Invariants in machine learning are mathematical measures used to reduce the amount of data needed for training models. With invariants, machine learning algorithms can identify patterns and symmetries to generate accurate models with fewer examples.