ML at @UBC_CS and @AmiiThinks. Trans 🏳️⚧️, she/her.
Vancouver
Joined March 2010
- Tweets205
- Following661
- Followers883
- Likes2.7K
coming back to x, the everything app, to say:
Submit work, sign up to review, come to the workshop! This is a great chance to bring together a lot of really cool work that's been happening, but not all as connected (nor as easy to publish) as it should be!
testing.ml
Excited to share our ICML 2026 Hypothesis Testing Workshop in Seoul, this July! @icmlconf
🎉This workshop aims to bring together researchers developing modern hypothesis testing methodology and applying it to machine learning problems such as robustness, distribution shift, security, medicine, and LLM evaluation. In other words, if you care about how we make ML claims rigorous, this workshop is for you.
We now have four confirmed speakers:
Arthur Gretton @ArthurGretton,
Yao Xie @yaoxie21851119,
Bo Li @uiuc_aisecure, and
Yisong Yue @yisongyue.
The organizing team includes Xiuyuan Cheng (Duke), Feng Liu @AlexFengLiu1, Lester Mackey @LesterMackey, Shayak Sen @shayaksen, Danica J. Sutherland
@d_j_sutherland, and Nathaniel Xu (UBC).
📌 Submission deadline: 10 May 2026
📌 Notification: 26 May 2026
📌 Camera-ready: 17 June 2026
📌 Workshop date: July 10 or 11, 2026 (TBA)
🚩Check more information below!
🔗Website: testing.ml
🔗Submission Portal: openreview.net/group?id=ICML…
We’re also recruiting PC members/reviewers.
🔗 Reviewer interest form: docs.google.com/forms/d/e/1F…
🏁Please feel free to share this with colleagues, collaborators, and students who may be interested.
#ICML #ICML26
Danica Sutherland retweeted
Excited to share our ICML 2026 Hypothesis Testing Workshop in Seoul, this July! @icmlconf
🎉This workshop aims to bring together researchers developing modern hypothesis testing methodology and applying it to machine learning problems such as robustness, distribution shift, security, medicine, and LLM evaluation. In other words, if you care about how we make ML claims rigorous, this workshop is for you.
We now have four confirmed speakers:
Arthur Gretton @ArthurGretton,
Yao Xie @yaoxie21851119,
Bo Li @uiuc_aisecure, and
Yisong Yue @yisongyue.
The organizing team includes Xiuyuan Cheng (Duke), Feng Liu @AlexFengLiu1, Lester Mackey @LesterMackey, Shayak Sen @shayaksen, Danica J. Sutherland
@d_j_sutherland, and Nathaniel Xu (UBC).
📌 Submission deadline: 10 May 2026
📌 Notification: 26 May 2026
📌 Camera-ready: 17 June 2026
📌 Workshop date: July 10 or 11, 2026 (TBA)
🚩Check more information below!
🔗Website: testing.ml
🔗Submission Portal: openreview.net/group?id=ICML…
We’re also recruiting PC members/reviewers.
🔗 Reviewer interest form: docs.google.com/forms/d/e/1F…
🏁Please feel free to share this with colleagues, collaborators, and students who may be interested.
#ICML #ICML26
Danica Sutherland retweeted
📢Curious why your LLM behaves strangely after long SFT or DPO?
We offer a fresh perspective—consider doing a "force analysis" on your model’s behavior.
Check out our #ICLR2025 Oral paper:
Learning Dynamics of LLM Finetuning!
(0/12)
Danica Sutherland retweeted
In our Spexphormer method, a smaller network estimates the attention scores of a larger one. This led us to a fundamental question: How small can the network be while still producing accurate estimates? We tackled this question through rigorous theoretical analysis. 1/
Danica Sutherland retweeted
As a reminder, we will have our poster session tomorrow:
📍 East Exhibit Hall, Poster #3010
📄 arxiv.org/abs/2411.16278
💻 github.com/hamed1375/Sp_Exph…
To motivate you further, we have some insights gained from the attention score analysis of this work, which I'll share in this thread:
Graph Transformers (GTs) can handle long-range dependencies and resolve information bottlenecks, but they’re computationally expensive. Our new model, Spexphormer, helps scale them to much larger graphs – check it out at @NeurIPSConf next week, or the preview here!
#NeurIPS2024
Danica Sutherland retweeted
LLM's self-play is ubiquitous. What will happen if M[t] iteratively learns from M[t-1] for too many generations?
