MAVI provides a flexible, numerical, and GPU-backed implementation of approximate basis computation of vanishing ideals.
MAVI includes the following methods:
- Numerical methods: Vanishing Component Analysis (VCA; [Livni+, '13]) and its variants Simple Basis Computaition (SBC; [Kera+, '19, '20, '21a]).
- Symbolic methods: Approximate Buchberger–M"oller algorithm (ABM; [Limbeck+, '13]) and its variant [Kera, '21b].
$git clone https://github.com/HiroshiKERA/monomial-agnostic-vanishing-ideal.git
The simplest usage is as follows:
from mavi.vanishing_ideal import VanishingIdeal
# X: data matrix, eps: threshold (positive real)
vi = VanishingIdeal()
vi.fit(X, eps, method='vca', backend='numpy')
vi.plot(X, target='vanishing', spiltshow=True)
Please refer to the Jupyter notebooks in example/
.
To meet the various usages, MAVI is implemented in three different backends numpy
, jax
, and pytorch
. Users are recommended to use jax
because it provides the fastest computation backed by GPU and JIT compilation. pytorch
can also work with GPU but symbolic methods are not implemented yet. If you use jax
backend only, you can uncomment two @partial
decorators at evaluate()
and gradient()
for further accelaration.
If you find this useful, please cite:
@misc{kera2021mavi,
author = {Hiroshi Kera},
title = {{MAVI: Monomial Agnostic Vanishing Ideal}},
url = {https://github.com/HiroshiKERA/monomial-agnostic-vanishing-ideal},
}
@article{kera2021monomial,
title={Monomial-agnostic computation of vanishing ideals},
author={Hiroshi Kera and Yoshihiko Hasegawa},
journal={arXiv preprint arXiv:2101.00243},
year={2021}
}
- [Livni+, '13]: Livni, R., Lehavi, D., Schein, S., Nachliely, H., Shalev-Shwartz, S., Globerson, A., 2013. Vanishing component analysis. In: Proceedings of theThirteenth International Conference on Machine Learning (ICML). PMLR, pp. 597–605.
- [Limbeck+, '13]: Limbeck, J., 2013. Computation of approximate border bases and applications. Ph.D. thesis, Passau, Universit ̈at Passau.
- [Kera+, '19]: Kera, H., Hasegawa, Y., 2019. Spurious vanishing problem in approximate vanishing ideal. IEEE Access 7, pp. 178961–178976.
- [Kera+, '20]: Kera, H., Hasegawa, Y., 2020. Gradient boosts the approximate vanishing ideal. In: Proceedings of the Thirty-Fourth AAAI Conference on ArtificialIntelligence (AAAI). AAAI Press, pp. 4428–4425.
- [Kera+, '21a]: Kera, H., Hasegawa, Y., 2021. Monomial-agnostic computation of vanishing ideals. arXiv preprint arXiv:2101.00243
- [Kera, '21b]: Kera, H., 2021. Approximate border basis with gradient normalization. arXiv preprint arXiv:2101.00401