Parameters are the values a model learns for itself during training.
In a simple linear model the parameters are the weight and the bias.
A neural network can have millions or billions of parameters.
Training just means searching for parameter values that give good predictions.
You never set parameters by hand, which is what separates them from hyperparameters.
Parameters are the knobs the model turns for itself.
You never set them, and you rarely look at them.
A simple straight line has two: how steep it is and where it crosses.
A large language model has hundreds of billions.
Training is nothing more than a search for good values for those knobs.
They all start out as small random numbers.
Random matters, because if they all started equal they would all learn the same thing.
More parameters means the model can represent more complicated patterns.
It also means more ways to memorise instead of learn.
The right number depends on how much data you actually have.
A million parameters on a thousand rows will simply memorise the thousand rows.
Parameters are what gets written to disk when you save a model.
Loading a model means loading those numbers back in.
Two models with the same shape but different parameters behave completely differently.
The shape is your choice, and the values are the model's.
One question on this card. Every option explains itself once you pick, right or wrong.
55 concepts in this module, then a graded test with every answer explained.