If you are interested to know how ”Redundant features can hurt robustness to distribution shift” check out our virtual poster at ICML'20 UDL Workshop. This is joint work with @ApostolosModas, Seyed Moosavi and Pascal Frossard. 📄 Paper: bit.ly/3eGqy3A (1/11)

Very excited with the new blog post of our NeurIPS 2020 paper “Neural Anisotropy Directions” where we study the role of architecture in deep learning. Co-authors: @modasapo, Seyed Moosavi and Pascal Frossard. 💻 Blog: bit.ly/3lI81rO 📄 Paper: bit.ly/3jWhP0y 1/n

Our new "tutorial" on what we've learned from adversarial robustness beyond security is online! Still a pre-print, so any comment/criticism is more than welcome! :) With Guillermo Ortiz-Jiménez, Seyed Moosavi, Pascal Frossard

Nicolas Flammarion is looking for #PhD student. Find out more about his #research epfl.ch/labs/tml/, and learn more about our EPFL #EDIC #computerscience PhD program: go.epfl.ch/phd-edic

Our framework provides a geometric connection between adversarial examples and data features! Find us today 6-8pm CET (9-11am PST) @ Town A0, spot A0. #NeurIPS2020

A huge DenseNet vs an old-fashioned Logistic Regression, solving a linearly separable problem. Who is gonna win??

Hours of thorough and hard work, now published in the Proceedings of IEEE 💣🥳 Many credits to my collaborators Guillermo Ortiz-Jiménez Seyed Moosavi Pascal Frossard! 💪

Apparently adversarial robustness improves transfer learning even in non-trivial classification tasks!🧐 Check our new paper "Improving filling level classification with adversarial training" 🍷🥃 📄arXiv: bit.ly/3kFZZ38 Pascal Frossard Andrea Cavallaro Alessio Xompero Ricardo Sánchez Mati

When there exist multiple valid solutions for a given task, do neural networks always go for the “simplest” one? with Guillermo Ortiz-Jiménez, Itamar, @modasapo and Pascal Frossard. Happy to hear your thoughts on that :)

Guillermo Ortiz-Jiménez @modasapo Seyed Moosavi Pascal Frossard "Deep nets are commonly believed to always converge to the simplest possible solution. This study elegantly demonstrates that they sometimes prefer complex solutions to simple ones, depending on the inductive biases in their architecture" ~Review by Anon rentry.co/rev_fs_0002

If you can donate money to support Ukraine now, here are some vetted funds 👇

Are you amazed by NeRFs, SIRENs or FFNs but don’t know why they work? 🤔 Then you’re going to love our work 🚀: “A structured dictionary perspective on implicit neural representations” #CVPR2025 which explains how INRs manage to look so crisp! 🤓 📄: bit.ly/3adThjE 🧵 1/9

Are you amazed by weight interpolation techniques to edit large pre-trained models? 🤔 Then you’re going to love our work 🚀: “Task Arithmetic in the Tangent Space”🏄 which explains why they work and gives a simple way to make them better! 🤓 📄: rb.gy/q9djm 🧵1/9

Do you want to learn more about ethical and responsible image data collection? Then come find us at #NeurIPS2023 🤓 Oral: Thu. 10:30 CST (Oral 5B) 🚀 Poster #1602: Thu. 10:45 CST (Poster session 5) 👨‍💼

Glad to have two papers accepted to #EMNLP2024 as main! 🎉 1) "Resampled Datasets Are Not Enough: Mitigating Societal Bias Beyond Single Attributes" arxiv.org/abs/2407.03623 w/ Jerone Andrews, Dora Zhao, Orestis Papakyriakopoulos, Apostolos Modas, Yuta Nakashima, Alice Xiang 🧵1/2

Nature research paper: Fair human-centric image dataset for ethical AI benchmarking go.nature.com/4hSUagA