Lab 7: Uncertainty estimation

CIFAR-10N: uncertainty and human label noise

Problem: uncertain predictions and noisy labels

CIFAR-10N augments the CIFAR-10 training images with human-provided labels that include real annotation errors. It provides a setting for studying how model confidence and predictive uncertainty relate to correctness when supervision can be unreliable. Label noise and predictive uncertainty are related concepts, but they are not interchangeable.

See the official CIFAR-10N dataset release and the authors’ Learning with Noisy Labels Revisited.

Lab instructions

TODO(instructor): Lab instructions will be added here.