Deep Learning for Computer Vision Quiz 1 26 Oct 2025 — Question 19

Select all that apply.

1. A

   Momentum maintains an EMA-like velocity of gradients/updates.

2. B

   RMSProp normalizes step sizes by a running RMS of gradients.

3. C

   In the Adam optimizer, the hyper-parameters β1 and β2 control the exponential decay rates for the moving averages of the gradient and its squared vale. Adam with reduces to SGD with momentum. β1 = 0 and β2 = 0 reduces to SGD with momentum.

4. D

   Cosine-annealed warm restarts can help escape sharp minima.