David T. Frazier

@TakuoMatsubara @fx_briol Well done @TakuoMatsubara!!

@fx_briol Thanks a lot @fx_briol and everyone! The last four years (with you and everyone) was really an exhilarating journey to me - looking forward to my next chapter in Edinburgh and our collaboration in future!

Congratulations to Dr Takuo Matsubara (@TakuoMatsubara) for passing his PhD viva with flying colours!! Great to have worked closely together over the last four years and looking forward to seeing all his future achievements!

Our paper on improving marginal approximations in likelihood-free inference now online at JCGS. @dtfraz @QUTDataScience tandfonline.com/doi/full/10.10…

@statsgen Chat GPT knows me not! Internet win!

@statsgen Haha. A bit scared to see what the supreme intelligence thinks it me?!?

LOL😂 Your turn @dtfraz

@JulyanArbel @Science_Academy @robjhyndman Thanks, Julyan!

@dtfraz @Science_Academy @robjhyndman Wow, congrats David! 🍾

I just wanted to thank everyone for all the well-wishes I've received over the last week. I especially want to thank @Science_Academy for awarding me the Moran medal, and @robjhyndman for nominating me. Also, a big congratulations to my co-winner Rachel Wang. Thanks again!

@MarvinSchmittML @Science_Academy @robjhyndman Thanks, Marvin!

@dtfraz @Science_Academy @robjhyndman Congrats! 🎉

@AleAviP @Science_Academy @robjhyndman Thanks, Alejandra!

@dtfraz @Science_Academy @robjhyndman Congrats, David! 🎉

@SirioLegramanti @Science_Academy @robjhyndman Thanks, Sirio!

@dtfraz @Science_Academy @robjhyndman Huge congrats!! 👏

@LauchLab @Science_Academy @robjhyndman Thanks, J!

@dtfraz @Science_Academy @robjhyndman Well deserved! Congrats man!

@PierreAlquier @Science_Academy @robjhyndman Thanks, Pierre!

@TresBienHien My brain, my heart, it hurts...

Hmmm 🤔

In short, if your goal is inference, then you probably shouldn't use fixed form VB approximations in SSMs. Even very good ones are inconsistent. However, if your goal is prediction, and you don't have a long time series, then VB is surprisingly accurate.

Who's interested in a cautionary tale of when you should and shouldn't use VB methods in SSMs? If that's the kinda thing you're into, check out our recent paper in #JCGS with @RubenLoaizaMay1 and Gael Martin: tandfonline.com/doi/abs/10.108…

@PierreAlquier @ESSEC_AP @essec Congrats, Pierre!! Best of luck at the new department!!

@TresBienHien @smallscientist_ @TresBienHien"Canadian" geese, which are absolutely everywhere in the US, are legit scary. They are big and chase you with little warning. "Apparently, they can break a man's arm!"

@smallscientist_ Is there a big difference between geese in the US and UK?

1st question: why does this data exist? 2nd question: have these people seen a goose??

Amsterdam is absolutely lovely, how have I never been here before?!?!?

@PierreAlquier Kingfishers have to be my favourite bird. So delicate looking, yet such adept hunters!

@SeBayesian Nice to see that this problem is finally getting the recognition it deserves. I'll send you through any comments I have!!

What's the takeaway? SBI methods can perform very well when real data looks like simulated data. If not there is a danger of wild inaccuracy. Future work should look for methods to 1) identify and 2) counter misspecification.

To make simulation-based inference methods viable for imperfect simulators (simulators that only partially explain real outcomes) we need new and robust methods. We experimented with popular SBI approaches using coverage as a metric and ..oh boy 🧵

@pw_cannon @SeBayesian Yes, the answer is yes. The extrapolation problem is alive and well!

Is model misspecification the biggest challenge for modern SBI algorithms? In our new work, we show that many neural SBI techniques can fail dramatically when simulators are even slightly misspecified. For details, see arxiv.org/abs/2209.01845