Gated Recurrent Units Quiz

Answers and explanations appear as you take the quiz.

1. How many gates does the GRU cell in the guide use?
   • Two: an update gate and a reset gate
   • Three plus a separate cell state
   • One gate per sequence position
   • No gates
2. What does the update gate control?
   • Which input token is removed permanently
   • The number of sequence layers
   • Whether to compute a separate cell state
   • How much of the previous hidden state to carry versus new candidate information
3. What does the reset gate control?
   • The final sequence length
   • The learning rate of the optimizer
   • How much past information is used when forming a candidate state
   • Whether the update gate exists
4. How does the GRU represent memory compared with an LSTM?
   • It uses a separate cell state and three gates
   • It uses one hidden state rather than a separate cell state
   • It has no recurrent state
   • It stores the entire training sequence in the gates
5. For a GRU update gate z, what does a value near one do?
   • Favor the new candidate state over the previous hidden state
   • Control how much past information is used in the reset-gated candidate
   • Decide when the sequence should stop processing
   • Carry most of the previous hidden state into the next state
6. Why can gating help with long-range dependencies?
   • It gives the model a learned route to preserve or update hidden information across steps
   • It eliminates all recurrent computation
   • It guarantees perfect memory at any length
   • It turns a sequence into an unordered set
7. What distinguishes a GRU candidate state from the prior hidden state?
   • It is always the ground-truth next token
   • It is computed from the current input and reset-gated past information
   • It is copied unchanged from the input
   • It stores a second LSTM cell
8. What should be concluded from fewer GRU gates than LSTM gates?
   • GRUs always outperform LSTMs
   • GRUs cannot model sequences
   • The architecture is simpler in this respect, but task performance still needs evaluation
   • Fewer gates guarantee better accuracy