Lecture
This lecture delves into the evaluation of machine and human accuracy and robustness on ImageNet, showcasing the explosive growth in machine learning. It discusses the progress over the past decade, the impact of overfitting, and the implications of accuracy drops. The lecture also explores the shortcomings of current metrics, the importance of multi-label accuracy, and the challenges in evaluating ImageNet. It further examines the robustness notions in image classification, the implications for evaluating machine learning, and the need for broader evaluation beyond ImageNet. The lecture concludes by discussing the substantial room for improvement in machine learning benchmarks and the importance of new datasets for training and testing.