Come and chat with us at NeurIPS
🗓️Friday, Dec 13
📍East Exhibit Hall A-C, Poster #3305
⏰11:00 AM–2:00 PM PST.
[1/7]
Danica Sutherland retweeted
Graph Transformers (GTs) can handle long-range dependencies and resolve information bottlenecks, but they’re computationally expensive. Our new model, Spexphormer, helps scale them to much larger graphs – check it out at @NeurIPSConf next week, or the preview here!
#NeurIPS2024
Grokking on modular arithmetic: early (kernel) phase can overfit but cannot generalize, but with small regularization GD eventually escapes the kernel regime and can provably generalize. Poster #913 this afternoon at #ICML, come hear about it!
What causes 𝙜𝙧𝙤𝙠𝙠𝙞𝙣𝙜 on modular addition problems?
Our #ICML2024 work identifies the 𝙥𝙚𝙧𝙢𝙪𝙩𝙖𝙩𝙞𝙤𝙣 𝙚𝙦𝙪𝙞𝙫𝙖𝙧𝙞𝙖𝙣𝙘𝙚 of the task as the root cause of poor generalization early in training.
Paper: arxiv.org/abs/2407.12332
Danica Sutherland retweeted
Previous live link is now private. Cut version is here: youtu.be/O2gpl5l2eQA?si=1WMA…
Danica Sutherland retweeted
Testing for conditional independence is notoriously difficult.
Here's o̶n̶e̶ three weird tricks to:
👾 make testing work better on limited data,
👾 get correct false positive rates
arxiv.org/abs/2402.13196
@rmnpogodin,@AntoninSchrab,@yazhe_li,@d_j_sutherland
Danica Sutherland retweeted
We're hiring a Postdoc and two Ph.D. students to work together on privacy problems in ML at DTU. (and yes, I moved again, lol)
efzu.fa.em2.oraclecloud.com/…
efzu.fa.em2.oraclecloud.com/…
Come say hi at #ICML2023!
- This afternoon (#837): orders-of-magnitude speedup in evaluating empirical NTKs, with guarantees arxiv.org/abs/2206.12543
- Thursday morning (#229): Exphormers, scalable graph transformers based on expander graphs arxiv.org/abs/2303.06147
Plus, secret bonus: this morning (#704) I'll tell people about goodness-of-fit testing for sequential models (eg strings), arxiv.org/abs/2210.10741 by Jerome Baum, Heishiro Kanagawa, @ArthurGretton.
Teaser thread about Exphormer :)
1/ 🧵I'm pleased to announce Exphormer: Sparse Transformers for Graphs (with my fantastic coauthors @HamedShirzad13 @ijalabv @d_j_sutherland and Ali Sinop), which appears at ICML 2023.
Come check out our poster at 10:30am HST on July 27!
arxiv.org/abs/2303.06147
Danica Sutherland retweeted
1/ 🧵I'm pleased to announce Exphormer: Sparse Transformers for Graphs (with my fantastic coauthors @HamedShirzad13 @ijalabv @d_j_sutherland and Ali Sinop), which appears at ICML 2023.
Come check out our poster at 10:30am HST on July 27!
arxiv.org/abs/2303.06147
If you want to avoid accidentally uploading comments in your tex source to arXiv, consider using a script like github.com/djsutherland/arxi… that strips them automatically (among other niceties). :p
You might know that MSFT has released a 154-page paper (arxiv.org/abs/2303.12712) on #OpenAI #GPT4 , but do you know they also commented out many parts from the original version?
🧵: A thread of hidden information from their latex source code
[1/n]
Danica Sutherland retweeted
Turn your HSIC dependence statistic into a conditional dependence statistic with CIRCE!
Learn NN features that are independent of distractors/protected attributes, conditioned on labels.
Used for domain invariant learning and fairness with equalized odds.
arxiv.org/abs/2212.08645
Danica Sutherland retweeted
📢We are excited to launch our 2022-23 Queer Experiences in AI Survey!📢
Consider filling it if you identify as queer🏳️🌈 and work in AI! form.jotform.com/22248072096…
The responses will aid in understanding issues faced by our community, to inform relevant policies in the near future.
Danica Sutherland retweeted
fuck this is so sad, rest in power Daniel. he was a a trans man and employee of Club Q. he was killed last night. here’s one of his recent tweets about the club that shows why queer spaces are so important. this is what they want to take from us
