Wouter Kool

@BlackHC You may also like our paper arxiv.org/abs/2110.01515 where we relate the uniform, exponential and Gumbel parameterizations of the trick. @IamHuijben

Reading up on Gumbel-Softmax distribution (yet again): It's "just" Gumbel-noising the softmax: with a Gumbel sample G ~ Gumbel(0,1) for each class, which can be drawn using a transformation of a uniform sample U ~ U[0,1]: g = -ln(-ln(u)), [or as the -ln of a sample of an exponential variable because e ~ Exp(1) can be written as e = -ln(u)]. If pi_i is itself a softmax distribution of logits l_i, we have: (We can ignore the partition constant of pi because Softmax is invariant to constant shifts.) That is, we can just add Gumbel noise to the logits and temper them. The biggest question maybe is: how important is it to use a Gumbel noise? For example, could we also add Gaussian noise to the logits? I don't think so: The Gumbel-Softmax uses the Gumbel-Max trick to relax the argmax. As \tau -> 0, we approach the argmax which together with the Gumbel noise allows us to sample from \pi_i. A different noise distribution likely wouldn't induce the same distribution (I guess?). Papers: - "Categorical Reparameterization with Gumbel-Softmax" - "The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables" #PaperSummary

The proceedings of the EURO Meets NeurIPS 2022 Vehicle Routing Competition are online: proceedings.mlr.press/v220/kool22a.h…. Thanks again to everyone making the competition a success!

Don't miss the final workshop of the EURO Meets NeurIPS 2022 Vehicle Routing Competition! We'll have great methods presented by the finalists and we'll announce the winners!

@chaitjo Possibly yes! You can find the link here: uva.nl/en/content/eve…

@wouter_kool Wow, congratulations in advance! Is the defence possibly streamed?

In (less than) one week from now (Nov 23, 11:00), I will publicly defend my PhD thesis 'Learning and Optimization in Combinatorial Spaces' at the UvA Aula in Amsterdam! hdl.handle.net/11245.1/b8d128…

After that, I'll be (hopefully) celebrating at NeurIPS so let me know if you're there and want to catch up!

Aaaand it's done! The qualification phase has ended! More than 800 submission were done resulting in almost 50 teams on the leaderboard! Congratulations to the top 10 teams qualifying for the FINALS with their impressive results! :)

Update: 100 teams have registered! 🎉

The competition is going well with 99 teams (!!) registered, 25 already submitted to the leaderboard and over 140 people in the Slack! Thanks all! Who will be participant #100? 🙌 Registration will close September 30th! #leaderboard" target="_blank" rel="nofollow noopener">…neurips-vrp-2022.challenges.ortec.com/#leaderboard

With more than 50 teams registered, 100 participants on Slack, the code submission platform is now available! Check the instructions on #submitting-your-solver" target="_blank" rel="nofollow noopener">github.com/ortec/euro-neu… and secure your place on the public #leaderboard" target="_blank" rel="nofollow noopener">…neurips-vrp-2022.challenges.ortec.com/#leaderboard

Thank you all for your active involvement! Almost 50 teams registered! One week from now (August 1st) the code submission platform will open! Please keep an eye on Twitter, Slack and GitHub to stay updated! …neurips-vrp-2022.challenges.ortec.com

While #EURO2022 is taking place... Already more than 20 teams registered for the EURO Meets #NeurIPS2022 Vehicle Routing Competition! Thanks for joining us in solving this challenging problem! …neurips-vrp-2022.challenges.ortec.com

A few more days until the quickstart will be released! Registration for the EURO Meets NeurIPS 2022 Vehicle Routing Competition is now open! …neurips-vrp-2022.challenges.ortec.com

Excited to announce that we are organizing the EURO Meets NeurIPS 2022 Vehicle Routing Competition! Follow this account to stay updated! Starting July 1st, more info on …neurips-vrp-2022.challenges.ortec.com.

Finally, thanks to @wouter_kool for the most hilarious description of HGS that I have ever seen. I wished I had thought of such a great presentation before! The videos of the workshop should be online at this point, so I'm looking forward to watching it again 😅!

Proud to share that our @ORTEC team Wouter & Co won the @DIMACSChallenge for the VRP with time windows!

The Challenge workshop starts in two weeks! In addition to talks by top competitors and announcement of CVRP & VRPTW winners, there will be keynotes by @MGC_Resende, @SperanzaGrazia, Michel Gendreu, and Renato Werneck. Register: dimacs.rutgers.edu/events/details…

I'm proud to announce that our review paper on the Gumbel-max trick and its extensions for machine learning has been accepted as a survey paper in IEEE TPAMI.🥳 The work covers sampling algorithms, gradient estimators, links to Poisson processes, and fields of applications.

Proud to share that with a team of @ORTEC colleagues we scored 1st place in Phase 1 of the DIMACS challenge for the Vehicle Routing Problem with Time Windows (VRPTW)! @DIMACSChallenge

@hen_str @srush_nlp Thanks! I was thinking about the same! I thought I could show likes, color the dots according to the predicted 👍-lihood and when selecting a region in the plot use this to filter the papers view as well.

#ICLR2020 @srush_nlp @hen_str thanks for building ICLR virtual! I made a page to (privately) like papers and get instant maximum 'like'-lihood recommendations, with some other extra search/browse functionalities! wouterkool.github.io/ilikeclr