Gradient descent
Improving a model by stepping each weight a little downhill on the error, again and again.
Learn it through a story
- Fog on Nandi Hills (Jugnupur Tales Vol 2, Ch 2)
Short on time? Read the 5-minute version.
Related terms
- Error, or loss A number that measures how wrong a model’s answers are.
- Feature A clue the AI looks at, like colour or shape.
- Generalising Doing well on new examples, not just old ones.
- Held-out set Data kept aside, like a sealed exam paper.
- Input A number that goes into a neuron, like how clear the sky is.
- Label The right answer given with an example.