What is dropout and how does it prevent overfitting?
Updated May 17, 2026
Short answer
Dropout randomly disables neurons during training to reduce overfitting.
Deep explanation
It prevents co-adaptation of neurons by randomly zeroing activations with probability p during training, forcing redundancy and improving generalization.
Real-world example
Used in fully connected layers of NLP and vision models.
Common mistakes
- Applying dropout during inference instead of training only.
Follow-up questions
- How does dropout compare to L2 regularization?
- What is inverted dropout?