## How a deep neural network learns

A model learns by updating and improving its weights and biases every epoch (when we call the `fit()` function).

It does so by comparing the patterns its learned between the data and labels to the actual labels.

If the current patterns (weight matrices and bias values) don't result in a desirable decrease in the loss function (higher loss means worse predictions), the optimizer tries to steer the model to update its patterns in the right way (using the real labels as a reference).

This process of using the real labels as a reference to improve the model's predictions is called [**backpropagation**](https://en.wikipedia.org/wiki/Backpropagation).

In other words, data and labels pass through a model (**forward pass**) and it attempts to learn the relationship between the data and labels.

If this learned relationship isn't close to the actual relationship or it could be improved, the model does so by going back through itself (**backward pass**) and tweaking its weights and bias values to better represent the data.