Each model is comprised of several layers. Each layer is a data transformation. This transformation is captured using a bunch of numbers, called layer weights. This is not a complete truth though, since most layers often have a mathematical operation associated with them, for example, convolution or an affine transform. A more precise perspective would be to say that a layer is parameterized by its weights. Hence, we use the terms layer parameters and layer weights interchangeably.
The state of all the layer weights together makes the model state captured in model weights. A model can have anywhere between a few thousand to a few million parameters.
Let's try to understand the notion of model learning in this context: learning means ...
