A model-specific hyperparameter is like a special setting on a toy that changes how it plays, and it depends on what kind of toy you have.
Imagine you're playing with two different toys: one is a robot, and the other is a train set. Each has its own way of working, so each needs its own special settings to run best.
The Robot Toy
The robot might need you to choose how many steps it takes at once, that’s like a model-specific hyperparameter for the robot. If you pick 2 steps, it moves slower but more carefully; if you pick 10 steps, it zooms ahead but might miss things.
The Train Set
The train set might need you to decide how many trains can run on the track at once, that’s a model-specific hyperparameter for the train. If only one train runs, it's simple and quiet; if five trains run, it's busy and exciting, but maybe there are more collisions.
Each toy (or model) has its own special settings, which are called model-specific hyperparameters, they help you get the best play out of each one!
Examples
- A neural network uses a 'learning rate' like how fast it learns, while a decision tree might use the number of splits per layer, these are model-specific hyperparameters.
Ask a question
See also
- What are adaptive step sizes?
- How Does Regularization Work?
- How Does Parameters vs hyperparameters in machine learning Work?
- Who is Expected SARSA?
- What are policy gradients?