Risk Management in AI Models - Confidence Estimation in Machine Learning Classifiers

This Hebrew lecture by Gil Einziger from Ben-Gurion University explores confidence estimation in machine learning classifiers and risk management in AI models. Discover how the distinction between model accuracy and confidence values is crucial for determining AI reliability boundaries and risk assessment in decision-making systems. Learn how geometric properties of training data can be leveraged to estimate confidence, with the intuition that models should be more confident when classifying inputs similar to their training examples. The presentation demonstrates how geometric approaches can outperform existing confidence estimation techniques, especially when using appropriate representations like those from transformers. Scheduled for Thursday, April 24th, 2025, at 11:00 AM in room B220, this talk is presented by Dr. Einziger, an Assistant Professor in Computer Science at Ben-Gurion University whose research spans networked systems, algorithms, computer security, and artificial intelligence.