Very happy to publish this work with Oliver Friedrich (USM/LMU/MCML) and Daniel Gruen (USM/LMU/MCML)!
Simulation-based inference has its own Dodelson-Schneider effect (but it knows that it does)
https://t.co/z5yl3fvOzw
๐งต[1 / n]
@nicolasvelezb Simulation-based inference is a key use case of generative models that have exact likelihoods. The claim is that we can write down the likelihood of the data without a theoretical model built from first principles. It's not that simple however: https://t.co/58dPiAQQt3
@zhaisf Quick question because I plan to use this on a project (it's good to have reliable likelihoods again)
What exactly is the key/value caching doing? Is it simply propagating the k, v caches through the flow between the attention blocks? They're updated per datapoint right? Thanks!
Very happy to publish this work with Oliver Friedrich (USM/LMU/MCML) and Daniel Gruen (USM/LMU/MCML)!
Simulation-based inference has its own Dodelson-Schneider effect (but it knows that it does)
https://t.co/z5yl3fvOzw
๐งต[1 / n]
Here we show SBI posteriors derived with a linear compression (pink, parameterised by a simulation-estimated covariance) and a NN compression (yellow) with a Gaussian likelihood analysis (green).
Note this is only in one universe - the data and widths will scatter!
5
@AugustLamm You see a LOT more in artist's drawings than anything else, so it's hard not to search for them, who else is in the ink cult favourites? Find the Zorn drawing featuring a sea... Unbelievable skill
@AugustLamm Yeah his drawings are my favourite tbh, see Anders Zorn's drawings too, his chaotic hatching is mind blowing! What's the ISBN of this book?