Recurrent Neural Networks and Their Implications in Learning

Page 1 sur 18Lecteur de document UniversityLib

Recurrent Neural Networks and Their Implications in Learning

Deep Learning - Recurrent Neural Networks (RNNs) · exam

Voir tous les documents en intelligence artificielle et données

Question 1

Correct

Note de 1,00 sur 1,00

To which of these tasks would you apply a many-to-one RNN architecture? (Check all that apply).

A. Gender recognition from speech

B. Sentiment classication (input a piece of text and output a 0/1 to denote positive or negative

sentiment)

C. Image classication

D. Machine Translation (input a piece of text and the output is a translated text)

Votre réponse est correcte.

Les réponses correctes sont :

Sentiment classication (input a piece of text and output a 0/1 to denote positive or negative sentiment),

Gender recognition from speech

Question 2

Correct

Note de 1,00 sur 1,00

Consider this RNN: This specic type of architecture is appropriate when:

A.

T < T

y

x

B. T = Ty

x

C. T > T

y

x

D.

T = T =1

y

x

Votre réponse est correcte.

La réponse correcte est :

T = Ty

x

Question 3

Correct

Note de 1,00 sur 1,00

You have nished training a language model RNN and are using it to sample random sentences, as

follows:

What are you doing at each time step t:

A. (1) Pick the highest probability from the output probabilities of the RNN  for the time step

, (2) the selected word is passed to the next step t+1.

B. (1) Randomly sample a chosen word according to the given output probabilties  of the RNN  for

the time step

, (2) the selected word is passed to the next step t+1.

C.

D.

(1) Use the probabilities given by the RNN to sample a word for the time step

, (2) the

ground truth word from the training set is passed to the next step t+1.

(1) The highest probability word of the time step t as

is picked from output probabilities,

(2) the ground truth word from the training set is passed to the next step t+1.

Votre réponse est correcte.

La réponse correcte est :

(1) Randomly sample a chosen word according to the given output probabilties  of the RNN  for the time step

, (2) the selected word is passed to the next step t+1.

Question 4

Partiellement correct

Note de 0,67 sur 1,00

Which of these is the most likely cause of problems for RNNs:

A. Exploding gradients.

B. Vanishing gradients.

C. Dicult to process longer sequences.

D. Activations are too large.

Votre réponse est partiellement correcte.

Publicité

Vous en avez sélectionné correctement 2.

Les réponses correctes sont :

Exploding gradients.,

Vanishing gradients.,

Dicult to process longer sequences.

Question 5

Correct

Note de 1,00 sur 1,00

The Gated Recurrent Units (GRU) helps avoiding vanishing/exploding gradients

Where:

Check all that apply.

A. z  is a gating vector that determines how much of the past information should be passed along to

t

the future and it depends on the reset gate vector r

t

B. The reset gate r  is used to decide how much of the past information to forget.

t

C. The nal output state h is a combination of h  and ϕ (W x +U (r ⊙h )+b ) via z and it is the

t−1

h t

h t

t−1

t

t

h

h

representation of the input x

t

D. The nal output state h is a combination of h  and ϕ (W x +U (r ⊙h )+b ) via z and it is the

t−1

h t

h t

t−1

t

t

h

h

representation of the hidden state h

t -1

Votre réponse est correcte.

Les réponses correctes sont :

The nal output state h is a combination of h  and ϕ (W x +U (r ⊙h )+b ) via z and it is the

t−1

h t

h t

t−1

t

t

h

h

representation of the input x ,

t

The reset gate r  is used to decide how much of the past information to forget.

t

Question 6

Correct

Note de 1,00 sur 1,00

Join the RNN architecture to the right application:

Many-to-many

Many-to-many

Many-to-one

Name entity recognitionMachine translationSentiment classicationOne-to-one

One-to-many

Votre réponse est correcte.

La réponse correcte est :

Many-to-many

Publicité

Many-to-many

→ Name entity recognition,

Traditional neural networkMusic generationMany-to-one

One-to-one

One-to-many

→ Machine translation,

→ Sentiment classication,

→ Traditional neural network,

→ Music generation

Question 7

Correct

Note de 1,00 sur 1,00

Join the gure to the right activation function:

Votre réponse est correcte.

La réponse correcte est :

SigmoidRELUTanh → Sigmoid,

→ RELU,

→ Tanh

Question 8

Correct

Note de 1,00 sur 1,00

What is true about RNN (check all that apply)

A. Computation takes into account historical information

B. Can consider any future input for the current state

C. Weights are shared across time

D. Possibility of processing input of any length

Votre réponse est correcte.

Les réponses correctes sont :

Possibility of processing input of any length,

Computation takes into account historical information,

Weights are shared across time

Question 9

Correct

Note de 1,00 sur 1,00

You are training an RNN, and nd that your weights and activations are all taking on the value of NaN

(“Not a Number”). Which of these is the most likely cause of this problem?

A. Vanishing gradient problem

B. Exploding gradient problem

Votre réponse est correcte.

