Ground truth quality for machine learning models

This invention describes a way to automatically find problems in the labeled data used to train artificial intelligence models. It works by taking your labeled data, training a model with it, and then converting each piece of data into a numerical representation. These representations are then grouped together, and the system identifies data points that don't fit well within their groups or are far from others, indicating they might be mislabeled, confusing, or simply unusual. The system then tells you which data points are problematic. The claims specifically mention using images as data and neural networks to create these numerical representations.