Machine Learning and Nonparametric Statistics Researcher
Berlin, Germany
Joined December 2018
- Tweets271
- Following202
- Followers142
- Likes1.9K
Robert A. Vandermeulen retweeted
Congratulations: The paper “Set Learning for Accurate and Calibrated Models” is accepted at the #ICLR2024. Authors: @lukas_mut, @robvdm, Q. Zhang, T. Unterthiner, K.-R. Müller.
Preprint: arxiv.org/abs/2307.02245
Wrap up: t1p.de/ijaiw
@GoogleDeepMind @ml_tuberlin
ALT Set Learning for Accurate and Calibrated Models” is accepted at the 12th International Conference on Learning Representations (ICLR) 2024. Authors: Lukas Muttenthaler, Robert A. Vandermeulen, Qiuyi Zhang, Thomas Unterthiner, Klaus-Robert Müller Wrap up: Are you worried about miscalibration of your model outputs but tired of post-hoc temperature scaling and think label smoothing is a hack? Then just train with OKO - a principled set learning approach for more accurate and better calibrated models. OKO is very simple to use: Just change your sampling strategy from standard single data point sampling to set sampling and compute cross-entropy for the sum of set data point logits rather than for single data point logits. Because this forces a model to bring all class logits onto the same scale, your model will be much better calibrated!
Robert A. Vandermeulen retweeted
OKO got accepted to #ICLR2024🎉🥳 OKO is very simple to use: Just change your sampling from single data point sampling to set sampling and compute cross-entropy for the sum of set data point logits rather than for single data point logits. You will get much better calibration! 🔥
🚨 Preprint alert🚨 Are you worried about miscalibration of your model outputs but tired of post-hoc temperature scaling and think label smoothing is a hack? Then just train with OKO - a principled set learning approach for better model calibration! (1/5)
arxiv.org/abs/2307.02245
Robert A. Vandermeulen retweeted
Fun work between @ml_tuberlin @bifoldberlin @GoogleDeepMind 🎆 Thanks to my co-authors @robvdm @TomUnterthiner @richard_qiuyi_z and Klaus-Robert Müller!
Robert A. Vandermeulen retweeted
🔥The CfP for Re-Align is up 🔥
If you are broadly interested in the alignment of two or more representation spaces, consider submitting to our workshop at #ICLR2024 and join us in beautiful Vienna!
🚨Call for Papers🚨
The Re-Align Workshop is coming to #ICLR2024
Our CfP is finally up! Come share your representational alignment work at our interdisciplinary workshop at @iclr_conf: representational-alignment.g…
1/4
Robert A. Vandermeulen retweeted
At the poster for @lukas_mut's paper "Improving neural network representations using human similarity judgments"
xitter.gq/AndrewLampinen/status/…
Wed 13 Dec 10:45 a.m. CST — 12:45 p.m. CST
Great Hall & Hall B1+B2 (level 1) #300
nips.cc/virtual/2023/poster/…
What aspects of human conceptual knowledge do machine learning objectives fail to adequately capture? Could we use small datasets of human similarity judgements to improve model representations and get better downstream performance? @lukas_mut's cool new work has the answers!
Robert A. Vandermeulen retweeted
Freshman: I don't know how to solve this problem. It's too hard!
Tutor: OK, first let's do this tiny step.
Freshman: Yeah, yeah. I know /that/. But that doesn't solve my problem.
Tutor: OK, now let's take the next tiny step.
...
Robert A. Vandermeulen retweeted
Bruh as a professional writer/artist you just sit around wondering if tomorrow everyone is going to stop liking your stuff, and every vacation is unpaid vacation, including weekends.
What you actually want is to be an idle aristocrat living on inherited wealth.
Robert A. Vandermeulen retweeted
It's almost as if the attack on standardized testing is a way for rich white people to defend their children's social position from the kids of hard-charging Asian, Middle Eastern, and African immigrants, rather than an actual way of promoting equity in admissions
Not sure how you can look at this data, ostensibly be interested in either meritocracy or equality, and want to move away from standardized tests. It's the subjective measures that are most slanted in favor of the rich kids.
nytimes.com/interactive/2023…
Robert A. Vandermeulen retweeted
🔊📃
“Set Learning for Accurate and Calibrated Models”
Check out our new preprint together with @GoogleAI, @GoogleDeepMind and @bifoldberlin 👉 arxiv.org/abs/2307.02245
Congratulations to @lukas_mut, @robvdm, @richard_qiuyi_z, @TomUnterthiner!
#MachineLearning
🚨 Preprint alert🚨 Are you worried about miscalibration of your model outputs but tired of post-hoc temperature scaling and think label smoothing is a hack? Then just train with OKO - a principled set learning approach for better model calibration! (1/5)
arxiv.org/abs/2307.02245
Robert A. Vandermeulen retweeted
"Mr. Oppenheimer, what if, instead of bombing Japan, we, instead, tried to love them?"
Robert A. Vandermeulen retweeted
🚨 Preprint alert🚨 Are you worried about miscalibration of your model outputs but tired of post-hoc temperature scaling and think label smoothing is a hack? Then just train with OKO - a principled set learning approach for better model calibration! (1/5)
arxiv.org/abs/2307.02245
Robert A. Vandermeulen retweeted
🚨Beep beep 🚨 Have you ever been wondering about how to use human similarity judgments for improving neural network representations? We have something for you! We found a transform that improves representational alignment and downstream task performance!
arxiv.org/abs/2306.04507
Robert A. Vandermeulen retweeted
🔖Whether or not humans & machines see alike is a long-standing question at the intersection of cognitive science and #AI. BIFOLD researchers tried to pin down the factors that impact the degree of alignment of AI systems with human behavior. 👇bifold.berlin/news-events/ne…
Robert A. Vandermeulen retweeted
Natural disaster deaths were much more common in the past when the world was poorer and medical technology was worse
Robert A. Vandermeulen retweeted
Every day you see some new version of people being weirdly resistant to the idea that what bothers swing voters about the other political party isn't the thing that bothers you, a highly partisan political obsessive.
But like this is just obviously true.
