#minibatch-size

Question 1: Single choice

Under a fixed wall-clock time budget, you train the same MLP with SGD using small and large minibatches. Which statement describes the trade-off?

Larger batches give less noisy gradients and allow a larger learning rate, but fewer updates fit into the budget1
Larger batches give noisier gradient estimates but allow more parameter updates per second2
The batch size has no effect on how training progresses; only the learning rate matters here3
Very small batches are slow to compute, but they give the most accurate gradient estimates4