Media Summary: Verification of neural networks, Box convex approximation, complete vs incomplete methods, sound vs unsound methods, ... Introductory lecture of an ETH Zurich graduate course on topics such as adversarial attacks, experimental and provable defenses, ... This talk with Prof. Ben Shneiderman is the first event in a series of online talks by the Dyson School of Design Engineering.
Reliable And Interpretable Artificial Intelligence - Detailed Analysis & Overview
Verification of neural networks, Box convex approximation, complete vs incomplete methods, sound vs unsound methods, ... Introductory lecture of an ETH Zurich graduate course on topics such as adversarial attacks, experimental and provable defenses, ... This talk with Prof. Ben Shneiderman is the first event in a series of online talks by the Dyson School of Design Engineering. A surprising fact about modern large language models is that nobody really knows how they work internally. At Anthropic, the ... The speaker will discuss the importance of human Certification of Neural Networks, Complete Certification with MILP (Mixed-Integer Linear Solvers), Combination of Complete and ...
Visualization of the decision process in neural networks, connection to adversarial robustness. Adversarial Defenses, PGD defense, min-max optimization, adversarial accuracy vs. natural accuracy. Randomized Smoothing for Robustness Certification, Statistical Certification of Deep Neural Networks, Confidence Intervals.
