… and yet they do. 🤖✨

Deep learning relies on highly complex loss functions. Algorithms should optimize them, but we know they can’t. So why are the results still so impressive?

Our new paper offers a mathematical explanation. 📘🧠 We rigorously prove that deep-learning algorithms don’t actually need to find the true optimum. Being close to a local optimum is already enough. In nerdier terms: We show that every reasonable stationary point of certain neural networks — and all points nearby — generalize essentially as well as the global optimum. 🔍📈

The paper has been accepted at TMLR! 🎉 Find it here.

Huge congratulations to Mahsa and Fang—two rising stars in machine learning. 🌟🌟

We all love sparsity: it makes computations faster, guarantees tighter, and interpretations easier. In our paper, , which will appear in TMLR, we introduce a new type of sparsity, which we term “cardinality sparsity”. We show that cardinality sparsity has all the usual perks, and more importantly, we demonstrate that it is also a very powerful concept for matrix-matrix multiplications. Indeed, cardinality sparsity can speed up such computations and reduce memory usage dramatically. Well done, Ali! 👍👍👍

How many samples are needed to train a deep neural network? Our recent paper explores this question and comes to the conclusion that: it takes many, many samples. The paper will appear in ICLR 2025. Congratulations to Pegah and Mahsa, who did a wonderful job on this! 🍺🍺🍺

Our paper “Single-Model Attribution of Generative Models Through Final-Layer Inversion” has now been accepted at ICML. Well done, Mike and Jonas, and thanks for the great collaboration, Asja! 🍻

Our paper “Benchmarking the fairness of image upsampling methods” is now accepted at ACM FAccT 2024. Congratulations everyone! 🇩🇪🇨🇭🏴‍☠️

Shih-Ting’s paper on targeted deep learning has been accepted at Stat. Good job, Shih-Ting! ⛰️

Our paper “Balancing Statistical and Computational Precision and Applications to Penalized Linear Regression with Group Sparsity” has been accepted at the IEEE Transactions on Information Theory. Well deserved, Néhémy and Mahsa! 🥇🥇

Our paper “DeepMoM: Robust Deep Learning With Median-of-Means” has been accepted at the Journal of Computational and Graphical Statistics. Awesome job, Shih-Ting! 🕺🕺

Our paper “Copula-Based Normalizing Flows” has been accepted at ICML. Cheers, Mike and Asja! 🥂