#long-term-dependencies

Question 1: Single choice

mediumAI-checkedRecurrent Neural Networks

You build a stripped-down GRU that keeps only one of its two gates. Which gate should you keep if the task needs information carried across many time steps?

The update gate, because $\mathbf{Z}_t \approx 1$ copies the old state forward almost unchanged1
The reset gate, because $\mathbf{R}_t \approx 1$ lets the whole old state into the update2
Either one, because the two gates become interchangeable once they are trained3
Neither, because a recurrent unit with only one gate has no hidden state at all4

#long-term-dependencies

Timed test