Mentimeters

Question 1

A CNN filter is applied to

Answer 1

all channels across layer

Question 2

Stride is

Answer 2

step with which the filter is applied

Question 3

Padding

Answer 3

increases the size of input data

Question 4

Pooling

Answer 4

• combines feature values within a region
• downsamples feature maps

Question 5

CNN activation is applied to

Answer 5

channel

Question 6

Hyperparameters can be learned with

Answer 6

validation

Question 7

FC layer is typically used

Answer 7

close to the output side of the network

Question 8

Typical loss for multiclass classification

Answer 8

• cross entropy
• softmax
• negative log likelihood

Question 9

ReLu can be applied

Answer 9

before or after max-pooling

Question 10

Learning rate is

Answer 10

step of weights update

Question 11

Weights are not updated once per

Answer 11

epoch

Question 12

All training data is used to update weights in one

Answer 12

epoch

Question 13

Averaging updates over iterations is called

Answer 13

momentum

Question 14

first and second order moments of gradients and used in

Answer 14

• adadelta
• RMSProp
• Adma
• Adagrad

Question 15

Batch normalisation is applied to

Answer 15

channels

Question 16

Dropout is an effective regularisation of

Answer 16

fully connected layers

Question 17

L2 regularisation of weights is called

Answer 17

decay

Question 18

finetuning is a process of

Answer 18

updating parameters pretrained on another dataset

Question 19

data augmentation consists of

Answer 19

generating new samples from existing ones

Question 20

hard negative is a

Answer 20

negative example which is similar to a positive one

Question 21

hard positive is a

Answer 21

positive sample which is dissimilar to positive ones

Question 22

to debug a model

Answer 22

overfit on a small dataset

Question 23

bias in a dataset is

Answer 23

confusing noise introduced during data collection

Question 24

VGG uses

Answer 24

3x3 filters and max pooling

Question 25

VGG is widely used because of its

Answer 25

effective feature representation

Question 26

efficiency of 1x1 filters was exploited in

Answer 26

inception

Question 27

inception block uses

Answer 27

parallel filters with concatenated outputs

Question 28

skip connections are used in

Answer 28

ResNet

Question 29

skip connections in ResNet

Answer 29

do not change data

Question 30

best performing word embedding is

Answer 30

Bert

Question 31

Which unit is least effective in remembering sequences

Answer 31

RNN

Question 32

Gating mechanism uses

Answer 32

sigmoid

Question 33

in GRU hidden state and input are

Answer 33

concatenated

Question 34

Language modelling uses architecture type

Answer 34

many to many

Question 35

Transformers self attention uses

Answer 35

linear projections

Question 36

What is the goal of reinforcement learning?

Answer 36

maximise expected return

Question 37

The discount factor in value function is used to

Answer 37

weigh immediate and future rewards

Question 38

which behaviour is exploration in game exploring

Answer 38

play an experimental move

Question 39

what is the main drawback to monte carlo sampling approach to RL

Answer 39

needs to run an entire episode before updating

Question 40

what is the main different between Q learning and SARAR

Answer 40

SARAR uses epsilon greedy update policy

Question 41

main problem of Q-learning

Answer 41

not scalable

Question 42

in deep Q learning 'deep' is mainly used to

Answer 42

approximate Q function

Question 43

in policy based methods do we select actions according to value function

Answer 43

no

Question 44

policy optimisation can only be performed using gradient based methods

Answer 44

false

Question 45

REINFORCE is based on

Answer 45

monte carlo

Question 46

in REINFORCE with baseline the baseline is used to

Answer 46

reduce variance

Question 47

which method is not designed to reduce variance

Answer 47

REINFORCE

Question 48

in actor-critic methods, critic is similar to which part of a GAN

Answer 48

discriminator

Question 49

compared to value-based methods, policy-based methods can handle continuous action easily?

Answer 49

true

Question 50

which parameters are not hyperparameters

Answer 50

weights of convolutional kernel

Question 51

which hyperparameters optimization method is more efficient

Answer 51

random search

Question 52

in successive halving, the number of configurations n indicates

Answer 52

exploration

Question 53

in meta learning only training tasks contain training set and test set

Answer 53

false

Question 54

in meta learning total loss is computed using

Answer 54

test examples

Question 55

meta learning and multi task learning are the same

Answer 55

false

Question 56

which colour representation can be used to compute colour similarities

Answer 56

RGB colour space

Question 57

unsupervised representation learning can't be used for

Answer 57

learning a mapping function from dataset and labels

Question 58

autoencoder is an

Answer 58

unsupervised method

Question 59

in autoencoder the decoder must be symmetric to the encoder

Answer 59

false

Question 60

as long as an autoencoder can reconstruct the input, this autoencoder can learn useful representations of the input

Answer 60

false

Question 61

what objective function is used to train autoencoder

Answer 61

reconstruction loss

Question 62

which is not an autoencoder?

Answer 62

disruptive

Question 63

which autoendcoder can be used to perform dimensionality reduction

Answer 63

undercomplete autoencoders

Question 64

in autoencoders which technique is used for anomaly detection

Answer 64

reconstruction error

Question 65

which autoencoder should be used to recover noisy data

Answer 65

denoising autoencoder

Question 66

an image classification model is a

Answer 66

discriminative model

Question 67

VAEs

Answer 67

explicit methods

Question 68

how are VAEs trained

Answer 68

maximising likelihood

Question 69

the reparameterisation trick in VAEs is used for

Answer 69

training

Question 70

GANs are

Answer 70

implicit methods

Question 71

which loss is better for training the generator of GANs

Answer 71

non-saturating heuristic

Question 72

what do GANs and VAEs have in common

Answer 72

both are generative models

Question 73

Are VAEs easier to train but generate less sharp images?

Answer 73

yes