Machine Learning
Systems that learn the rules from examples — the foundation everything in this glossary is built on.
Machine learning is the discipline of building systems that learn from data rather than executing hand-written rules: instead of engineers specifying how to recognize a character, classify a document, or spot a fraudulent statement, the system infers the patterns from examples — labeled ones in supervised learning (the paradigm behind most document AI: inputs paired with correct outputs), structure-seeking in unsupervised learning (clustering, anomaly detection), and reward-guided in reinforcement learning. Deep learning, the neural-network branch that scales with data and compute, is the engine behind essentially every capability this glossary describes.
The document AI relevance is genealogical: the field's history is the replacement of rules by learning, layer by layer. Classical OCR's hand-crafted features gave way to learned recognition; template extraction gave way to models that learned layouts; keyword classification gave way to learned semantics; and the current era's foundation models — pretrained on vast corpora, adapted by fine-tuning and instruction — completed the shift: document capabilities now come from data (pretraining corpora, labeled datasets, correction streams) far more than from code. That is why so many entries here concern data disciplines — annotation, ground truth, benchmarks, feedback loops — rather than algorithms: in a learned system, the data is the specification.
The concepts a document AI practitioner borrows daily from ML's core: training/validation/test discipline and its contamination hazards, overfitting and generalization (the model that memorized its templates versus the one that learned documents), distribution shift and drift (production documents wandering from the training population), calibration (scores meaning what they claim), and the bias-variance-data triangle that governs whether the next accuracy point comes from more data, better labels, or a different model. The entries around this one apply each of these to documents specifically; this one marks where they all come from.
The generation gap in text recognition — neural networks that learned to read where rules used to try.
Teaching machines language — the discipline whose modern form is the L in LLM.
Vision, language, and more in one model — the paradigm that made documents a native input.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo