Model Training
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
The phase where a machine learning model learns from data: an algorithm adjusts the model's parameters to shrink the gap between predictions and truth.
What is Model Training?
Model training is the phase in which a machine learning model learns from data: an algorithm feeds it examples and adjusts its internal parameters to reduce the error between its predictions and the known answers.
Everything a deployed model does traces back to this phase. Questions about a model’s bias, its blind spots, or the provenance of its knowledge are questions about what happened during training.
How Model Training Works
Training follows a repeating cycle. The data is cleaned and split into training and validation sets. The model then makes predictions on the training data, and a loss function measures how wrong those predictions are.
An optimization method such as gradient descent uses that error signal to nudge the parameters in the direction that reduces the loss. The cycle repeats over the data many times, and the error shrinks with each pass.
Finally, the model is checked against the validation set – data it never saw during training. Strong results on training data but weak results on validation data signal overfitting: the model memorized its examples instead of learning the pattern.
Model Training vs Inference
Training is when the model learns; inference is when the finished model is put to work on new data. The two phases have opposite economics. Training is a large one-time (or periodic) compute bill; inference is a small cost paid on every single request, forever.
| Model training | Inference | |
|---|---|---|
| What happens | Parameters change with every batch of examples | Parameters are frozen; the model only produces outputs |
| When it runs | Before deployment, and again for retraining | Every time the deployed model receives a request |
| Compute profile | Large batches on GPU clusters, for hours or weeks | Single requests, where milliseconds of latency matter |
| Typical failure | Overfitting, data leakage, mislabeled data | Latency spikes, data drift, serving a stale model |
Example of Model Training
Suppose you train a classifier to tell cat photos from dog photos. You start with thousands of labeled images and hold some back for validation. Early in training, the model’s guesses are no better than coin flips.
Each pass over the data, the loss function scores the errors and gradient descent adjusts millions of parameters by tiny amounts. The model gradually latches onto the shapes and textures that separate the two animals.
Training stops when accuracy on the held-out validation images stops improving. What remains is a fixed set of parameters, ready for inference on photos the model has never seen.
Related AI terms: Inference · Overfitting · Gradient Descent · Training Set · Fine-Tuning
Did you like the Model Training gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About