Blair Bilodeau (@blairbilodeau) header
Blair Bilodeau (@blairbilodeau) profile picture

quant

Toronto, Ontario
Joined August 2011
Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)Photo by Blair Bilodeau (@blairbilodeau)
You can’t take our country — and you can’t take our game.
47,507
42,941
14,881
363,337
55,334,469
Blair Bilodeau retweeted
Join us next week (June 20) for the Theory of Interpretable AI seminar series, where @blairbilodeau will discuss the ** fundamental theoretical limitations ** of attribution methods and its implications for interpretability! 🌐tverven.github.io/tiai-semin… @ML_Theorist @tverven
1
6
3
36
5,013
Our recent PNAS paper shows that widely used interpretability methods, when used to ask simple counterfactual questions about models like “if I pay down this credit card will my credit score increase?”, are provably no better than random guessing. This is really problematic bc...
Excited to finally share that "Impossibility Theorems for Feature Attribution" is published in PNAS. TL;DR Methods like SHAP and IG can provably fail to beat random guessing. w/ @natashajaques @PangWeiKoh @_beenkim PNAS: pnas.org/doi/10.1073/pnas.23… arXiv: arxiv.org/abs/2212.11870
1
6
42
11,827
Blair Bilodeau retweeted
Many previous work of mine and others hinted ‘something fishy’ about saliency-based methods. But we never had a rigorous proof of what we saw. This work “Impossibility Theorems for Feature Attribution", now published in PNAS, to me marks a point of new beginnings.
Excited to finally share that "Impossibility Theorems for Feature Attribution" is published in PNAS. TL;DR Methods like SHAP and IG can provably fail to beat random guessing. w/ @natashajaques @PangWeiKoh @_beenkim PNAS: pnas.org/doi/10.1073/pnas.23… arXiv: arxiv.org/abs/2212.11870
3
46
4
347
84,429
Excited to finally share that "Impossibility Theorems for Feature Attribution" is published in PNAS. TL;DR Methods like SHAP and IG can provably fail to beat random guessing. w/ @natashajaques @PangWeiKoh @_beenkim PNAS: pnas.org/doi/10.1073/pnas.23… arXiv: arxiv.org/abs/2212.11870
4
27
6
183
73,748
Our theory applies to many models, including neural nets, which we empirically validate. Thm 3.3 is equivalent to saying your ROC curve will be a diagonal, and when we use real methods to conduct hypothesis tests about models trained on ML datasets, that's what we see!
1
5
1,160
Where do we go from here? We now know we can't always trust the intuitive conclusions of feature attributions. But we can use hypothesis testing to understand these methods. This opens up a new direction: design methods that reliably test properties of trained models. n/n, n = 6
1
8
929
3.5 years after we started this project while visiting @the_IAS, it’s been accepted to the Annals of Statistics. I won’t be at #JSM2023 but make sure to stop by Jeff’s session to hear about our work!
Super excited that this work has been accepted for publication in AoS. For those at #JSM2023, I'll be speaking about it in the session on Advances in Statistical Machine Learning, scheduled for Sunday 8/6/2023 2:00:00 PM. I'm speaking first so don't be late 😉.
2
2
23
10,314
Blair Bilodeau retweeted
Feature visualizations are widely used interpretability tools - but can we trust them? We investigate this question from an adversarial 🥷, empirical 🔬 and theoretical 📝 perspective. The result: Don’t trust your eyes! (1/6) Paper: arxiv.org/abs/2306.04719 🧵
3
75
3
303
87,593
Celebrating #Statistics PhD Students for #WomenInScience in #Ontario: Emma Kroell, @UofTStatSci 👤 emmakroell.ca/ Featured Work 📚 https://xitter.gq/t.co/ZRqMOuBpa4 More ℹ️ https://xitter.gq/t.co/Ns1KW2PZP1 #February11 #WomenInSTEM @UofT @researchuoft @UofTNews @UofTArtSci @CANSSIINCASS
5
11
2,136
Blair Bilodeau retweeted
Started my PhD with a pandemic, about to end it with a global recession. Is there a support group or something?
61
167
40
3,769
356,928
Interested in adaptivity, bandits, or causality? Come say hi today at #NeurIPS2022 4pm-6pm Hall J #805
Is it possible to efficiently identify the optimal intervention while remaining agnostic to assumptions about the causal structure? In new work with Linbo Wang and @roydanroy, we study adapting to the presence of a d-separator using multi-armed bandits. arxiv.org/abs/2202.05100
14
I'm at #NeurIPS2022 this week! Come check out my poster for arxiv.org/abs/2202.05100 on Tuesday, 4pm-6pm Hall J #805 Reach out to chat about adaptivity, uncertainty, coffee, or music! Looking forward to finally putting an in-person face to some twitter handles and zoom screens
I am on the job market for faculty and research scientist positions. I develop statistical theory and algorithms to make data-based decisions that balance the need for robustness in high-stakes settings with strong performance in practice. Some highlights below: 1/n
1
10
I am on the job market for faculty and research scientist positions. I develop statistical theory and algorithms to make data-based decisions that balance the need for robustness in high-stakes settings with strong performance in practice. Some highlights below: 1/n
3
31
5
146
This work studies online learning in between IID and adversarial data. We show how to optimally aggregate information without knowing the structure of the data in advance. xitter.gq/blairbilodeau/status/1… We published follow-up work at NeurIPS 2021. xitter.gq/jeffNegrea/status/1453… 7/n
Did you ever wish you could get Quantile/sqrt{KL} Regret Bounds AND use FTRL? Now you can! Our new work on a general family of FTRL algorithms shows how. arxiv.org/abs/2110.14804 We are excited to announce that this work will appear at #NeurIPS2021!
1
4
More of my work and my CV can be found at blairbilodeau.ca. In addition to research, I have a demonstrated history of commitment to undergrad teaching and academic service. DM me here or email [email protected] with any opportunities. 8/n, n=8
5