Ryan Giordano

I'm an assistant professor of statistics at UC Berkeley. He / him.

[Ryan Giordano @rgiordan]

So annoying that Brahms Lullaby + Carmen + Hall of the Mountain King type stuff is the stationary point of Spofity autoplay starting from absolutely any classical music

[Ryan Giordano @rgiordan]

@wfithian I've always liked "An Introduction to Probability and Inductive Logic" by Ian Hacking. It's more philosophy than statistics, but maybe that's fine.

[Ryan Giordano @rgiordan]

In statistics we sure make a big deal about data exchangeability. But once we've established it, we /never actually exchange it/.

I'm concerned that people will notice and stop taking us seriously.

[Ryan Giordano @rgiordan]

@ShenRaphael Wolves are really tough reviewers, good luck

[Ryan Giordano @rgiordan]

Is there an easy-to-use and safer alternative to pickle for serializing PyTrees (specifically blackjax samples, but also in general)?

[Daniel Severo @_dsevero]

Best poster award

[Ryan Giordano @rgiordan]

@ShenRaphael Funny, when I searched on Google, in addition to the actual result, one of the top news hits had the headline "'A very disturbing picture': another retraction imminent for controversial physicist."

I see what you're doing there, Google, and I like it.

[Ryan Giordano @rgiordan]

@ShenRaphael @skdeshpande91 @AdvikSh We were out there in the alley confessing to each other our feelings about Dynkin's pi-lambda theorem

(true story)

[Ryan Giordano @rgiordan]

@ProbFact In Stigler's History of Statistics, he makes the opposite case --- that the 18th century's initial focus on the binomial distribution impeded the development of inference since it's not naturally amenable to fiducial arguments. (Ch. 2, esp. page 91 of the 1986 edition)

[Ryan Giordano @rgiordan]

@unsorsodicorda @ml_angelopoulos Great point, thanks!

[Ryan Giordano @rgiordan]

1/n The (great) Distribution-Free, Risk-Controlling Prediction Sets (2021) paper differs from traditional conformal in at least three ways: they use a generic loss, separately control randomness in calibration and test data, and control an "point estimate" expected loss.

[Ryan Giordano @rgiordan]

@ShenRaphael What about the tooth?

[Ryan Giordano @rgiordan]

3/n This result is well known to the papers' authors, and probably won't be surprising to people already very familiar with conformal. But I feel it's worth pointing out --- the paper's methods are even more flexible than it might at first appear.

[Ryan Giordano @rgiordan]

2/n I wrote a short blog post pointing out that these three differences can be easily disentangled. For example, one can guarantee (WHP) that the loss on a particular test point is less than some value, rather than controlling expected loss.

https://t.co/IDEG5xHKKx

[R-nimble @R_nimble]

We're happy to announce a major new version of nimble. It includes Hamiltonian Monte Carlo (package nimbleHMC, configurable with other MCMC samplers), Laplace approximation, and automatic differentiation for algorithm programmers: https://t.co/75t1UQNPUk. https://t.co/kTd3RYLzQy

[Ryan Giordano @rgiordan]

@SomeoneSerge @xenophar Good question --- this is about handling the intractable expectation in the variational objective, not the size of the data.

We assume you can evaluate the log joint distribution on the whole dataset.

[Ryan Giordano @rgiordan]

I have never written a VB paper using SGD --- instead, I have always fixed the draws in advance and optimized the deterministic objective.

Thanks in part to amazing co-first-author @xenophar, there is finally a paper about it!

https://t.co/ktJGijDxuT

[Ryan Giordano @rgiordan]

@renberlinghieri @ta_broderick @brianltrippe @TOzgokmen @renberlinghieri, you've been been doing such stellar work on this exciting project. Scientifically motivated priors FTW!

[Renato Berlinghieri @renberlinghieri]

So exciting to see MIT news featuring our work (accepted at ICML) on how to better predict currents and identify divergences in the ocean from sparse buoy data 🌊
@ta_broderick @brianltrippe @rgiordan @TOzgokmen David Burt, Kaushik Srinivasan & Junfei Xia

https://t.co/hEZpznIZU8

[Ryan Giordano @rgiordan]

Berkeley forms a college of data science.

"Colleges can hire their own faculty, for example, and award degrees to students."

https://t.co/KbHj6rzN5i

[Ryan Giordano @rgiordan]

RIP the mighty Ian Hacking, who completely transformed and rejuvenated my idea of what philosophy can be.