Dropout
Domain: Evaluation
🎯 Motivation
Randomly turning off neurons during training. Forces the network to be robust, not dependent on any single neuron.
📋 Concrete Example
Dropout rate 0.2: during training, 20% of neurons are randomly turned off each step. The network learns redundant paths. At test time, all neurons are on. More robust.
🔗 Analogy
A basketball team practicing with players sitting out - the team learns to play well even when a player is missing. Dropout makes the network robust to missing neurons.