Do ImageNet Classifiers Generalize to ImageNet? - Analyzing ML Progress and Challenges

Explore a thought-provoking colloquium presentation by Ludwig Schmidt from UC Berkeley, focusing on the generalization capabilities of ImageNet classifiers. Delve into the analysis of machine learning progress, examining the challenges in achieving safe, dependable, and secure AI systems. Investigate the nature of overfitting in ML benchmarks through reproducibility experiments on ImageNet and other key datasets. Discover insights on distribution shift as a major obstacle for reliable machine learning. Learn about adversarial examples and innovative methods inspired by optimization and generalization theory to address robustness issues. Gain valuable knowledge from a comprehensive experimental study of current robustness interventions, highlighting the main challenges in the field. Benefit from the expertise of Ludwig Schmidt, a postdoctoral researcher at UC Berkeley with a focus on making machine learning more reliable.