Hugh Salimbeni
129 posts
Hugh Salimbeni
@HSalimbeni
PhD student in machine learning at Imperial. Currently thinking about Gaussian processes and their hierarchical extensions
London, England Katılım Mayıs 2017
349 Takip Edilen735 Takipçiler
@xenophar Natural gradient descent gives a significant boost in the Gaussian observation case (minibatch training). I recommend the gpytorch/gpflow demos
English
What are people's favourite ways of fitting an inducing point GP to big Gaussian data these days? I think I have a good overview of the non Gaussian case but not quite as sure about the Gaussian one.
English
@thomaspinder94 @stefanos_ele @markvanderwilk @jameshensman In jax it should be simpler with built-in forward mode, though what you already have is the most useful
English
Nice paper! And an important contribution to interpretable ML. Thoughts, @ChristophMolnar?
James Hensman@jameshensman
Our new ICML paper shows how to build more interpretable Gaussian-Process Generalized Additive Models (GP-GAMs). With @xiaoyulu2022 and @Boukouva1Alexis. Here’s our GP-GAM on the SUSY dataset.
English
@louistiao It is visually quite similar to the great circle method. A difference is that the samples are not constant density in time, and the periodicity is decoupled from how the time dimension is traversed @scien_ti_st @PhilippHennig5
English
@louistiao Nice work! Here's yet another way of getting animated sample via a spatio-temporal GP
English
A Gaussian process is a collection of (infinitely many) random variables such that the marginal distribution over any finite subset is a multivariate Gaussian distribution.
This somewhat arcane description is easier to conceptualize when we focus on a subset of two variables:
English
Hugh Salimbeni retweetledi
Neural Network Gaussian Processes (NNGPs) correspond to wide Bayesian neural networks! In arxiv.org/abs/2006.10541 we show that the posterior distribution over functions computed by a Bayesian neural network converges to the posterior of the NNGP as layer width grows large.
English
Hugh Salimbeni retweetledi
Excited to share our #ICML2020 paper on
*Inter-domain Deep Gaussian Processes*
which I will be presenting at #ICML today:
⏲️Time: July 16, 7am & 6pm AOE
💻Room: icml.cc/virtual/2020/p…
📄Paper: proceedings.icml.cc/static/paper_f…
🕸️Website: sites.google.com/view/inter-dom…
w/ @yaringal @sejDino
GIF
English
@stefanos_ele @PhilippHennig5 @lawrennd @fhuszar @maosbot @carlhenrikek @jameshensman The best thing is that the basis for these jokes is infinite (except for the degenerate jokes)
English
It felt so great to take part in my first G̶a̶u̶s̶s̶i̶a̶n̶ ̶P̶r̶o̶c̶e̶s̶s̶ ̶A̶p̶p̶r̶e̶c̶i̶a̶t̶i̶o̶n̶ ̶C̶e̶r̶e̶m̶o̶n̶y̶ ML@CL Group Meeting with @lawrennd @carlhenrikek and others. These will get boring quickly as my GP jokes get old (the best GP jokes are all from before 2013)
Español
Hugh Salimbeni retweetledi
I will be giving an #ICML2020 tutorial on Machine Learning with Signal Processing. The links between the two are many and old; 4x30min is just for scratching the surface. I'm grateful for the opportunity @icmlconf and for all the help I've received.
icml.cc/Conferences/20…
English
Hugh Salimbeni retweetledi
At #ICML2020: SDEs with Variational Wishart Diffusions.
We introduce a Bayesian framework with Wishart processes, which learns richer noise models in SDEs, both state-dependent and correlations between outputs. Joint work with @mpd37 and @HSalimbeni
arxiv.org/abs/2006.14895
GIF
English
Hugh Salimbeni retweetledi
@mpd37 @andrewgwils @HSalimbeni GPyTorch recently released an example of the Doubly Stochastic DeepGP using PyTorch that was implemented in the above paper.
gpytorch.readthedocs.io/en/latest/exam…
English
Happy to announce that I will be giving a NeurIPS tutorial on “Deep learning with Bayesian principles”. I feel very fortunate to have this opportunity and thank the organizers for their efforts.
Danielle Belgrave@DaniCMBelg
#NeurIPS2019 tutorials are now out! @aliceoh and myself wrote a blog on how we went about selecting this year's tutorials @NeurIPSConf/behind-the-curation-of-the-neurips-2019-tutorials-82e0ebc08c56" target="_blank" rel="nofollow noopener">medium.com/@NeurIPSConf/b…
English
Video of the presentation can be found here: slideslive.com/38916615/gauss…
English
Delighted to share a new paper, Deep Gaussian Processes with Importance-Weighted Variational Inference arxiv.org/abs/1905.05435 appearing at @icmlconf. With @vdutor @jameshensman @mpd37. Code: github.com/hughsalimbeni/…
English
@BurtDavidR presenting at ICML. Work with Carl Rasmussen and @markvanderwilk. Huge congratulations!
English
Hugh Salimbeni retweetledi
Thanks to @icmlconf for the best paper award for our work on "Rates of Convergence for Sparse Variational Gaussian Process Regression"! First author @davidrburt will be giving the invited talk on Thursday 3PM Hall A. We'll be at poster #237 that evening! arxiv.org/abs/1903.03571
Cambridge MLG@CambridgeMLG
Congratulations to our student @BurtDavidR for receiving an ICML 2019 Best Paper Award for “Rates of Convergence for Sparse Variational Gaussian Process Regression”, jointly with Carl E. Rasmussen and @markvanderwilk! (arxiv.org/pdf/1903.03571…) @icmlconf #ICML2019 #bestpaper
English
Hugh Salimbeni retweetledi
We have a cool new manuscript "ODE2VAE: Deep generative second order ODEs with Bayesian neural networks" arxiv.org/abs/1905.10994 with @CgtyYldz @HLahdesmaki where we embed neural ODE's in VAEs with position-velocity decompositions @AaltoCS @FCAI_fi
English