RNNs Quiz

RNNs Quiz

Check your understanding of recurrent networks — how weight sharing across time works, why gradients vanish, and how the LSTM's gated cell state fixes it.

Exercise

How are an RNN's weights used across time steps?

Need a hint?

Unroll the network — how many distinct weight matrices do you see?

Exercise

What is backpropagation through time (BPTT)?

Need a hint?

First unroll, then apply something familiar.

Exercise

Why do gradients vanish in vanilla RNNs over long sequences?

Need a hint?

What happens when you multiply 50 numbers that are each less than 1?

Exercise

How does the LSTM cell state combat vanishing gradients?

Need a hint?

Compare 'multiply at every step' with 'add at every step'.

Exercise

What does the LSTM forget gate control?

Need a hint?

It outputs values in (0, 1) that multiply the previous cell state.

Exercise

State-space models (S4, Mamba) train much faster than RNNs on long sequences. What property makes that possible?

Need a hint?

What does a LINEAR recurrence unroll into?

Exercise

An RNN applies the same weights at every time step (weight sharing across time). Which slot of the ML project loop is that, and what does it buy?

Need a hint?

It's a structural assumption about the model family — an inductive bias for sequences.

Exercise

Transfer test — a technique this course did NOT teach: temperature scaling, applied after training to make a classifier's probabilities calibrated. Which slot does it modify?

Need a hint?

It's a post-hoc fix measured and applied at evaluation time.