Picard stands for “Preconditioned ICA for Real Data”. This algorithm quickly solves maximum-likelihood ICA. It is detailed in Faster ICA by preconditioning with Hessian approximations.
Picard-O is an adaptation of Picard which solves the same problem as FastICA, while being much faster on real data. It can also separate super- and sub- Gaussian sources. It is detailed in Faster ICA under orthogonal constraint.
Python and matlab code is available online at https://pierreablin.github.io/picard/.
Qndiag is a very fast algorithm for joint diagonalization of symmetric matrices. Python and matlab code is available at https://pierreablin.github.io/qndiag/
Momentum resnets are a drop-in replacement for residual networks, that consume much less memory. The code is available at https://michaelsdr.github.io/momentumnet/
KSD descent is an algorithm to draw samples from a distribution when only its score is available. The code is at https://pierreablin.github.io/ksddescent/, and it also contain a tested implementation of the celebrated SVGD algorithm.
mvlearn is a general purpose library for machine learning on multiview data: it closely mimics the scikit-learn API, but handles multi-view datasets. It is available at https://mvlearn.github.io/