Robust Vision: State of the Art

This lecture covers the importance of robust vision in the context of visual intelligence, focusing on the challenges posed by distribution shifts, both adversarial and non-adversarial. It delves into benchmarking robustness, improving models through training-time mechanisms like data augmentation and architectural changes, and test-time mechanisms such as adaptation signals. The instructor discusses failure examples, the significance of robust statistics, and the impact of non-robust features. The lecture also explores the reasons for failure, including texture vs shape biases, spurious correlations, and biased data distributions. Various types of distribution shifts, imperceptible and perceptible adversarial shifts, and active benchmarks are explained, along with strategies to enhance robustness through diverse data pretraining and mixing datasets.