gemp2020eigengame,
title={EigenGame: PCA as a Nash Equilibrium},
author={Ian Gemp and Brian McWilliams and Claire Vernade and Thore Graepel},
year={2020},
eprint={2010.00554},
archivePrefix={arXiv},
primaryClass={cs.LG}
}