The IIT Madras BS Deep Learning (Deep Learning) Quiz 1 paper sat on 26 Oct 2025, in the September 2025 term, set 2: 22 questions for 44 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.
THIS IS QUESTION PAPER FOR THE SUBJECT "DEGREE LEVEL : DEEP LEARNING (COMPUTER BASED EXAM)" ARE YOU SURE YOU HAVE TO WRITE EXAM FOR THIS SUBJECT? CROSS CHECK YOUR HALL TICKET TO CONFIRM THE SUBJECTS TO BE WRITTEN. (IF IT IS NOT THE CORRECT SUBJECT, PLS CHECK THE SECTION AT THE TOP FOR THE SUBJECTS REGISTERED BY YOU)
YES
NO
Correct answer
YES
Correct answer
A → 0.01 , B → 0.001 , C → 1.0
A → 0.001 , B → 0.01 , C → 1.0
A → 1.0 , B → 0.001 , C → 0.01
A → 0.01 , B → 1.0 , C → 0.001
Correct answer
A → 0.001 , B → 0.01 , C → 1.0
Which of the following statements best describes the primary representational advantage of a Multi-Layer Perceptron (MLP) over a Single-Layer Perceptron (SLP)?
An MLP can learn non-linear decision boundaries, allowing it to classify datathat is not linearly separable.
An MLP converges faster than an SLP on linearly separable data because itshidden layers accelerate learning.
An MLP is less prone to overfitting than an SLP because it has moreparameters to capture the data distribution.
An MLP can only be used for classification tasks, whereas an SLP can be usedfor both classification and regression.
Correct answer
An MLP can learn non-linear decision boundaries, allowing it to classify datathat is not linearly separable.
A data scientist is training a deep learning model on a massive dataset containing billions of data points. They find that using traditional Batch Gradient Descent (GD) is computationally infeasible. What is the primary reason that makes Batch GD impractical for this task, thus requiring an alternative like Stochastic Gradient Descent (SGD)?
It is impossible to fit the entire large dataset into the computer’s memory(RAM) at once.
The gradient calculated by Batch GD is often too noisy and inaccurate, leadingto poor convergence.
Batch GD is guaranteed to get stuck in sharp local minima, whereas SGD canescape them.
Batch GD requires calculating the gradient of the loss function with respect tothe parameters over the entire dataset for a single update, which is computationally very expensive.
Correct answer
Batch GD requires calculating the gradient of the loss function with respect tothe parameters over the entire dataset for a single update, which is computationally very expensive.
Correct answer
You are adapting a neural network that was originally designed for a 10-class image classification problem to perform a regression task, specifically to predict the price of a car. Which of the following modifications are essential or standard practice for this conversion?
Change the activation function of the final output layer to a linear function.
Set the number of neurons in the output layer to one.
Change the loss function from Categorical Cross-Entropy to Mean SquaredError (MSE).
Use accuracy as the primary metric to evaluate the model’s performance.
Keep the softmax activation function in the output layer to scale the predictedprice.
Correct answers
Change the activation function of the final output layer to a linear function.
Set the number of neurons in the output layer to one.
Change the loss function from Categorical Cross-Entropy to Mean SquaredError (MSE).
A written answer, not marked automatically.
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
What is the value of the output of the hidden layer? (Answer correct upto two digits after the decimal)
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
What is the value of the cross entropy loss? (use natural log). (Answer correct upto two digits after the decimal)
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
Use cross entropy loss and compute the gradient of W1 which is the weight between input and hidden layer. (consider upto two digits after the decimal for all the calculations)
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
Correct answer
Based on the above data, answer the given subquestions.
Correct answer
Based on the above data, answer the given subquestions.
Correct answer
Based on the above data, answer the given subquestions.
Correct answer
Suppose you build a neural network with one hidden layer that uses sigmoid as an activation function and an output layer with a softmax activation function. You initialize the hidden layer weights W1 randomly and set the output layer weights W2 to all zeros. Assume the bias is set to be zero. You train the network using stochastic gradient descent. Based on the above data, answer the given subquestions.
After one iteration of gradient descent, will the new weight for W1 be the same as the previous weight?
Yes
No
Correct answer
Yes
Suppose you build a neural network with one hidden layer that uses sigmoid as an activation function and an output layer with a softmax activation function. You initialize the hidden layer weights W1 randomly and set the output layer weights W2 to all zeros. Assume the bias is set to be zero. You train the network using stochastic gradient descent. Based on the above data, answer the given subquestions.
After one iteration of gradient descent, will the new weight for W2 be the same as the previous weight?
Yes
No
Correct answer
No
Based on the above data, answer the given subquestions.
A written answer, not marked automatically.
Based on the above data, answer the given subquestions.
A written answer, not marked automatically.
Suppose you build a neural network with one hidden layer that uses sigmoid as an activation function and an output layer with a softmax activation function. You initialize the hidden layer weights W1 randomly and set the output layer weights W2 to all zeros. Assume the bias is set to be zero. You train the network using stochastic gradient descent. Based on the above data, answer the given subquestions.
A written answer, not marked automatically.
