Training a model is an iterative loop. The model makes a prediction, compares it with the right answer, measures the error and adjusts its internal parameters to reduce it. Over millions of such steps, performance moves from random to expert.
But a model that memorises its training data is useless. The data are therefore split: a training set teaches the model, a validation set tunes its settings, and a held-out test set measures true performance on images the model has never seen. That final test is what the published numbers describe.
The three phases
- Training: the model learns patterns from labelled examples
- Validation: settings are tuned, overfitting is detected
- Test: final performance is measured on unseen images