La réponse correcte est :

Exploding gradient problem

Question 10

Correct

Note de 1,00 sur 1,00

Then evne the world "Ecstatic" does not appear in your training set, the RNN might reasonably be

expected to recognize "I'm ecstatic" as deserving the label=1

A. False

B. True

Votre réponse est correcte.

La réponse correcte est :

True

Question 11

Correct

Note de 1,00 sur 1,00

Suppose you are training a LSTM. You have a 10000 word vocabulary, and are using an LSTM with 100-

dimensional activations a<t>. What is the dimension of Update gate Γ at each time step?

u

A. 10000

B. 10

C. 100

D. 100000

Votre réponse est correcte.

La réponse correcte est :

100

Question 12

Partiellement correct

Publicité

Note de 0,67 sur 1,00

Some word embedding techniques are:

a. Skip-gram

b. GloVe

c. Word2vec

d. GliVer

Votre réponse est partiellement correcte.

Vous en avez sélectionné correctement 2.

Les réponses correctes sont :

Word2vec,

Skip-gram,

GloVe

Question 13

Correct

Note de 1,00 sur 1,00

Join the type of the gate and where it is used.

Update gate Γu : How much past should matter now?

Forget gate Γ :Erase a cell or not?

f

Relevance gate Γ :Drop previous information?

r

Output gate Γ : How much to reveal of a cell?

o

Votre réponse est correcte.

La réponse correcte est :

Update gate Γu : How much past should matter now? → GRU, LSTM,

Forget gate Γ :Erase a cell or not? → LSTM,

f

Relevance gate Γ :Drop previous information? → GRU, LSTM,

r

Output gate Γ : How much to reveal of a cell?

o

→ LSTM

Question 14

Correct

Note de 1,00 sur 1,00

Gated Recurrent Unit (GRU) are  generalization of  Long Short-Term Memory units (LSTM)

with combination of the forget and input gates into a single “update gate.”

A. False

B. True

Votre réponse est correcte.

La réponse correcte est :

False

GRU, LSTMLSTMGRU, LSTMLSTMQuestion 15

Correct

Note de 1,00 sur 1,00

Sort the LSTM process in the right order.

The putput gate decides what we’re going to output.

The input gate and the update gate decide what new information we’re going to store in the

cell state

The forget gate decides what information we’re going to throw away from the cell state.

Votre réponse est correcte.

La réponse correcte est :

The putput gate decides what we’re going to output. → 3,

The input gate and the update gate decide what new information we’re going to store in the cell state → 2,

The forget gate decides what information we’re going to throw away from the cell state. → 1

Question 16

Correct

Note de 1,00 sur 1,00

What is the purpose of training RNN model?

A. Learn patterns in ABC music and generate a new piece based on model.

B. Predict following character in each time step

C. Given character & sequence whats the most probable next character

D. Generate a new piece of music based on model

E. Maintaining internal state that depends on previously seen elements

Votre réponse est correcte.

Publicité

La réponse correcte est :

Learn patterns in ABC music and generate a new piece based on model.

321Question 17

Correct

Note de 1,00 sur 1,00

The forget gate  looks at h  and x , and outputs a number between 0 and 1 for each number in the

t−1

t

cell state C . A 1 represents “completely keep this” while a 0 represents “completely get rid of this.”

t−1

It is helpful to use tanh activation then sigmoid.

A. False

B. True

Votre réponse est correcte.

La réponse correcte est :

False

Question 18

Correct

Note de 1,00 sur 1,00

The GRU combines the gating functions of the input gate and the forget gate into a simple update gate.

Further, the cell state and the hidden output are combined into a single hidden state layer. Choose the

right answer.

A. LSTM trains a little faster than the GRU.

B. GRU trains a little faster than the LSTM.

Votre réponse est correcte.

La réponse correcte est :

GRU trains a little faster than the LSTM.

Question 19

Correct

Note de 1,00 sur 1,00

The gradient clipping (Coupure de gradient) is:

a. a technique used to cope with the vanishing gradient problem sometimes encountered when

performing backpropagation

b. a technique used to cope with the exploding gradient problem sometimes encountered when

performing backpropagation

Votre réponse est correcte.

La réponse correcte est :

a technique used to cope with the exploding gradient problem sometimes encountered when performing

backpropagation

Question 20

Correct

Note de 1,00 sur 1,00

Match the description to thier situation

Happens rarely, but the eect can be catastrophic and might destroy what has been

learned so far.

Means that learning long term dependencies in data is dicult for simple RNN

architecture

The gradients start to increase exponentially during backpropagation through the

recurrent weights

As we propagate the gradients back in time, usually their magnitude quickly

decreases

Votre réponse est correcte.

La réponse correcte est :

Happens rarely, but the eect can be catastrophic and might destroy what has been learned so far.

→ Exploiding gradients,

Means that learning long term dependencies in data is dicult for simple RNN architecture

→ Vanishing gradients,

The gradients start to increase exponentially during backpropagation through the recurrent weights

→ Exploiding gradients,

As we propagate the gradients back in time, usually their magnitude quickly decreases → Vanishing gradients

◄ Test Evaluation #3

Aller à…

Devoir Surveillé 23/12/2021 ►

Exploiding gradientsVanishing gradientsExploiding gradientsVanishing gradients