EXAMEN

Semestre : 1 2 X

Session : Principale X Rattrapage

Module: Deep Learning
Enseignant(s): UP GL-BD & Math
Classe(s): 4DS
Documents autorisés: OUI NON X Nombre de pages: 8
Calculatrice autorisée: OUI NON X Internet autorisée: OUI NON X

Date: 25/05/2022 Heure: 14h30 Durée : 1h30min

Question 1: In Keras, we set the inputdim parameter on which layer of the neural network?

A) Input layer.

B) Hidden layers.

C) Dropout layer.

D) Output layer.

Question 2: Which of the following is FALSE about Softmax function?

A) Softmax function is a generalized version of a sigmoid function.

B) Softmax function is usually used in the hidden layers of neural networks.

C) Softmax function is usually used for multi-class classification problem.

D) None of the above.

Question 3: Which of the following METHODS is used to train a Keras Sequential model?

A) model.compile

B) model.fit

C) model.evaluate

D) model.predict

Question 4: Gradient descendent computes derivative of loss function w.r.t:

A) Input

B) Activation value

C) Weight

D) Bias

Question 5: Batch Normalization is helpful because?

A) It normalizes (changes) all the input before sending it to the next layer.

1
 B) It returns back the normalized mean and standard deviation of weights

C) It is a very efficient backpropagation technique

D) None of these

Question 6: Which of the following frameworks is used as backend in Keras by default?

A) TensorFlow.

B) CNTK.

C) Theano

D) MXNet

Question 7: What is the purpose of an activation function?

A) To decide whether a neuron will fire or not.

B) To increase the depth of a neural network.

C) To create connectivity among hidden layers

D) To normalize the inputs

Question 8: Order the following steps to use a gradient descent algorithm?

1) Calculate error between the actual value and the predicted value
2) Reiterate until you find the best weights of network
3) Pass an input through the network and get values from output layer
4) Initialize random weight and bias
5) Go to each neuron which contributes to the error and change its respective values to reduce the error

A) 3, 2, 1, 5, 4.

B) 4, 3, 1, 5, 2.

C) 1, 2, 3, 4, 5

D) 5, 4, 3, 2, 1

Question 9: Which of the following is the structure of the input to an artificial neuron?

A) Weighted sum of inputs + bias.

B) Sum of inputs + bias.

C) Weighted sum of bias + inputs

D) Sum of bias + inputs

Question 10: Which of the following terms is NOT associated with CNN?

A) Zero and Valid Padding.

B) Strides.

2
 C) Filters(Kernels).

D) Forgot Gates

Question 11: Which of the following is FALSE about Stochastic Gradient Descent?

A) Performs model updates for each single observation.

B) Frequent model updates can also result in a noisy gradient signal.

C) Learning is slower than batch gradient descent.

D) None of the above.

Question 12: Which of the following is FALSE about ML and DL algorithms?

A) DL algorithms are more interpretable as compared to ML algorithms.

B) Data augmentation can be done easily in DL compared to ML.

C) Deep learning algorithms efficiently solve computer vision problems.

D) None of the above.

Question 13: Which of the following model is best suited for sequential data?

A) Autoencoders.

B) Convolutional Neural Networks (ConvNets).

C) Capsule Neural Networks (CapsNets)

D) RNN(Recurrent Neural Network)

Question 14: Output of which of the following activation functions is zero centered?

A) Hyperbolic Tangent.

B) Sigmoid.

C) Softmax.

D) ReLU.

Question 15: Which of the following is TRUE about Convolutional Layer in CNN?

A) It is used to extract features from the input images.

B) It uses a matrix filter and performs convolutional operation to detect patterns in the image.

C) We may lose some information about edges and corners of an image during convolution.

D) All of the above.

Question 16: Which of the following is FALSE about Learning Rate?

A) Instead of adaptive learning rate, we should prefer constant learning rate.

3
 B) Model may never converge if learning rate is too high.

C) High learning rate speeds up the training process.

D) None of the above.

Question 17: Output of sigmoid activation function ranges from:

A) 0 to 9.

B) 0 to 1.

C) -1 to 1.

D) -1 to 0.

Question 18: Which of the following is FALSE about Epochs in a neural network?

A) Time to complete one epoch depends upon the batch size.

B) Number of epochs is the number of times the whole training data is shown to the network.

C) Number of epochs is a learnable parameter.

D) None of the above.

Question 19: In a neural network, which of the following techniques is used to deal with overfitting?

A) Regularization.

B) Dropout.

C) Batch Normalization.

D) All of the above.

Question 20: Which of the following is TRUE about Backpropagation.

A) Helps to improve the performance of the network.

B) Is a technique to improve the performance of the network.

C) It backpropagates the error and updates the weights to reduce the error.

D) All of the above.

Question 21: You are building a binary classifier for recognizing cucumbers (y=1) vs. watermelons (y=0). Which one of
these activation functions would you recommend using for the output layer?

A) Tanh.

B) Sigmoid.

C) Leaky Relu.

