NEW
Font size
WorksheetsDEL Test
Total questions: 25
Worksheet time: 13mins
Which of the following is not required for learning?
A defined Task
A Defined Evaluation method
An adaptable automaton
A Testing Data Set
Why can linear regression be seen as machine learning?
It can linearly separate space
More inputs and targets (experience) improve its performance
It is used as the basis of Perceptrons
Which of these models overfits?
The middle one
The right one
The left one
None of them
All of them
In deep learning, an activation function of a hidden layer should not :
Have discontinuities
Be linear
Have a continuous derivative (except isolated points)
In order to monitor the capacity of a model we don't need:
A training dataset
A predetermined threshold (target value of J(w))
A validation dataset
If my model has to predict which type of image this is, the final activation should not be:
soft-max
tanh
sigmoid
linear
θ{t+1}<−θ{t}−ϵt B1b=0∑B−1∂θ∂L(θ,mb) An online training implies:
B== 1
B <= Number of Elements in Training Dataset
B < Number of Elements in Training Dataset
B == Number of Elements in Training Dataset
Which of the following is true about ReLU?
All neurons always activate positively in a ReLU layer
It only focuses on presence, not intensity
It is non-linear
It fully resolved the vanishing gradient problem
Vanishing gradient happens when:
The learning rate is too strong
The activations of some ReLU are set to 0
The activation functions are linear
Consecutive derivatives multiplications tend to 0
Dropout is equivalent to which type of normalization?
Elastic Net
L2
L1
L1.5
Batch Normalization
Has no learnable parameters
Needs Batches >8 to be robust
Needs batches <128 to be robust
Increases the dependence on the weights initialization
Convolutional Neural Networks don't
Use shared weights
Require pooling to work
Work without ReLU
Work on 1D images
A pooling layer does not:
Work with a window of odd size (ex: 3x3, 5x5)
Require training
Perform a compression of the image
Work with options other than max and average
Hyperparameters:
Are always fixed to a single value during training
Are optimizable by TPE / Gridsearch / randomsearch
Do not impact the convergence of the network
Cannot be set manually
Basic Recurrent Neural Networks don't
Have shared weights
Have an internal state
Have a forget gate
Work on time-series
Bi-dirrectional LSTMs:
Must predict the whole sequence
Take into account the whole time-sequence
Require a separate metric for each direction
Cannot be used to auto-encode time-series
I am dealing with temporal data. I need:
to use LSTMs
to use Conv1Ds
to predict the whole series, not only one time-step
to take care with data leakage
Res-nets
Can be considered to learn ODEs
Suffer from vanishing gradient a lot
Do not work without ReLU
Are only useful with convolutional networks
Variational Autoencoders cannot :
compress data
function without convolutions
be used to unclassify data
be used to generate new data
Generative Networks do not require:
A conditioning
Training
A latent random variable
Gradient descent
The adversarial part of a GAN ideally:
Performs a regression task
Starts unable to detect the real data
Should not change its weights during training
Should have a loss that fluctuates around 2log(0.5)
Diffusion Models
Learn during the backward process
Learn during both the forward and backward process
Learn during the forward process
Require a T<20 to function
Deep Image Prior
Is inexpensive
Does not require more than one image to train on
Does not require a stopping point to prevent overfitting
Can be effectively applied only on non-structed data
What is more important?
Validating your results
Lowering the training error
Having the minimum amount of parameters
Using all explanatory variables
You get a new database. What do you do last:
Visualize it
Train a deep network on it
Select how you split it in train / test / validation
Check documentation on it
