December 02, 2022
This paper investigates the failure cases and out-of-distribution behavior of trans- formers trained on matrix inversion and eigenvalue decomposition. I show that incorrect model predictions still retain deep mathematical properties of the solution (e.g. correct eigenvalues, unit norm of eigenvectors), and that almost all model fail- ures can be attributed to, and predicted from, properties of the problem or solution. This demonstrates that, when in doubt, math transformers do not hallucinate absurd solutions (as was sometimes proposed) but remain “roughly right”. I also show that the careful choice of a training dataset can accelerate training, while allowing the model to generalize out of its training distribution, invalidating the idea that transformers “merely interpolate” from memorized examples.
Written by
François Charton
Publisher
Neurips MAH-AI workshop
April 14, 2024
Heng-Jui Chang, Ning Dong (AI), Ruslan Mavlyutov, Sravya Popuri, Andy Chung
April 14, 2024
February 21, 2024
Tom Sander, Pierre Fernandez, Alain Durmus, Matthijs Douze, Teddy Furon
February 21, 2024
February 15, 2024
Danny Deng, Hongkuan Zhou, Hanqing Zeng, Yinglong Xia, Chris Leung (AI), Jianbo Li, Rajgopal Kannan, Viktor Prasanna
February 15, 2024
February 15, 2024
Adrien Bardes, Quentin Garrido, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, Nicolas Ballas, Jean Ponce
February 15, 2024
Product experiences
Foundational models
Product experiences
Latest news
Foundational models