D) ReLU.

4
Question 22: Pooling layers are used to accomplish which of the following?

A) To ovoid local minima.

B) To reduce the number of epochs while ensuring best accuracy.

C) To progressively reduce the spatial size of the representation.

D) To always select maximum value over pooling region.

Question 23: Why to use Pooling Layers?

A) To reduce the dimensions of the feature maps.

B) Improve image quality

C) To keep image structure

D) To change the structure of the image

Question 24: A neural network model is said to be inspired from the human brain. The neural network consists of many
neurons, each neuron takes an input, processes it and gives an output. Which of the following statement
correctly represents a real neuron?

A) A neuron has a single input and a single output only.

B) A neuron has multiple inputs and multiple outputs.

C) A neuron has multiple inputs but a single output only

D) All of the above statements are valid

Question 25: In a neural network, knowing the weight and bias of each neuron is the most important step. If you can
somehow get the correct value of weight and bias for each neuron, you can approximate any function. What
would be the best way to approach this?

A) Assign random values and pray to God they are correct.

B) Search every possible combination of weights and biases till you get the best value.

C) Iteratively check that after assigning a value how far you are from the best values, and slightly
change the assigned values values to make them better.

D) None of these.

Question 26: What is Vanishing Gradient?

A) It describes the situation where a deep multilayer feed-forward network or a recurrent neural network
is unable to propagate useful gradient information from the output end of the model back to the layers
near the input end of the model.

B) It describes the overfitting of the model

C) It describes the increase in the values of the gradient descent.

D) None of the above.

Question 27: Which of the following gives non-linearity to a neural network?

5
 A) Stochastic Gradient Descent.

B) Rectified Linear Unit.

C) Convolution function

D) None of the above.

Question 28: Which of the following is true about model capacity (where model capacity means the ability of neural
network to approximate complex functions) ?

A) As number of hidden layers increase, model capacity increases.

B) As dropout ratio increases, model capacity increases.

C) As learning rate increases, model capacity increases

D) None of these.

Question 29: Among the following, which ones are “hyperparameters”?

A) Weights.

B) Number of iterations.

C) Activation values.

D) B and C.

Question 30: Which of the following statement is the best description of early stopping?

A) Train the network until a local minimum in the error function is reached.

B) Simulate the network on a test dataset after every epoch of training. Stop training when the
generalization error starts to increase .

C) Add a momentum term to the weight update in the Generalized Delta Rule, so that training
converges more quickly.

D) A faster version of backpropagation, such as the ‘Quickprop’ algorithm .

Question 31: For a classification task, instead of random weight initializations in a neural network, we set all the weights
to zero. Which of the following statements is true?

A) There will not be any problem and the neural network will train properly.

B) The neural network will train but all the neurons will end up recognizing the same thing. .

C) The neural network will not train as there is no net gradient change.

D) None of these.

Question 32: For an image recognition problem (recognizing a cat in a photo), which architecture of neural network
would be better suited to solve the problem?

A) Multi Layer Perceptron.

B) Convolutional Neural Network. .

6
 C) Recurrent Neural network.

D) Perceptron.

Question 33: What are the factors to select the depth of neural network?

1) Type of neural network (eg. MLP, CNN etc)

2) Input data
3) Computation power, i.e. Hardware capabilities and software capabilities
4) Learning Rate
5) The output function to map

A) 1), 2), 4), 5).

B) 2), 3), 4), 5). .

C) 1), 3), 4), 5).

D) All of these.

Question 34: How many input nodes are required to process a colored image of 28 × 28?

A) 56 × 56.

B) 28 × 28 × 3. .

C) 84 × 84.

D) 28 × 28.

Question 35: Which of the following is an application of RNN?

A) Text mining.

B) Sentiment analysis. .

C) Text and Speech analysis.

D) All of the above.

Question 36: How many input nodes are required to process a grayscale image of 28?

A) 28 × 1.

B) 28 × 28. .

C) 56 × 56.

D) 56 × 1.

Question 37: Which of the following is FALSE about zero padding?

A) It is used to preserve the spatial size of the input valume.

B) It is used to preserve edge information of the image. .

C) It is used to preserve resolution of the image.

D) None of the above.

7
Question 38: Which of the following is FALSE about Kernels in CNN?

A) Kernels can be used in convolutional as well as in fully connected layers.

B) Kernels act as pattern detectors. .

C) Kernels lead to dimensionality reduction.

D) We may lose some information about edges and corners of an image using kernels.

Question 39: Which of the following is FALSE about loss functions?

A) Cross entropy can be used for both binary and multi-class classification problems.

B) One hot encoding of the target variable is not required with sparse cross entropy. .

C) Mean square error is commonly used for regression problems.

D) None of the above.

Question 40: Which of the following is FALSE about Deep Neural Networks?

A) These are computationally more complex as compared to shallow network.

B) These have less generalization capabilities as compared to shallow networks. .

C) These may suffer from overfitting problem.

D) These may suffer from vanishing gradient problem.

Bon travail