Notes from day 1

Here are notes on ideas we have...

Sklearn and Sklearn/dask

SAGA Paper: out of core variant [Olivier, Fabian, Tom]
parallel SGD-like/SAGA implementation [Olivier, Fabian, Tom]
pubsub API
adaptative hyperparameter search [Jim, Fabian, Olivier]
stress infrastructure and expermient [Gael, Olivier, Tom]
Joblib/dask integration: automated data broadcasting/scattering [Olivier, Jim]

scikit-image/dask

nd transforms [Kira, Juan]
benchmarks (useful for nd transforms) [Emma]
numba (precompilation?) [Emma, Juan]
- rewrite Cython
- parallelism
- stencil
- GPU support
Data types
automatically daskify (most?) functions? [Emma]
Big/many images with dask [John]

ASV

scikit-image [Tom]
scikit-learn

General

Matrix balancing on large scale matrices [Nelle]
Matrix factorization "at scale" [John]
- NMF
- online dictionnary learning
- See arthurmensch/modl

Infrastucture

~~figure out how to give access to give 32core access machines to Emma~~ [nelle]