Claas Voelcker
@c_voelcker
Followers
1K
Following
2K
Media
51
Statuses
2K
"All models are wrong, but some are useful" "Do not disfigure the soul" - PhD candidate @UofT, RL researcher unfocused on too many things, he/him, 🏳️🌈
Toronto, Canada
Joined October 2018
I genuinely can’t tell if this is a joke….
New Anthropic research: Adding Error Bars to Evals. AI model evaluations don’t usually include statistics or uncertainty. We think they should. Read the blog post here:
23
21
681
Undergrads should _never_ review. No discussion. Most master students and early stage PhD students should not review.
35
12
314
I have seen so many threads recently bemoaning the working culture in academia, and I fully agree with many things. BUT: y'all know you are academia and it is on you to come together and fix it, right?. I have a radical and simple first proposal: Don't work on the weekend. (1/9).
7
37
293
New grad students: in one of your interviews with the current group graduate students, ask "Have you ever had exceptional circumstances such as a personal loss or a health care problem and how has the PI handled it". The answer to this may be more important than anything else.
5
28
256
Can we stop calling positions in AI/ML "opportunities" and start calling them "rewards for winning the lottery against 10,000 hopefuls"? Feels more honest somehow. .
9
11
198
I am hyped/flustered/excited to share that I will be joining @GoogleDeepMind (formerly Brain Robotics 🤖) as a student researcher 👨🔬! I will work for the inimitable @JonathanTompson and learn all the cool big model things! Thanks for this insane opportunity, I am truly honoured.
14
3
190
Nothing drives you towards religion like running last minute ICLR experiments with research software on a large SLURM cluster.
4
3
164
CW: #iclr2022 . OMG, I can finally say this, for the first time ever!. I am so happy to announce that our paper was accepted to ICLR 2022!!! (Am I allowed to de-anonymize it?).
12
0
133
Today I'm officially remote-beginning my PhD at @UofT with @animesh_garg and @SoloGen. So excited to see what new discoveries we will make, and what great people I will meet! And I hope I will be able to come to Toronto soon too 🇨🇦.
4
3
119
I am proud and humbled to announce that I got my first ever rejection in my PhD. I will take questions after a lengthy Netflix binge.
4
0
108
Only in RL will the unanimous feedback from reviewers be "more experiments" even if you provide 7000 experiments in total in your paper. .
7
1
92
Proposal @icmlconf : we should only have poster sessions and maybe a few orals interspersed. Most people I have talked to feel that the main benefit of in person conferences are spontaneous discussions, large talks seem anachronistic now that we are well set up online.
4
4
91
I am super excited to announce that I'll be joining @animesh_garg and @SoloGen at the @VectorInst / @UofT in September as a PhD student!.
11
5
80
Getting rejected from #NeurIPS is ok, but getting rejected with the ACs comment listing stuff that was thoroughly addressed in the rebuttal is a bit rough. We spend a lot of time on these things, please read them carefully.
6
1
80
Just got common law partnered! (The strange Canadian not-really-but-kind-of-married). Love you @Miterion ❤️🏳️🌈.
11
0
73
This is not about gate keeping. Reviewing is not an honor or opportunity, it’s a service and a duty that has serious weight. It needs preparation and qualification which is essentially impossible to get as an undergrad. Undergrads do not loose anything if they do not review.
2
1
66
The common denominator among (almost) all happy PhD students are a strong in-person lab community (which includes advisors sitting close by) and a present and engaged advisor. Cherry on top is having strong senior lab members (post-docs, candidates) for hands-on mentoring.
I must be a radical outlier. On the whole, I loved grad school. I learned more than any other time in my life. Was taught by brilliant profs. Had endless discussions with other amazing students. Tons of fun times. Made life long friends. Great city. Met my wife. Would do it again.
1
5
62
I am taking a break from shitposting to announce that I have (if the protocol is submitted and I didn’t forget something) become a PhD candidate today. Thanks to my amazing supervisors and committee @SoloGen @igilitschenski and @WilCunningham !.
6
0
58
@apsarathchandar I somehow take more offense at them framing this as "research" than the actual blog post 😂. big "rediscovering trapezoid rule" vibes
2
1
58
#RepresentationLearning can help training strong RL agents on a variety of tasks. But which feature learning method should you pick? State reconstruction (like Dreamer), or latent self prediction (like SPR)?. With @tylerkastnr @igilitschenski @sologen
3
10
55
One additional reason that is rarely promoted: Canada enables spouses and common-law partners of PhD students to get a work permit fairly easily! That could mean the difference between starting/continuing a life together and being long distance for a loooooong time.
