mi1_05_rnn

Did you attempt the last exercise sheet?

• Yes

• No

How is this course working out for you so far?

• Boring

• Need time to digest everything

• I'm completely lost

• ok and not so ok

I usually have questions (after/during) the lecture/tutorial...

• afterwards but don't end up asking

• I still need time to go over the material

• I can ask them

• during but ...

I understand things better when...

• I go over the material on my own time

• I browse for alternative material

• ...I've tried everything and I'm still lost

• ...I'll deal with understanding later. Let's move on.

• ... we discuss them in lecture and/or tutorial

An RNN can be used to classify an entire time series (T/F)

• True

• False (we must see the prediction at every time step)

• False, RNNs are not classifiers

• What's a time series?

The Backpropagation algorithm...

• is a learning algorithm

• computes the EXACT gradients.

• APPROXIMATES the gradients

• is a classification model

Backpropagation through time...

• is a learning algorithm

• computes the exact gradients.

• approximates the gradients

• is a classification model

Unfolding an RNN in time means

• turning the RNN into a feedforward network

• recomputing the EXACT same operation at t=1 over and over

• Making an RNN go back in time.

Weights in an unfolded RNN...

• The weight matrices between sucessive layers are duplicates of W

• The W's between time steps are always different.

RNNs are difficult to train because of...

• exponential forgetting

• unfolding an RNN is very tricky

• no one really understands how to implement an RNN

• vanishing gradient problem