wayground logo

Free Printable Worksheets

NEW

Font size

S
M
L
XL
Worksheets

DEL Test

Total questions: 25

Worksheet time: 13mins

Name
Class
Date
1.

Which of the following is not required for learning?

a)

A defined Task

b)

A Defined Evaluation method

c)

An adaptable automaton

d)

A Testing Data Set

2.

Why can linear regression be seen as machine learning?

a)

It can linearly separate space

b)

More inputs and targets (experience) improve its performance

c)

It is used as the basis of Perceptrons

3.

Which of these models overfits?

a)

The middle one

b)

The right one

c)

The left one

d)

None of them

e)

All of them

4.

In deep learning, an activation function of a hidden layer should not :

a)

Have discontinuities

b)

Be linear

c)

Have a continuous derivative (except isolated points)

5.

In order to monitor the capacity of a model we don't need:

a)

A training dataset

b)

A predetermined threshold (target value of J(w))

c)

A validation dataset

6.

If my model has to predict which type of image this is, the final activation should not be:

a)

soft-max

b)

tanh

c)

sigmoid

d)

linear

7.

θ{t+1}<θ{t}ϵt 1Bb=0B1L(θ,mb)θ\theta_{\left\{t+1\right\}}<-\theta_{\left\{t\right\}}-\epsilon_t\ \frac{1}{B}\sum_{b=0}^{B-1}\frac{\partial L\left(\theta,m_b\right)}{\partial\theta}  An online training implies:

a)

B== 1

b)

B <= Number of Elements in Training Dataset

c)

B < Number of Elements in Training Dataset

d)

B == Number of Elements in Training Dataset

8.

Which of the following is true about ReLU?

a)

All neurons always activate positively in a ReLU layer

b)

It only focuses on presence, not intensity

c)

It is non-linear

d)

It fully resolved the vanishing gradient problem

9.

Vanishing gradient happens when:

a)

The learning rate is too strong

b)

The activations of some ReLU are set to 0

c)

The activation functions are linear

d)

Consecutive derivatives multiplications tend to 0

10.

Dropout is equivalent to which type of normalization?

a)

Elastic Net

b)

L2

c)

L1

d)

L1.5

11.

Batch Normalization

a)

Has no learnable parameters

b)

Needs Batches >8 to be robust

c)

Needs batches <128 to be robust

d)

Increases the dependence on the weights initialization

12.

Convolutional Neural Networks don't

a)

Use shared weights

b)

Require pooling to work

c)

Work without ReLU

d)

Work on 1D images

13.

A pooling layer does not:

a)

Work with a window of odd size (ex: 3x3, 5x5)

b)

Require training

c)

Perform a compression of the image

d)

Work with options other than max and average

14.

Hyperparameters:

a)

Are always fixed to a single value during training

b)

Are optimizable by TPE / Gridsearch / randomsearch

c)

Do not impact the convergence of the network

d)

Cannot be set manually

15.

Basic Recurrent Neural Networks don't

a)

Have shared weights

b)

Have an internal state

c)

Have a forget gate

d)

Work on time-series

16.

Bi-dirrectional LSTMs:

a)

Must predict the whole sequence

b)

Take into account the whole time-sequence

c)

Require a separate metric for each direction

d)

Cannot be used to auto-encode time-series

17.

I am dealing with temporal data. I need:

a)

to use LSTMs

b)

to use Conv1Ds

c)

to predict the whole series, not only one time-step

d)

to take care with data leakage

18.

Res-nets

a)

Can be considered to learn ODEs

b)

Suffer from vanishing gradient a lot

c)

Do not work without ReLU

d)

Are only useful with convolutional networks

19.

Variational Autoencoders cannot :

a)

compress data

b)

function without convolutions

c)

be used to unclassify data

d)

be used to generate new data

20.

Generative Networks do not require:

a)

A conditioning

b)

Training

c)

A latent random variable

d)

Gradient descent

21.

The adversarial part of a GAN ideally:

a)

Performs a regression task

b)

Starts unable to detect the real data

c)

Should not change its weights during training

d)

Should have a loss that fluctuates around 2log(0.5)

22.

Diffusion Models

a)

Learn during the backward process

b)

Learn during both the forward and backward process

c)

Learn during the forward process

d)

Require a T<20 to function

23.

Deep Image Prior

a)

Is inexpensive

b)

Does not require more than one image to train on

c)

Does not require a stopping point to prevent overfitting

d)

Can be effectively applied only on non-structed data

24.

What is more important?

a)

Validating your results

b)

Lowering the training error

c)

Having the minimum amount of parameters

d)

Using all explanatory variables

25.

You get a new database. What do you do last:

a)

Visualize it

b)

Train a deep network on it

c)

Select how you split it in train / test / validation

d)

Check documentation on it