Two readings of one shape

The workshop’s entry check: X.shape == (20, 8, 8).

Give two possible meanings for its axes. Both (images, rows, columns) and (time, rows, columns) fit. Then predict X[:, 3, 4].shape — it is (20,): one pixel across images, or that same pixel across time.

The selection is identical in both readings. Only the name of the remaining axis changes, and the name is the part that decides what you may do next.