Yale Chang

machine learning for healthcare. research scientist @ Philips clinical research

[Adarsh Subbaswamy @_asubbaswamy]

New preprint w/ (co-first author) @royjamesadams and @suchisaria: "Evaluating Models Robustness Under Dataset Shift" https://t.co/ykOu7G9EQE

How can we evaluate *ahead of time* whether or not a model's performance will generalize from training to deployment? 1/

[Tengyu Ma @tengyuma]

Some features can spuriously correlate with labels but do not cause them. Models relying on these features are biased and non-robust to varying correlations. We show self-training on an *unlabeled*, more diverse dataset can avoid using spurious features. https://t.co/4ryC4Leegs

[Ofir Nachum @ofirnachum]

Policy evaluation via duality/Lagrangian methods presents a lot of choices (how to setup the LPs, regularize them, etc). In https://t.co/Ics1296x4U we examine how these choices affect accuracy of final eval. Lots of insights in this paper, many of which I didn't expect....

[Sergey Levine @svlevine]

In the real world, the test distribution never actually matches the training distribution. Adaptive risk minimization (ARM) addresses distributional shift by adapting to it, without labels -- just from seeing a group of test inputs rather than a single individual test input.

[Emtiyaz Khan @EmtiyazKhan]

Must read blog by @beenwrekt. Robust/reliable ML is extremely important and should be mainstream. Can’t agree more.

https://t.co/F4n3wwjniL

[Max Welling @wellingmax]

A simple framework that unifies many well known MCMC algorithms. If you read this you can clean up your brain to make place for new stuff (useful for when you are over 50).

[Ben Recht @beenwrekt]

Some reflections on the learning-to-control renaissance. https://t.co/zQ4L5lUtCV

[Andrew Gordon Wilson @andrewgwils]

The inductive biases of normalizing flows can be more of a curse than a blessing. Our new paper, "Why Normalizing Flows Fail to Detect Out-of-Distribution Data", with @polkirichenko and @Pavel_Izmailov: https://t.co/rIQ1Pv14Se

[Sergey Levine @svlevine]

We (Aviral Kumar, A. Zhou, @georgejtucker) released conservative Q-learning (CQL). CQL is an offline RL algorithm, and it works very well. Much better than I thought offline RL could work, on many tasks (see below). [1/7]

https://t.co/Y7gZfbU7DT
https://t.co/u0eVrncvAF

[Sergey Levine @svlevine]

To learn more about offline RL, check out our tutorial: https://t.co/kv4mtVMScG

And if you want to try it out, check out our D4RL offline RL benchmarks: https://t.co/hu268l8wPL

[Anshul Kundaje @anshulkundaje]

Congrats to Amr and @avshrikumar. Their paper on 'EM with bias-corrected calibration for label shift adaptation' is accepted at #icml2020 . Important in many prediction tasks where baseline class prevalence is different between training and test scenarios
https://t.co/icJIPV0uIC

[Chelsea Finn @chelseabfinn]

Offline RL may make it possible to learn behavior from large, diverse datasets (like the rest of ML). We introduce:

MOPO: Model-based Offline Policy Optimization
https://t.co/6MlXg4Y5qc

w/ Tianhe Yu, Garrett Thomas, Lantao Yu @StefanoErmon @james_y_zou @svlevine @tengyuma

[Francisco Massa @fvsmassa]

I'm super excited to share our work on End-to-End Object Detection with Transformers. @PyTorch code, pre-trained models and colab notebooks available at https://t.co/hhNsTphkHr, check it out!

[Sasha Rush @srush_nlp]

Introducing🤗nlp viewer: https://t.co/UnPYaZDPW7

A simple website for browsing SQuAD, MNLI, IMDB, SST, Yelp, drop, CommonsenseQA, billsum, CoQA, BoolQ, Jeopardy, RACE, SCAN, Super Glue, qangaroo, winogrande, CFQ, BLIMP, MLQA, opinosis, hellaswag, ...

cc @huggingface @Thom_Wolf

[Dustin Tran @dustinvtran]

Excited to release rank-1 Bayesian neural nets, achieving new SOTA on uncertainty & robustness across ImageNet, CIFAR-10/100, and MIMIC. We do extensive ablations to disentangle BNN choices.@dusenberrymw @Ghassen_ML @JasperSnoek @kat_heller @balajiln et al https://t.co/aMfBvVkl0v

[Graham Neubig @gneubig]

I've finished uploading the lecture videos for CMU CS11-747 "Neural Networks for NLP"'s 2020 edition: https://t.co/H1jHhqwTWz

Check it out if you're interested in a comprehensive graduate-level course on modern NLP methods!

[Edward Grefenstette @egrefen]

Love this. Similar result (different domain) to Melis et al 2017: https://t.co/gHArwPVQ9X.
That paper was pearls before swine for the EMNLP reviewers that read it. Also echos stuff that @Smerity has observed, I believe.

[Eric Jang @ericjang11]

Every once in awhile a paper comes out that makes you breathe a sigh of relief that you don't publish in that field...

https://t.co/56heAufhGA

"Our results show that when hyperparameters are properly tuned via cross-validation, most methods perform similarly to one another"

[Kamalika Chaudhuri @kamalikac]

Excited to teach a class this quarter on Trustworthy Machine Learning. Lecture notes available at https://t.co/319LAwY7pT. So far we have notes on the Statistical Learning Framework and Robustness!

[Robert Lange @RobertTLange]

Puuuh. What are you up to these days? 💭 I try to stay sane, clean my place 🧹& write✍️. Todays edition - 'Getting started with #JAX'. Learn how to embrace the 'jit-grad-vmap' powers 💻 and code your own GRU-RNN in JAX. Stay safe & home. 🤗
https://t.co/lYqwpYvCiW