What changes when a neural network is trained
Training a neural network adjusts numerical parameters so its outputs better match a chosen objective on training data. An error measure guides those adjustments through repeated examples. The process can produce useful patterns, but good training performance alone does not show that the model will handle unfamiliar cases well.
Go deeper
Adjust parameters using an objective
A neural network transforms inputs through layers of numerical operations. Its parameters, often called weights and biases, influence the output.
During training, an objective or loss function measures how the output compares with the desired behaviour. Optimisation methods adjust parameters to reduce that loss across examples.
Learning here means changing the parameters, not storing a human explanation of every answer. The resulting behaviour depends on the data, architecture and objective.
- Training data provides examples.
- The loss defines what improvement means.
- Evaluation on separate data tests performance beyond the training examples.
A model can fit its training data too closely and perform poorly elsewhere, a problem called overfitting. Lower training loss is therefore not sufficient evidence that the model has learned a broadly useful relationship.