The IIT Madras BS Deep Learning (Deep Learning) Quiz 2 paper sat on 23 Nov 2025, in the September 2025 term: 15 questions for 40 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.
Choose
, since it gives the lowest training loss and allows the model to fit the data best.
Choose
, as it provides a reasonable compromise between model complexity and generalization.
Choose
, since it yields the lowest validation loss and achieves the best bias–variance trade-off.
Choose
, as it produces a highly regularized model that minimizes the risk of overfitting.
Correct answer
Choose
, since it yields the lowest validation loss and achieves the best bias–variance trade-off.
Consider a neural network as shown in the figure with 2 input neurons, a hidden layer with 2 neurons (ReLU), and 1 output neuron (sigmoid).
Weights are:
with no biases. Dropout probability is 0.5 on the hidden layer.
If the input vector
is fed to this network, what is the expected output while testing?
0.77
1.25
2.5
0.92
Correct answer
0.77
Which of the following statements are true about a Convolutional (CONV) layer?
The number of parameters depends on the depth of the input.
The number of filters determines the number of output channels.
The total number of parameters depends on the stride.
The total number of parameters depends on the padding.
Correct answers
The number of parameters depends on the depth of the input.
The number of filters determines the number of output channels.
Which of the following optimizers adapts the learning rate for each parameter during training?
Stochastic Gradient Descent
Nesterov Accelerated Gradient Descent
Adagrad
RMSProp
Correct answers
Adagrad
RMSProp
Which of the following techniques can be used to reduce model overfitting?
Data augmentation
Dropout
L2 regularization
Using Adam instead of SGD
Correct answers
Data augmentation
Dropout
L2 regularization
Why is bias correction needed in the Adam optimizer?
To prevent the learning rate from becoming too large in the early iterations.
To correct the underestimated moving averages of the first and second moments at the beginning of training.
To increase the momentum effect for faster convergence.
To ensure the learning rate remains constant throughout training.
Correct answer
To correct the underestimated moving averages of the first and second moments at the beginning of training.
After training a linear regression model on a large training set of size
, it achieves a training
error of
. Analysis of the residual plot shows a clear non-linear pattern, suggesting the model is underfitting the data. Which two of the following modifications are most likely to improve the model’s performance by increasing its capacity to capture non-linear relationships?
Adding a regularization term (such as L2 or L1 penalty) to the mean squared error loss function.
Using a 2-hidden layer feedforward network with ReLU activation functions in place of linear regression.
Applying a polynomial feature transformation of degree
to the input variables.
Standardizing all training samples to have mean zero and unit variance.
Using a 5-hidden layer feedforward network without non-linear activation functions in place of linear regression.
Correct answers
Using a 2-hidden layer feedforward network with ReLU activation functions in place of linear regression.
Applying a polynomial feature transformation of degree
to the input variables.
Consider an Inception module in GoogLeNet with an input feature map of size
. The module has four branches:
•
Branch 1:
convolution with 64 filters.
•
Branch 2:
convolution with 96 filters followed by
convolution with 128 filters.
•
Branch 3:
convolution with 16 filters followed by
convolution with 32 filters.
•
Branch 4:
max pooling followed by
convolution with 32 filters.
Assume that all convolutions use no padding, stride 1, and that biases are ignored. What is the total number of learnable parameters in this module?
Correct answer: 163328
Consider a fully connected neural network as follows:
•
One input neuron
•
A single hidden layer with two neurons
•
One output neuron.
The network uses the ReLU activation function in both the hidden and output layers. The loss function used is the squared error:
where
is the true target and
is the network output.
Let the initial weights of the network be
.
Based on the above data, answer the given subquestions.
Choose the option which can represent the loss for single input data.
, if
, if
, if
None of these
Correct answers
, if
, if
Consider a fully connected neural network as follows:
•
One input neuron
•
A single hidden layer with two neurons
•
One output neuron.
The network uses the ReLU activation function in both the hidden and output layers. The loss function used is the squared error:
where
is the true target and
is the network output.
Let the initial weights of the network be
.
Based on the above data, answer the given subquestions.
Select the gradient of the loss function with respect to weight
from the following.
Note:
represents the indicator function.
Correct answer
Consider a fully connected neural network as follows:
•
One input neuron
•
A single hidden layer with two neurons
•
One output neuron.
The network uses the ReLU activation function in both the hidden and output layers. The loss function used is the squared error:
where
is the true target and
is the network output.
Let the initial weights of the network be
.
Based on the above data, answer the given subquestions.
Suppose the network uses the ELU activation function with
in the hidden layer and ReLU activation function in the output layer. Using the same initial weights, compute the updated value
of
after one step of stochastic gradient descent on the data point
with a
learning rate of
. Enter the answer correct to one decimal place.
Hint: ELU activation function is given by
Correct answer: -0.4
Based on the above data,answer the given subquestion.
What will be the input to the Pooling Layer.
Correct answer
Based on the above data,answer the given subquestion.
Find the output after average pooling. Answer correct up to 2 digits after the decimal.
Correct answer: 2.25 (accepted within ±0.03)
Based on the above data,answer the given subquestion.
If
, compute
and submit the answer correct up to 2 digits after the decimal
Correct answer: 4 (accepted within ±0.2)
Based on the above data,answer the given subquestion.
If
, compute
. where
is the top left element of the filter. Submit the answer correct up to two digits after the decimal.
Correct answer: 2 (accepted within ±0.1)
