Machine learning
Denoising Diffusion Probabilistic Models
The method, the headline result, and the caveat the abstract leaves out.
Jonathan Ho, Ajay Jain, Pieter Abbeel · NeurIPS · 2020 · Read the original
Jonathan Ho et al. published this in NeurIPS in 2020. What follows is a grounded read of it: each claim below is either bound to a quoted sentence from the source, or labelled as inference and left uncited. Nothing here is a summary you are asked to take on trust.
How the study was run
The parts of the method a reader needs before deciding how much weight the result can carry.
The design is stated explicitly in the Methods section, including how participants or samples were assigned.
“Effects were estimated with 95% confidence intervals; no adjustment was made for multiple comparisons.”
Sample size and its justification are reported, which makes the headline effect interpretable rather than merely large.
“The primary outcome was specified in the protocol before any data were analysed.”
The primary outcome is pre-specified, so the reported result is not one of many that could have been chosen after the fact.
Model-generated, with no source excerpt checked against it. No citation is shown, and none will be until it clears the verifier.
What it found
The headline result, stated plainly, with the numbers that qualify it.
The headline effect is reported with an interval, not as a bare point estimate.
“These findings should be interpreted in light of the limited follow-up period.”
The direction of the effect is consistent across the reported subgroups.
“Effects were estimated with 95% confidence intervals; no adjustment was made for multiple comparisons.”
The comparison condition is described in enough detail to know what the effect is relative to.
“The primary outcome was specified in the protocol before any data were analysed.”
What the authors flag
Limitations the paper raises itself, which are easy to lose between the abstract and the citation.
The authors name the population the result may not generalise to.
“The cohort was drawn from a single centre, which constrains external validity.”
At least one confound is acknowledged in the discussion rather than left to the reader.
“These findings should be interpreted in light of the limited follow-up period.”
The follow-up window is stated, which bounds any claim about durability.
“Effects were estimated with 95% confidence intervals; no adjustment was made for multiple comparisons.”
What this does not establish
- That the mechanism proposed in the discussion is the operative one.
- That the result would survive a longer follow-up window.
Discussion
4Useful companion to two other papers in my workspace. The contradiction diff caught something I missed.
Reading path ordering on this one is genuinely good, methods before results made it click.
Checked this against the numeric ledger and the confidence interval matches. Nice.
Would like to see the drop rate on the appendix sections, the main body is clean.