Machine Learning Foundations

Gradient Descent, explained

Gradient descent is an optimization method that updates model parameters in the direction that most reduces the loss locally.
Back to Glossary

Part of the Semantic Notion technical glossary.

Short Definition

Gradient descent is an optimization method that updates model parameters in the direction that most reduces the loss locally.

Intuition

Imagine walking downhill on the loss surface. The gradient tells you which direction is uphill, so gradient descent steps the other way.

Technical Definition

Gradient descent iteratively updates parameters by subtracting a learning-rate-scaled gradient of the objective with respect to those parameters.

Example

A neural network uses gradients from backpropagation to reduce prediction error over many batches.

Common Misunderstandings

Gradient descent is local; it does not guarantee the globally best solution.
Bigger steps are not always faster because they can overshoot good regions.

Start here

Need the broader concept map?

Return to the glossary index to move from this definition into adjacent AI, ML, mathematics, and computer science terms.