Dropout
Randomly switching off part of the network during training so it can’t over-rely on any one path.
Think of it like
Making a team practice with random players benched, so no single star becomes a crutch.
Example
A model prone to memorizing its training data might use dropout to force more robust, spread-out learning.
How it actually works
During training, dropout zeroes a random fraction of activations each step, so neurons must learn redundant, general features instead of brittle co-dependencies. It’s a cheap, effective regularizer against overfitting. At inference it’s turned off and the full network runs.
For product teams
A standard anti-overfitting knob — it trades a little training speed for better generalization.
For engineers
Stochastically masks activations with probability p at train time; a regularizer disabled at inference.
Related
- Overfitting — Fights overfitting.
- Generalization — Improves generalization.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome