Deep Learning – IIT Ropar Week 2 Assignment Answers

2 months ago
Sanket Kumar
2 minutes
Deep Learning - IIT Ropar

Deep Learning – IIT Ropar Week 2 Assignment Answers (Jan-Apr 2026)

Course Link : Click Here

  1. Deep Learning – IIT Ropar Week 3

1. Which property of the sigmoid function makes it suitable for modeling fraud probability?

  • It produces only binary outputs
  • It is non-monotonic
  • It is continuous and differentiable
  • It does not depend on model parameters
Answer : c

2. For transaction T2, compute the value of z. :

Answer : 0.6

3. The decision rule p^≥0.5 is equivalent to which condition?

  • z≥1
  • z≥0
  • z≤0
  • z≤−1
Answer : b

4. Which transactions trigger step-up authentication?
(Select all that apply.)

  • T1
  • T2
  • T3
  • T4
Answer : b, c, d

5. Holding x2constant, increasing b shifts the sigmoid activation curve along the x1 -axis in which direction?

  • Right
  • Left
  • Up
  • Down
Answer : b

6. What is the objective of training this model?

  • Maximize the loss
  • Minimize the loss
  • Keep the loss constant
  • None of the above
Answer : b

7. If y=1 and p^=0.2, what is the loss?

  • 0.4
  • 0.5
  • 0.32
  • 0.8
Answer : c

8. Which statement about the sigmoid output is correct?

  • It can be interpreted as a probability
  • It always equals 0 or 1
  • It is undefined for negative inputs
  • It is non-differentiable
Answer : a

9. Fill in the blank

If w⊤x+b=0, then p^=_______.

Answer : 0.5

10. Which statements are correct for this learning setup?

  • Model parameters are learned from labeled data
  • Gradient descent can be used
  • Sigmoid activation is differentiable
  • Outputs are always binary
  • Outputs are always binaryLoss is minimized during training
Answer : a, b, c, e

11. How many trainable weights connect the input layer to the hidden layer (excluding biases)?

  • 9
  • 20
  • 25
  • 10
Answer : b

12. How many bias parameters exist in the hidden layer?

  • 1
  • 4
  • 5
  • 0
Answer : b

13. How many trainable weights connect the hidden layer to the output neuron (excluding bias)?

  • 1
  • 4
  • 5
  • 20
Answer : b

14. Why is a sigmoid activation function used in the hidden layer?

  • It produces binary outputs
  • It guarantees exact rank prediction
  • It is differentiable and supports gradient-based learning
  • It removes the need for a loss function
Answer : c

15. Which statements about representational capacity are correct?

  • A perceptron network with one hidden layer can represent any Boolean function
  • A single sigmoid neuron can represent XOR exactly
  • A sigmoid network with one hidden layer can approximate any continuous function (given enough neurons)
  • The approximation guarantee holds with a fixed small number of neurons
  • Increasing hidden neurons can increase approximation accuracy
Answer : a, c, e

For other NPTEL Assignment Answer : Click Here