C5W3 (Sequence Models And Attention Mechanism) Flashcards

1
Q

What is Beam search?

A

Algorithm for sampling most likely output sentence from model

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Input parameters for Beam algorithm?

A

B - Beam width
A - degree of normalisation

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

Why do you need to normalise Beam search?

A

Because otherwise Beam search will prefer shorter sentences

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

How to carry error analysis?(RNN)

A

Run model with “correct” and “actual” output.
If probability of “correct” is greater, then problem is in Beam search. Otherwise, in the model itself

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

Bleu score

A

Bilingual Evaluation Understudy.
Used to evaluate model’s performance against equally good sentences

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

What is attention models?

A

Improvement of regular RNN.
It has attention weights applied to the Encoder output. The weights are learned by another NN

How well did you know this?
1
Not at all
2
3
4
5
Perfectly