NPTEL INTRODUCTION TO ARTIFICIAL INTELLIGENCE WEEK 12 ASSIGNMENT ANSWERS

Q 1. What is the height of the output image after an input image of size 128 x 128 is passed through a convolution filter of size 32 x 32 and stride 8 followed by a convolution filter of size 3 x 3 and stride 1? All dimensions are specified as height x width ?

Answer :   11

Q 2. Consider a two-layer neural network with identical weights W and bias b in each layer and activation function Sigmoid. If the input to the network is x, what is the sum of elements in the output vector? Round your answer to two decimal places

    

Answer  :      1.50

Q 3. If g(z) is the sigmoid function, which of the following is the correct expression for its derivative with respect to z

 g(z)

 1-g(z)

 g(z) * (1 - g(z))

 g(z) / (1 - g(z))

Answer:    g(z) * (1 - g(z)) 

Q 4. Which of the following logical expression(s) involving two variables can be implemented using a single perceptron with the sign function as activation function (sign(x) = 1 if x >= 0 and -1 otherwise)? ~ denotes the NOT operator.

 (x1 ^ x2)

 (x1 V x2)

 x1->x2

 x1 <-> x2

Answer:  ABC

Q 5. Suppose we are using a neural network to classify handwritten digits (0-9). Which of the following activation functions is most suitable at the output layer?

 Softmax

 ReLU

 Tanh

 Sigmoid

  

Answer :  A

Q 6. Which of the following statement(s) is(are) correct about fat + short vs. tall + thin neural networks?

 Fat + short neural networks can theoretically compute all possible continuous functions provided enough neurons are available

 There is a certain class of continuous functions that can only be computed by tall + thin neural networks and not fat + short ones no matter how many neurons are available

 For the same number of parameters, we should prefer tall + thin networks over fat + short ones

 For the same number of parameters, it does not matter whether we choose tall + thin networks or fat + short ones

Answer : AC

Q

1 point

     Q 7.What is the fundamental difference between classical ML (machine learning) and modern DL (deep learning)?

 DL involves engineering features by hand whereas classical ML discovers them automatically

 Classical ML involves engineering features by hand whereas DL discovers them automatically

 There is no fundamental difference between the two, DL is just classical ML with large neural networks

 Classical ML focuses on classification tasks whereas DL cannot be used for classification tasks

Answer :B

1 point

   Q 8. Which of the following is(are) the motivation behind using neural networks to approximate Q values instead of using                   a table to store these values?

 Neural networks can be used as an efficient form of storage for large tables

 Neural networks support efficient lookup of Q-values using GPUs

 Neural networks can infer values of new states similar to the ones it has already seen even when the Q-value of these new states are not explicitly stored

 When state spaces are large, it becomes infeasible to calculate the Q-value of each state-action pair and store them in a table

 

Answer :CD

1 point

   Q 9.Which of the following statement(s) is(are) true about DQNs for Atari game ?

 They learn to play the games by watching videos of expert human players

 The DQN only has access to the video frames on the game screen and the reward function. It cannot directly observe the mechanics of the game simulator.

 They learn to play the games by random exploration

 The DQN only considers one video frame of the game at a time when playing the game

 

Answer :BC

1 point

   Q 10.Which of the following statement(s) is(are) true for Deep Q-learning?

 It can be used for continuous control i.e. when the action and state spaces are continuous

 To prevent training instability an experience replay buffer is used to store experiences

 Stochastic gradient descent is used to optimize the DQNs?

 For convergence, DQNs must always follow a greedy policy determined by the current Q-values

Answer :ABC

1 point

Q 11.Which of the following is the most accurate characterization of the robustness issue in deep learning?

 Neural networks are highly sensitive to errors in numerical computation

 Neural networks are highly sensitive to small changes in the training dataset

 Neural networks can give completely incorrect predictions if tiny adversarial perturbations are added to the input

 Neural networks often give results that are sexist or racist even if the data has no gender or racial bias

Answer : C