Tech-ai

כנס iSpeech - ההרשמה נפתחה!
הצטרפו אלינו ב-23.7.25 לשמוע את פריצו הדרך בתחום הבינה המלאכותית, מולי השפה וטכנולוגיות הדיבור!
https://www.techaiconf.org/

💫 Another Spotlight on AI Research at the Technion! 💫
Generalization in deep neural networks remains a fundamental mystery—despite their ability to fit even random labels, they often generalize well in practice, seemingly in defiance of the classical theory of generalization.
This raises the question—how well do neural networks generalize when trained to perfectly fit a noisy dataset?
In a recent paper published at NeurIPS 2024, Itamar Harel, William M. Hoza (University of Chicago), Gal Vardi (Weizmann Institute of Science), Itay Evron, Prof. Nati Srebro (Toyota Technological Institute at Chicago), and Prof. Daniel Soudry consider fully connected neural networks with binary weights, analyzing both minimal networks (with the fewest possible weights) and typical networks (randomly initialized ones that interpolate the data), proving that in both cases generalization error is proportional to the noise in the data, similar to empirical evidence. Unlike previous work, their results apply to overparameterized networks with input dimension that is neither very large nor very small.

You can read the full article here: https://arxiv.org/abs/2410.19092