Login to unlock
Deep Learning · Phase 3: Optimization · Gradient Descent Variants

Gradient Descent Variants — Checkpoint Quiz

Checkpoint quiz 3 questions
Back to course

Check your understanding of this module before moving on.

?

Checkpoint quiz 3 questions

A quick check on this module's key ideas before you move on
0/3
Question 01

What is the main practical difference between batch gradient descent and stochastic gradient descent (SGD)?

A is correct. Batch gradient descent gives an exact gradient but requires a full pass over the data per update, which is slow on large datasets. SGD (and mini-batch SGD) trades some gradient noise for the ability to make many more updates per epoch.
Question 02

How does momentum help gradient descent in practice?

A is correct. Momentum keeps a moving average of recent gradient directions and adds it to the current update. This damps oscillations in noisy directions and builds speed in consistent directions, like a ball rolling downhill with inertia.
Question 03

Why might mini-batch gradient descent (e.g. batches of 32 or 64 examples) be preferred over both full-batch and single-example SGD in most deep learning practice?

B is correct. Mini-batches give a gradient estimate that's far less noisy than a single example but far cheaper to compute than the full dataset, and the batch size maps well onto GPU parallelism, which is why it's the default in practice.