This repository hosts several second order algorithms to solve maximum likelihood ICA.
The algorithms can be found in ml_ica/algorithms
. They are:
- Picard (Preconditioned ICA for Real Data): A preconditioned L-BFGS algorithm. The fastest algorithm in the repository.
- Simple Quasi-Newton method.
- Relative Trust region method.
- Truncated Newton method.
These algorithms have all been rewritten in Python. They call the same gradient, Hessian and likelihood functions, which makes time comparison meaningful.
The algorithms come with a benchmark at ml_ica/example/benchmark.py
. This script runs each algorithm on the same real dataset (fMRI or EEG) and times it.
Clone the repository:
$ git clone https://github.com/pierreablin/faster-ica.git $ cd faster-ica $ pip install -e .
The easiest way to get started is to run the benchmark:
$ python examples/benchmark.py
These are the dependencies to run the algorithms:
- numpy (>=1.8)
- matplotlib (>=1.3)
- numexpr (>= 2.0)
- scipy (>=0.19)
If you use this code in your project, please cite this paper:
Pierre Ablin, Jean-Francois Cardoso, and Alexandre Gramfort Faster independent component analysis by preconditioning with Hessian approximations Arxiv preprint, 2017