🇨🇦⚗️Grad school apps are due soon, so I thought I'd share my article on CS grad school in Canada. Canada is an *amazing* place to do research in CS, with amazing schools like UofT, Waterloo, UBC, SFU, McGill, Alberta, UdeM, among many many others!.
0
8
52
I am on day 3 of the journey to record a 5 minute video for an online conference. Am I uniquely bad at this, or does a pre-recorded video need far more time to prep than a regular talk?.
14
0
52
Tell your department you won't work on the weekend, tell your professors you won't work on the weekend, tell your student you won't work on the weekend, tell your ACs and your reviewers you won't work on the weekend. Demand personal boundaries and respect them (2/9).
1
0
48
Arrived safely in Canada 🇨🇦 Didn't know you could be so happy and sad at the same time 😭😃.
6
0
46
Promise your partner, your kids, your friends, your dog, your piano, your favorite mountain, your art, and your games that your weekend is yours and theirs! You can love your scholarship and something else, that is valid and important. If you don't, you will stop loving. (3/9).
1
0
45
Also, dear America, decide whether PhD students are students, researchers, teachers or community members and help them prioritize these roles! This is beyond ridiculous.
1
0
47
How certain are you that your #DeepRL algorithms actually work?. Maybe you tested them on one of the standard benchmark tasks like the beloved OpenAI Gym Hopper … or maybe you chose the DMC Hopper? Does this choice influence your conclusions?.
4
3
47
The grad application time of the year is upon us and I am so conflicted about everything. I love seeing all the great people advertising, but I also remember how grueling, brutal and disheartening this process was. Take time for yourself and mute Twitter!.
2
1
46
Students are an investment, but not a passive one! |Even good student require a lot of additional time and attention to make them great students, and they deserve that investment from you. Many of them move countries and leave all they know behind to work with you. Honor that!.
New faculty have been asking me for tips on starting out with PhD students, and I thought I'd share some of them here. Be warned, this advice is just what I've picked up from experience, and may or may not translate to your personality or the kind of work you do. 🧵1/?.
2
0
45
I thought I was a harsh reviewer. Now I found out I gave consistently above average scores (measured over each paper) for each of my reviews. Especially on borderline papers, I seem to have more of a "benefit of doubt" attitude? Give people the chance to present! #NeurIPS2022.
1
0
43
Claas (he/him) 🏳️🌈.I am a PhD student working on reinforcement learning, probabilistic inference and robotics 🤖 @UofTRobotics @VectorInst .I love cooking 🍔, reading 📔, playing games on- and offline 🎲 and I just had my relationship notarized to reunite with my boyfriend ❤️🎉🎊.
🏳️🌈🏳️⚧️🎉Happy Pride month! 🎉🏳️🌈🏳️⚧️.Let's do a queer in computer science (and related fields) roll call! 📜Please QT with:. -name/pronouns (if any) 🙋.-research area, specialization, or interests 🔬.-hobbies ⚽️🚵🎮📖.-fun fact ⁉️.
0
2
36
@willie_agnew This reminds me of how the most commonly used word filter library for node.js has a default list of "bad words" which includes "lesbian". This default is used in many systems at setup before people realize this. .
1
4
36
Have you ever wondered why increasing the number of update steps destabilizes #DeepRL training? In our newest paper we show that estimating the values of unseen actions causes instability in continuous state-action domains.
2
8
40
The saddest thing in the discussion to me is that people assume undergrads should have to learn the full skills of a researcher. That is what a PhD is for. Give the kids time and space, please!.Funny thing is, most grad students would love the excuse to get out of reviewing 😅.
Undergrads should _never_ review. No discussion. Most master students and early stage PhD students should not review.
2
0
38
"But my students depend on me!" Take on as many students as you can handle without sacrificing your life, they will thank you. Stop trying to personally fix years of mismanagement and underfunding in teaching roles! Demand more from departments and politicians! (4/9).
1
1
36
Writing a successful paper does not qualify one for reviewing.
1
0
37
If I hear anyone say again that academia needs more than a 40 - 50 hour workweek I am going to flip. Why would you need more than any other job? Do you need to feel better than your peers so badly? And most importantly: 1 hour of good work is worth more than 10 mediocre! (8/9).
1
0
37
By the way, for all you "academia first, everything for the scholarship" people:.Think about how much more we can get done if we don't loose students and professors to mental health threats and burnout! Having only the hardened people stay does not produce good scholarship! (9/9).
1
1
36
Since I am in a rant thread mood today:.Telling people they get to work on their passion is a red flag for exploitation. Telling academics they don't get payed because of the great degrees they will get is the same as paying artists in exposure. Telling PhD workers they (1/5).
