Out-of-Distribution Detection
The capability of a machine learning system to reliably identify when an input significantly deviates from the distribution of its original training data.
Think of It Like This
Like a dog that knows exactly how to herd sheep, but immediately stops and barks for help when asked to herd a flock of ostriches.
Models notoriously make highly confident but completely wrong predictions on OOD data. Detecting these anomalies is vital for AI safety in critical fields like autonomous driving or medicine. Techniques include analyzing softmax entropies, using distance-based metrics in latent space, or training explicit generative models to estimate data likelihood.