patrick-kidger / patrick-kidger/diffrax
Why can't UBP be backproped?
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 2.1k
- Forks
- 189
- Avg merge
- 3d 18h
- Merged PRs (30d)
- 1
Description
In the docs it says "You do not need to backpropagate through the differential equation." for UBP usage. However, this doesn't seem to be theoretically necessary, you can just backprop through the solver with the added noise right? What's the motivation requiring this to be the case?
It says "Internally this operates by just sampling a fresh normal random variable over every interval, ignoring the correlation between samples exhibited in true Brownian motion. Hence the restrictions above. (They describe the general case for which the correlation structure isn't needed.)" which makes sense inre adaptivity (since you need a brownian bridge or something of the like), but not for differentiation.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the UBP usage documentation and the passages quoted in the issue, then trace the documented assumptions about solver differentiation and fresh normal samples. Clarify whether backpropagation through the solver is supported and why the stated restriction exists. Done means the documentation directly explains the motivation and any relevant limitations.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- documentation, machine-learning
- Issue type
- Documentation
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100