1
1
37
"But I won't get everything done!" Then do the rest next week. Prioritize radically in the tasks you agree too. Clearly tell people that you won't be able to do something. Normalize having a life! Being overworked harms you and those around you, esp. students and coauthors (5/9).
1
1
33
It’s been a while since I have been this tired 😴 at the airport, head swirling after one of the most intense weekends ever, I just want to say a huge thank you to the organizers @RL_Conference for putting together an amazing event, and to the whole RL community for filling it.
1
2
35
You know you are too tired when: you are happy that a student got everything right and then realize that you started grading the solutions manual.
1
0
34
If you write an empirical RL paper and you outperform alternate approaches, PLEASE give at least a hypothesis for why!? Parts of the literature are useless because they don't prioritize this insight! Most compare qualitatively to one baseline and then add the others for what?.
2
0
32
@EugeneVinitsky Given the immense competitiveness of a PhD application these days, you are more likely to chose “bad situation” vs “no PhD”. Also, how are you supposed to know if a supervisor is good beforehand?.
4
0
34
"The culture won't let me!" Then demand the culture to change. There is no board on academia that can decide for you. Hire somebody with love and less papers to their name. Ask them what they prioritized and celebrate them. Diversity in passion makes scholarship richer! (6/9).
1
0
31
I have been playing around with environment models and it is fascinating that nn ensembles (PETS, MBPO, MOPO, etc) actually do manage to produce some reasonably looking uncertainty quantification (yes, this is only a toy task ^^) without regularization or (explicit) priors
3
3
31
Tiny RL pet peeve: stop switching off exploration for evaluating RL agents! It makes papers really really hard to compare with each other and skews performance evaluation. And for Christ’s sake, MENTION IT IN THE TEXT!.
6
2
33
@EugeneVinitsky RL struggles as a field because “benchmark” is an ill defined concept and vision was honestly just pretty lucky with imagenet. Almost no other domain has a similar thing, yet all of ML keeps believing in this.
1
0
32
Crab cakes are good, Baltimore is hot and in-person interaction was incredible after two years of online meetings! So grateful for the chance to visit #icml2022 !.
0
1
31
What exactly is the purpose of a #benchmark? To compare algorithms of course! But what does performance on one benchmark tell us about the capabilities of an algorithm elsewhere? In RL, turns out, surprisingly little…. With @marcel_hussing @EricREaton.
1
9
28
I still have the idea of the "no DL" consultancy. You come with your problem and the hype DL solution somebody peddled to you, and we make a working solution without DL. Anybody want to invest/join?.
Am very, very ready for the deep learning hype curve to start dying down. 😅. *So much* business value can be claimed with basic statistics, exploratory data analysis, and traditional machine learning–and those techniques require just a fraction of the time and complexity of DL.
2
2
27
Thanks to all of the friends, teachers and colleagues who made the last 6 years at TU Darmstadt so wonderful. It was a great farewell ceremony! @d120de.
1
0
24
This tweet was brought to you by: "Writing scholarship applications and wondering about the life choices I made".
1
0
27
Question for people working on policy-gradients: if you assume access to a simulator that spits out trillions of samples, shouldn't you also assume (and build in) oracle access to it? It seems strange to me (and maybe only explainable from history) to have one but not the other?.
9
0
28
How have we normalized a world where applying to top NA schools costs you an arm and a leg, when places like the ELLIS program manage to need neither fees nor GRE? (Go apply to the ELLIS program if you want a PhD in great locations without NA fees :D).
6
2
27
Semi-serious question: what do y'all do while waiting on experiments. I have finished mails, admin and am thinking of taking up crocheting :D I could try to read even more, but my brain is coming out my ears.
12
0
27
Ok, more nuanced actual take. Your first n reviews should be done under heavy guidance and with careful feedback from an established member of the community. Basically a “reviewing drivers licence”.
Undergrads should _never_ review. No discussion. Most master students and early stage PhD students should not review.
6
0
25
@pcastr Also, can we automate house cleaning and laundry before we automate the things I actually love like poetry, human interaction, and messing around with dumb Deep RL ideas to see stick figures make funny movements :D.
2
0
26
Overall, if you use the interviews with other grads to discuss only research, you are wasting valuable time that you have to explore lab culture, supervisor style, struggles etc. Save the research talk for the PI interview and maybe a short portion of the grad student interview.
2
1
23
A PhD can be a long time and something unforeseen will happen, so this is in all likelihood not a hypothetical. You might need time off, maybe even on short notice, rearrangement of supervision etc. You want to be prepared for this and trust that the PI is going to be attentive.
