Raw data isn't something a learning algorithm can work with directly. A spreadsheet of customer records, a folder of images, a corpus of text — these are useful to a person, but a model needs something more structured. The first job in building any AI system is turning raw information into a form the algorithm can actually learn from.
This is what feature engineering does. A feature is any measurable property of the data that might be relevant to the prediction you're trying to make. If you're predicting whether a loan will default, features might include income, debt level, employment history, and credit score. Feature engineering is the work of deciding which properties to include, how to measure them, and how to transform them so they're as informative as possible. It's more art than science, and historically it required deep domain expertise.
Once you have features, they get assembled into a feature vector — a structured numerical representation of a single data point. Every example in your training set becomes a vector, and those vectors are what the algorithm actually sees. For unstructured data like text or images, embeddings serve a similar purpose: they're dense numerical representations that capture meaning and relationships in a form the model can work with. The embedding for "king" and the embedding for "queen" end up close together in the mathematical space; the embedding for "king" and "bicycle" end up far apart.
With the data represented, training begins. The model makes predictions, compares them to the correct answers, measures how wrong it was, and adjusts its internal parameters to be a little less wrong next time. This loop runs thousands or millions of times. The settings that control how this process unfolds — how fast the model adjusts, how many examples it sees at once, how long it trains — are called hyperparameters, and tuning them well makes a significant difference in the quality of the result.
Two variations on the training process are worth knowing. Batch learning trains on the full dataset at once, then deploys a static model. Active learning is more selective: the model identifies which examples it's most uncertain about and asks for labels on those specifically, making the labeling process more efficient.
The result of all of this, if it goes well, is a model that has learned something real from the data — something general enough to apply to new examples it's never seen. That's the whole point.
