|
Michael Figurnov
@
mfigurnov
London, England
|
|
Research Scientist @ Deepmind
|
|
|
162
Tweetovi
|
361
Pratim
|
2.656
Osobe koje vas prate
|
| Tweetovi |
|
Michael Figurnov
@mfigurnov
|
13. pro |
|
[4/4] We hope that this class of estimators will find exciting machine applications! The paper is available online at bayesiandeeplearning.org/2019/papers/76…
|
||
|
|
||
|
Michael Figurnov
@mfigurnov
|
13. pro |
|
[3/4] It has low variance, similar to the reparameterization gradients, and works with non-differentiable functions and discrete distributions, just like REINFORCE. The downside is the higher computational complexity that grows with the number of parameters.
|
||
|
|
||
|
Michael Figurnov
@mfigurnov
|
13. pro |
|
[2/4] Measure valued derivatives are a class of Monte Carlo gradient estimators that has been introduced 30 years ago by Georg Pflug, but is almost unknown in the machine learning community.
|
||
|
|
||
|
Michael Figurnov
@mfigurnov
|
13. pro |
|
[1/4] I will be talking about Measure Valued Derivatives for Approximate Bayesian Inference, our joint work with @elaClaudia @shakir_za @AndriyMnih, at the Bayesian Deep Learning workshop at 16:05 tomorrow.
|
||
|
|
||
|
Michael Figurnov
@mfigurnov
|
10. pro |
|
I’m at #NeurIPS2019 this week. Let me know if you’d like to catch up!
|
||
|
|
||
|
Michael Figurnov
@mfigurnov
|
26. stu |
|
Cool paper from @artygadetsky @k_struminsky: REBAR-like control variates for Plackett-Luce, a distribution over permutations, with application to learning of causal graphs. Check it out! twitter.com/bayesgroup/sta…
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Mihaela Rosca
@MihaelaCRosca
|
22. stu |
|
The code reproducing the experiments is this paper is now available at: github.com/deepmind/mc_gr… twitter.com/shakir_za/stat…
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Oriol Vinyals
@OriolVinyalsML
|
30. lis |
|
#AlphaStar @nature: Grandmaster level as all 3 @StarCraft races on Battle.net, w/ a pro approved interface (camera & APM limits). 2 years ago I thought this was impossible!
How? Imitation learning (Diamond) -> multiagent League (Grandmaster)
deepmind.com/blog/article/A… pic.twitter.com/rcOrRsZ838
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Yaroslav Ganin
@yaroslav_ganin
|
3. lis |
|
A new paper on tweaking SPIRAL (proceedings.mlr.press/v80/ganin18a.h…).
What's new:
• Spectral normalization of discriminator (Miyato, 18) ⇒ sharper images
• Reward shaping by (Ng, 99) ⇒ longer episodes
• In-painting instead of stacking ⇒ better reconstructions
Lots of nice samples :) twitter.com/arkitus/status…
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
David Pfau
@pfau
|
6. ruj |
|
Thrilled to be able to share what I've been working on for the last year - solving the fundamental equations of quantum mechanics with deep learning!
arxiv.org/abs/1909.02487
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
DeepMind
@DeepMind
|
20. kol |
|
We’re excited to release episodes 1 - 4 of the #DMpodcast! Get the inside track on some of the big questions and challenges the field is wrestling with today. No need to be an expert - the amazing @FryRsquared speaks to the people behind the science. deepmind.com/podcast
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Shakir Mohamed
@shakir_za
|
31. srp |
|
Really excited to share our latest paper in @nature today on machine learning for health data to make early predictions of acute kidney injury. Has been an amazing journey over the last 2 years and with an amazing set of people. nature.com/articles/s4158… pic.twitter.com/XhMhdjArCb
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Shakir Mohamed
@shakir_za
|
23. srp |
|
After a short delay, the code in a notebook to reproduce the graphs in section 3 of our paper (arxiv.org/abs/1906.10652) is online. More to be come soon. See thread above👆🏾. github.com/deepmind/mc_gr… 👩🏾💻
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Danilo J. Rezende
@DeepSpiker
|
5. srp |
|
For anyone interested in constrained optimisation with DL models (e.g. as in arxiv.org/abs/1810.00597), we just released a few handy tools to deal with inequality constraints for Sonnet (tiny.cc/a9va9y). Thanks @fabiointheuk !
#Sonnet #ConstrainedOptimisation
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Olaya Álvarez
@oplahoma
|
3. srp |
|
Do you ever feel like a Bayesian distribution? #eeml2019 pic.twitter.com/x4Tdbzq7JU
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Shakir Mohamed
@shakir_za
|
26. lip |
|
Exited to share our new paper: 'Monte Carlo Gradient Estimation in Machine Learning', with @elaClaudia @mfigurnov @AndriyMnih. It reviews of all the things we know about computing gradients of probabilistic functions. arxiv.org/abs/1906.10652 🐾Thread👇🏾 pic.twitter.com/2eTPsFO7mZ
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Bayesian Methods Research Group
@bayesgroup
|
9. svi |
|
Right now @dmolch111, @senya_ashuha, @andrew_atanov and Oleg Ivanov with @mfigurnov present their works at #ICLR2019. Catch then while you can!
- The Deep Weight Prior, #48
- Variance Networks, #72
- VAE with Arbitrary Conditioning, #74 pic.twitter.com/KsNRcwKnlF
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
DeepMind
@DeepMind
|
10. tra |
|
Our new blog post overviews unsupervised learning, a paradigm for creating artificial intelligence that learns about data without a particular task in mind.
Read more about how we might teach computers to learn for the sake of learning:
deepmind.com/blog/unsupervi…
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Bayesian Methods Research Group
@bayesgroup
|
26. ožu |
|
Yesterday @mfigurnov successfully defended his PhD thesis! Congratulations! pic.twitter.com/NqxwTH0vtn
|
||
|
|
||
| Michael Figurnov proslijedio/la je tweet | ||
|
Sander Dieleman
@sedielem
|
13. ožu |
|
Likelihood is a great loss fn, it's all about the space you measure it in! Our latest work on hierarchical AR image models (w/ @JeffreyDeFauw, Karen Simonyan): arxiv.org/abs/1903.04933
We generated 128x128 & 256x256 samples for all ImageNet classes: bit.ly/2FJkvhJ (1/2) pic.twitter.com/4SsaOlqzV6
|
||
|
|
||