Excited to teach a class this quarter on Trustworthy Machine Learning. Lecture notes available at cseweb.ucsd.edu/classes/sp20/c…. So far we have notes on the Statistical Learning Framework and Robustness!
Kamalika Chaudhuri
3,673 posts
Kamalika Chaudhuri
@kamalikac
Researcher, Google Deepmind. Formerly, Director FAIR @ Meta. Former Professor at UCSD. Researcher in AI privacy, security, and generalization.
- See all videos from ICML 2019 here.New videos from ICML 2019 have been posted and linked into the schedule. See all videos here: icml.cc/Conferences/20…
- New blog post: ucsdml.github.io/jekyll/update/… Paper: arxiv.org/abs/2004.05675 TLDR; we formalize a type of overfitting called data-copying in generative models, and provide a three-sample non-parametric test to detect it. Work w/ @kc_meehan and S. Dasgupta. To appear in #AISTATS2020
- Do representation learning models memorize their training data? To understand this, we propose a new method called Deja vu to measure memorization in these models. #NeuRIPS2023
- At ICML 2019, @rsalakhu and I ran a new Code-at-Submit-Time Experiment. How did it do? Find out here:
- New paper at #ICML2020 on robustness of non-parametric methods. Paper: arxiv.org/abs/2003.06121 We take a closer look at robustness properties of general non-parametric methods in the large sample limit.
- How expressive are some normalizing flows? Find out at the #AISTATS2020 poster session for our paper "The Expressive Power of a Class of Normalizing Flow Models". Paper: arxiv.org/abs/2006.00392 Poster link: aistats2020.net/poster_1181.ht… Time: 3pm PDT Wed/9am PDT Fri Work w/ Z. Kong
- We know that LLMs learn facts, but how well do they properly unlearn pieces of knowledge? We investigate "deep unlearning" in our new paper.
- How do we better design classifiers that know when they don't know? There are two different kinds of uncertainty measures in the literature -- aleatoric, or due to inherent noise in the data from overlapping classes, and epistemic, or uncertainty due to atypical inputs.
- Excited to speak at the Adversarial Learning Workshop at KDD tomorrow!The 2nd workshop on Adversarial Learning Methods for Machine Learning and Data Mining @kdd_news will take place tomorrow! (8 am to noon Pacific time), live-streamed by Youtube (link will be posted tmr) sites.google.com/view/advml Invited speakers: @kamalikac @QuanquanGu @UnaMayMIT
- Thoughtful debate must begin with accurate facts. At @icmlconf there is gentle encouragement, but no "demand" for code, and no requirement whatsoever about data. See our FAQ: icml.cc/Conferences/20…Replying to @ilyasut
- Congratulations to my student @chhaviyadav_ for successfully defending her thesis today! Chhavi is doing super interesting work in trustworthy ml - keep an eye out for her. I am excited to see what she does next!Successfully defended my thesis today! Extremely grateful to my amazing advisor, @kamalikac, whose unwavering support, insights, teachings and motherly love made all this possible.. Even the darkest of days seemed lighter & hopeful after spending just 5 mins with her! It has been
