Towards the Universal Law of Robustness for Deep Neural Networks

Towards the Universal Law of Robustness for Deep Neural Networks

Explore fundamental developments in neural network theory through this 28-minute conference talk from Data Science Conference Europe 2022, where Luka Nenadović delves into why overparameterized models with billions or trillions of parameters perform exceptionally well, even when parameters exceed training sample sizes. Learn about the groundbreaking BLN Conjecture (2020) and examine extreme cases of the Bubeck-Sellke Theorem (2021) through an innovative pictorial approach that visualizes neural network behavior from a higher-dimensional geometric perspective. Gain insights into these complex theoretical concepts presented in an accessible format that challenges traditional machine learning paradigms.