1
1
23
More people I was truly impressed by (again) at @RL_Conference : @vdean314 and Erin Talvitie. Two people who radiate kindness, fun, and enthusiasm and who chose a path in academia that puts service in teaching first (and are apparently having a blast with it).
1
1
24
Happy #yalda to all my Iranian friends! One of my favorite unexpected parts of my PhD so far has been to learn so much about your culture, people and country! I hope you (and and everybody) find some hope and joy even in difficult times #WomenLifeFreedom.
0
2
23
@i_Forget_ ML as a whole is pretty terrible in the eval front, these things have been issues since imagenet. But every subcommunity rediscovers these things on their own.
1
1
23
Many things in the world are overrated. I think a total eclipse might actually be underrated. It was an incredible experience, unforgettable, strangely mysterious, truly magical.
1
0
24
And thanks for the student council farewell gift. You will always be the most amazing friends @eine_nana @bhaettasch @realjtl @real_maggi @einBambi @Miterion @tobi_geki @TobioEU @realFDamken @SchrdngrsZombie @Ladifaahri @haaase @atarrust @Xiphoseer @Krullichen @akwickert
1
0
21
Make your code expandable, not only run-able! So often I see codebases with no proper SWE and a README that just gives me the command to rerun the experiments in the paper. But that is not what I want to do when I take another work as a basis. Provide a rudimentary API doc!.
How to get your paper cited?.
1
0
23
@NeurIPSConf If you give each reviewer 6 papers, you have to give us more than one week to adequately address the rebuttal, especially if this is supposed to be a conversation. People can’t just clear out a week of work for reviews!.
1
4
21
I handed in my master thesis! A milestone I sometimes thought completely out if reach. It was an exhausting final run, but for now, I'm done 🎉.Now, hopefully I'll start a PhD soon.
4
1
22
Is anybody I know going to be at @icmlconf next week? Looking to finally turn some e-friends to e+irl-friends! Find me at @QueerinAI workshop and aimlessly wandering the conference!.
1
1
22
I am not telling you to not work hard or have no passion. I am asking you to treat yourself an others like the fully formed human being with countless passions, thoughts, feelings and stressors that you are. But most of all, stop the "culture" and spread a new one (7/9).
1
0
21
Maybe it's a bit misleading to advertise with the nicest building on campus when chemistry is located at the other end of the city in a really horrifying bunker :D.
Check out this open #PhD position in #OrganicChemistry / #MedicinalChemistry at @TUDarmstadt #PhDGermany #Chemistry [image: © TU Darmstadt, Thomas Ott]
1
2
22
Becoming a professor seems a bit like being a person who was trained solely to write beautiful poetry in French, but once you are hired your job is mostly getting funds to buy books for your students, running the school board meetings, and criticizing other people's French :D.
1
1
21
This week I learned what happens if you forget to bid on papers for review at @aistats_conf :D You get the papers no one else wanted to review. In total 1) a really enjoyable, well written paper in a niche data science topic 2) blatant plagiarism and c) I assume a class report?.
3
0
21
This might be an unpopular take: I don't think ML/DS people should attack a COVID-19 dataset without close coordination with medical experts, if the goal is actual scientific (and healthcare) impact. Don't try to solve another discipline with our tools, empower it!.
1
4
20
I rarely say that about LLM research, but I genuinely love how this is set up: instead of looking at models from unquantifiable data, carefully control the data and experiment. Most research should be set up like this in the current era.
Last day of the very special @COLM_conf !! Surprise, surprise, I am actually here to present a poster, than just tweet 😆. Stop by poster #3 this afternoon if you want to learn about training LMs entirely, and from scratch, on knowledge graphs! Why, How and What we learned.
0
2
21
So, we are going to have another straw fire discussion about “too many papers” in which most of the culprits (many who we knew of before) don’t participate. Then we keep giving them awards, funding, speaker invitations, adoration, and worship? Check, got it….
I deanonymized it. Some care needs to be taken to disambiguate names, but I think this is mostly correct up to ties. Levine submitted 22 papers more than half were rejected. A lot of other expected names.
2
0
20
My deep dark secret is out! The real reason I had to leave Germany and come to Canada is that I despise soccer and have no idea how the World Cup is going. .
3
0
19
@EugeneVinitsky No PhD this year overwhelmingly means no PhD ever since continuing publishing without academic support is nigh impossible and research positions on industry don’t exist outside of major US hubs. And we still should not put the blame on the applicants.
