Lots of hard work and many late nights went into the making of FSD 12.5, from across the team.
Many ideas were simplified and re-worked from first principles.
Hope everyone has a chance to try it out. It's a release we're proud of.
Last Fall I wrote a blog post collecting my personal notes to understand diffusion models.
To my surprise folks have told me they found it useful, so I'm sharing more widely. (Apologies in advance for any typos or errors).
We applied ML methods and found it can improve discrimination and calibration of personalized treatment effects for intensive blood pressure therapy. With Sanjay Basu,
@pranavrajpurkar
,
@DillonLaird
,
@AndrewYNg
.
Our work Randomized Smoothing of All Shapes and Sizes is at ICML this week. We show how to best choose randomized smoothing distns, esp for L1 adversarial robustness.
Thanks to MSFT collaborators
@TheGregYang
@edwardjhu
@hadisalmanX
@ilyaraz2
@jerryzli
.
Countdown Regression: probabilistic survival forecasting inspired by meteorology, applied to predict time to death from EHR data on 3 million patients! Work with
@anandavati
,
@drnigam
,
@AndrewYNg
.
Great paper comparing MLE vs CRPS for forecasting. Differences arise due to mis-specified data generating process (too light-tailed or heavy-tailed), and diagnosis can be done via calibration curves (W-shape or M-shape PIT).
@dpkingma
@TimSalimans
@poolio
@hojonathanho
Looks like some details in Appendix D of the VDM paper were removed in the update from v5 to v6 on arXiv.
Was this intentional, and if so what was the reason?
(Great paper by the way).
@andrey_kurenkov
@gradientpub
My take: a
@distillpub
"paper" is intended to be cited, whereas a "blog post" often accompanies a paper that's meant to be cited instead of the blog post.