Optimization · Level 2 of 5
SGD (SGD)
Gradient descent using sampled examples or mini-batches to estimate gradients.
The stochastic estimate introduces noise compared with a full-data gradient.
Listen to the definition and example
Audio transcript
SGD. Gradient descent using sampled examples or mini-batches to estimate gradients. The stochastic estimate introduces noise compared with a full-data gradient. For example: A model updates from a randomly selected batch rather than the entire dataset.
Explore this concept
A useful analogy
Estimating the downhill direction from a small sample of the landscape.
Why it matters
This helps you understand what parameter updates are trying to improve and when they can fail.
Technical detail
A mini-batch gradient averages ∇ℓ_i(θ) over sampled examples. Under suitable sampling it estimates the full-data gradient.
Common misconception
The stochastic gradient is an estimate, not necessarily the exact full-dataset direction.
Start with
Related concepts
Quick recall question
Try answering before looking back at the definition.