1
0
20
People regularly submit papers with 3 seeds and 5 benchmark envs. We provide 10-15 seeds across 20 environments, even reran some baselines cause they made mistakes in their comparisons. If you are an AC in RL @iclr_conf I truly ask you to take a stand on these kinds of requests.
1
0
19
The only purely positive thing about the pandemic is the amount of garlic I can eat guilt free. Three cloves in my dinner, why not :D.
1
0
17
@natolambert Company expos at big conferences feel pretty much just like petty taunting to me these days. Look at all the cool products we have, no we don’t know if we have headcount, why do you ask? I genuinely think we should only invite companies if they commit to hiring.
0
0
18
@pollti @chriswars @ein_flummi @groundop @Guhlius @Nendia1989 @KevinO_Chaos @luelistan @BluemlJ @matedealer @psycon @Treiundzwanzig and all those whose names were not on my cheatsheet. (I should not have tried to name you all, I will have missed someone. You may shame me ;)).
4
0
17
@animesh_garg Now only Twitter needs to go down and we can finally do research without distraction 😅.
0
0
18
So incredibly excited to be a (small) part of such an amazing effort! All the organizers at @QueerinAI truly deserve the recognition at #FAccT2023 and a huge shoutout to the folks driving the paper writing effort specifically!.
1
1
18
I have had this discussion now recurring: when is fixing a random seed actually helpful and not misleading? The only good reason to fix a seed that I can think of is debugging numerical issues. Other than that, it is pretty much always a bad idea imo? Any counterpoints?.
19
2
18
I’m gonna be @NeurIPSConf next week! I’m definitely going to be at the @QueerinAI workshop and I’m on the job market for next year (postdoc/faculty/industry, whatever you offer, I’m up for chatting 😁). Ping me if you are hiring or just want to meet!.
0
1
18
Freedom at last! I'm out of quarantine, and into the frying pan of the Ontario stay-at-home order 😂.
1
0
18
Hype!.
Thrilled to announce the first annual Reinforcement Learning Conference @RL_Conference, which will be held at UMass Amherst August 9-12!.RLC is the first strongly peer-reviewed RL venue with proceedings, and our call for papers is now available:
1
0
17
Our paper on learning object-aware sequence models will be presented as a poster at #iclr2020 🎉.Hope to see you in Addis Abeba!.(if the work looks interesting, at you are looking for PhD candidates interested in model-based rp, contact me 😉).
3
4
17
Hey, @iclr_conf . Standard policy for experimenting would be to ask for consent from participants and explain the setup (e.g. what systems are being used exactly) thoroughly. I don't think we should be legitimizing the use of LLMs in the review process.
1
3
16
@thegautamkamath Because if you do a PhD without “leading��� projects nobody will take you seriously. Genuine collaboration where you aren’t the “leader” is strongly disincentivized. Nobody rewards a good supporter in academia.
2
1
17
@ChristophR1996 I fully agree with this. But undergrads aren't (and shouldn't be) studying to be researchers (yet). Heaping more and more of research education into undergrad is a bad trend for the field. It weakens fundamental education and discourages robust breadth and exploration.
3
0
17
I asked these questions to all people I spoke to at @UofTRobotics and @VectorInst and the answers I got helped me immensely with making a good decision for my PhD.
0
1
17
These colabs between @marcel_hussing and I are the best example why research should not be a solitary effort. Since we have decided to collaborate almost full time 😅 we are on fire 🔥. Find a research buddy who looks at your ideas like @marcel_hussing looks at stable #DeepRL.
🎉Claas and I have another preprint out: MAD-TD🤪. After combatting divergence in high utd RL, we can now stabilize ood action value prediction using model-based synthetic data. This makes resetting networks unnecessary and leads to great performance. Check it out👇
0
0
17
Happy holidays and merry Christmas everyone! Wherever you are, whatever seasonal holiday you celebrate (or not), I hope you have time to relax, safely meet friends, family and loved ones and to leave 2020 behind.
0
0
16
I am currently rereading and I am wondering why this is not being talked about more? In my interpretation, this paper essentially calls into question a lot of our current research by showing that performance increase might be solely due to better hps?.
2
1
14
This is the best take about the NeurIPS high school track. There is nothing wrong at all with making conferences open to high school students. But people love to confuse rewards for a few privileged with widespread support and proper funding of basic education.
@marcel_hussing While I do think its nice to have a venue hosting high school research, but I predict that competition results will be dominated by support inequality, much like science fairs. In high school, I was so frustrated with how extremely unequal science support was. It filled my diary.
1
2
16
@andriy_mulyar I would rather wait than have feedback from somebody who is not prepared for the job. Reviews serve a purpose, they are not just a formal barrier for adding something to our CVs.
1
0
15