forked from JuliaTrustworthyAI/LaplaceRedux.jl
-
Notifications
You must be signed in to change notification settings - Fork 0
2023 06 01 Client Meeting Notes
Adelina Cazacu edited this page Jun 1, 2023
·
1 revision
- SB: do we want to stick to the pytorch convention or zygote convention?
- PA: if you can make it work in zygote without having to do tacky things to get it closer to the pytorch version, go for it, zygote convention is to be prefered
- IM: does the order of jacobians matter?
- PA: order is important for being able to compare results with the python library; do it in the way that seems cleaner to you.
- PA: work with simple examples, where you can derive the result by hand first. Full Hessian would be more computationally expensive.
- bayesian optimisation
- optimal subnetwork selection with a heuristic approach
- PA: branch coverage is a good idea, go for it
- ask anthony about branch coverage tools, both locally and on codecov.io
- PA: it is okay to do certain computations differently than what would be mathematically precise, as long as the final result is the correct one, but specify why we are computing them differently and document this