Glossary · Term

SGD

← all terms

Definition

Plain language

The basic training method that nudges a model's settings a little at a time, using small random batches of data.

As stated in the literature

Stochastic Gradient Descent, the foundational optimization algorithm for neural networks; its dynamics determine which of several equally-loss-fitting solutions a model settles into, e.g., the negation-neglect bias.

Also called: stochastic gradient descent

Why it matters: It's the engine behind training neural networks, and the way it nudges weights shapes which of many possible solutions a model ends up with.

For example, it adjusts a model's settings a tiny step at a time using a small random batch of examples, gradually reducing its errors.

Mentioned in 1 episode

  1. 043
    When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway

Related terms