A Python implementation of overcomplete independent subspace analysis.
This code implements an efficient blocked Gibbs sampler for inference and maximum likelihood learning in overcomplete linear models with sparse source distributions. A faster and more memory efficient implementation written in C++ can be found here:
https://github.com/lucastheis/cisa
- Python >= 2.6.0
- NumPy >= 1.6.2
- SciPy >= 0.11.0
I have tested the code with the above versions, but older versions might also work.
L. Theis, J. Sohl-Dickstein, and M. Bethge, Training sparse natural image models with a fast Gibbs sampler of an extended state space, Advances in Neural Information Processing Systems 25, 2012