Model Training & Adaptation
L1 regularization
A training rule that discourages large learned settings and can reduce some to zero.
Example
Training reduces the influence of some clues to zero.
Why people use it
It can encourage simpler predictors by reducing some learned influences all the way to zero.
What you'll hear
“Can we make the predictor rely on fewer measurements?”
What this means for you
Do not assume an excluded clue is unimportant outside this particular system.
Can you control it?
Developer-only
The people building or running the AI choose this setup. An everyday user generally needs their help to change how this part works.
Common questions
- Does a removed clue become unimportant in real life?
- No. Removal reflects this particular learning setup; the clue can still matter in other circumstances.
- Can changing the penalty change which measurements remain?
- Yes. A stronger or weaker penalty can change which influences are kept.
- Can similar measurements compete with one another?
- Yes. When they carry overlapping information, the method may favor one over another.