Training

In short

The process of fitting a Model to a specific use case by adjusting its parameters. An iterative process of prediction, evaluation, and updates.

It’s like learning at school. You’re taught something, then you have practice exercises, then flash tests to check understanding, and at the end a final exam. You never have repeating exercises — otherwise you’d just memorize answers like a poem, and that’s not the goal.

Training depends on the use case. There are two main approaches: Supervised Learning (you have the task and the solution) and Unsupervised Learning (you only have the task).

We always need Data for training, and preferably good data (see Data Quality). The data gets split into three parts: training, validation, and testing (see Data Splitting).

The process itself works like this:

1. Start with random parameters
2. Feed training inputs to the model, get predictions
3. Use a Cost Function to measure how bad the prediction was
4. Use Gradient Descent to update the parameters and improve
5. Validate on separate data to check real understanding
6. Repeat until performance is good enough

When the model starts memorizing instead of learning, that’s Overfitting. When it’s not adapting at all, that’s Underfitting. The sweet spot is somewhere in between.

Related

• Data - you always need data
• Supervised Learning - training with known answers
• Unsupervised Learning - training without known answers
• Data Splitting - train / validation / test
• Cost Function - measures prediction quality
• Gradient Descent - updates the parameters
• Overfitting - memorizing instead of learning
• Underfitting - not learning at all