The IIT Madras BS Deep Learning (Deep Learning) Quiz 2 paper sat on 2 Apr 2023, in the January 2023 term, set QPE2: 17 questions for 45 marks in 120 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
AdaGrad takes lesser number of iterations than vanilla Gradient Descent toconverge
AdaGrad takes more number of iterations than vanilla Gradient Descent toconverge
Both the algorithms will converge in equal number of iterations
Both the algorithms will keep oscillating around the minimum
Correct answer
AdaGrad takes more number of iterations than vanilla Gradient Descent toconverge
True, due to the presence of noise in the target
True, due to the high variance of the prediction model
False, due to the high variance of the prediction model
False, due to zero mean of the noise added to the target
Correct answer
False, due to the high variance of the prediction model
Which of the following statement(s) is (are) true about adaptive gradient descent algorithms?
They are more suitable for sparse data
They are more suitable for dense data
They add additional computational complexity to the vanilla gradient descentalgorithm
They are guranteed to converge to global minimum
Correct answers
They are more suitable for sparse data
They add additional computational complexity to the vanilla gradient descentalgorithm
Consider the assertion and the corresponding reason,
Assertion (A): The effective learning rate in RMSprop may increase or decrease Reason (R): This is due to exponential averaging of squared gradients
A is true but R is not the correct reason
A is True and R is the correct reason
A is False and therefore R
Correct answer
A is True and R is the correct reason
Suppose that we modify the update equation of ADAM optimizer as follows
Suppose that we modify the update equation of ADAM optimizer as follows
TRUE
FALSE
Correct answer
FALSE
High bias
High variance
Low bias
Low variance
Correct answers
High bias
Low variance
Suppose that we have a deep neural network with sigmoid neurons at all layers. Which of the following approaches, in general, help the network to overcome the vanishing gradient problem?
Replace sigmoid neuron by tanh
Replace sigmoid neuron by GELU
Use unsupervised layerwise pretraining and finally train the entire networkwith supervised objective
Use supervised training one-layer at a time with sigmoid functions andgradually add layers
Use Residual connections
Constrain the weight values with an appropriate regularization technique
Correct answers
Replace sigmoid neuron by GELU
Use unsupervised layerwise pretraining and finally train the entire networkwith supervised objective
Use Residual connections
Constrain the weight values with an appropriate regularization technique
An input image of dimension 55 × 55 × 3 is to be convolved with 10 filters (kernels) of spatial dimension 5 × 5. In all the sub-questions, assume a bias term for kernels.
Suppose that we pad zeros of size 1 (p = 1) around the sides of the image (in all channels). Assume stride 2 for all 10 convolution filters on the input image. All the 10 filters have bias associated with it. What is the dimension of the output volume? Enter the number of elements in the output volume.
Correct answer: 7290
An input image of dimension 55 × 55 × 3 is to be convolved with 10 filters (kernels) of spatial dimension 5 × 5. In all the sub-questions, assume a bias term for kernels.
How many number of parameters are there in the convolution layer?
Correct answer: 760
An input image of dimension 55 × 55 × 3 is to be convolved with 10 filters (kernels) of spatial dimension 5 × 5. In all the sub-questions, assume a bias term for kernels.
Correct answer: 8071
An input image of dimension 55 × 55 × 3 is to be convolved with 10 filters (kernels) of spatial dimension 5 × 5. In all the sub-questions, assume a bias term for kernels.
Suppose we applied 1D convolution operation with 10 kernels on the output of the first convolution layer and then applied the rest of the operations as in the previous question (flattening followed by fully connected layer), then the number of parameters in the network would have
Increased
Decreased
Remained same
Insufficient information
Correct answer
Increased
Correct answer: 12
Correct answer: 10.0
Correct answer: 10.0
Correct answer: 4
Enter the sum of updated bias (exclude weights).
Correct answer: -3
Correct answer: 0
