Here is a list of ideas / directions that I think would be good to start thinking about for v10. Feel free to add / amend.
- Adaptive / nested quadratures
- C1 continuity in piecewise polynomial trees
- Polynomial trees with subspace-based leaf nodes
- Formalising GPs with polynomial kernels
- Spline modelling
- Generalised linear models + theano / numpyro + MCMC
- Low rank tensor decompositions
- Multi-fidelity models
Nice ideas here. One to add to the list might be to explore more general speed-ups (although this depends on the size of datasets we want to support, and typical use-cases of users).
@bubald’s work on using cython/F2Py showed there is some potential there. Although I wonder if sprinkling some
numba decorators throughtout might give us some easier wins…
Seems like this is another reason to work with numpyro for just in time compilation.