#sequence-modeling

Question 1: Select all that apply

Compared with a recurrent layer, which statements about a 1D convolutional layer applied to a sequence are true? Select all that apply.

Its outputs at all time steps can be computed in parallel1
It keeps no state between time steps, so it suffers less from unstable gradients2
With a stride above 1 it can downsample a long sequence before an RNN3
A single such layer can relate time steps that are arbitrarily far apart4