Question 1
Based on the above data, answer the given subquestions.
Deep Learning Practice End Term: 13 September 2026, Set 1 (May 2026 term)
The IIT Madras BS Deep Learning Practice (Deep Learning Practice) End Term paper sat on 13 Sept 2026, in the May 2026 term, set 1: 31 questions for 50 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
- 31
- 50
- 180 min
- 18
- 4
- 9
Show answer
Correct answer: 16384
Question 2
Based on the above data, answer the given subquestions.
Show answer
Correct answer
Question 3
Based on the above data, answer the given subquestions.
Using the actual flattened feature size from previous question number 2, how many learnable parameters (weights + bias) would a fully connected layer with 10 outputs have?
Show answer
Correct answer: 163850
Question 4
Based on the above data, answer the given subquestions.
Which of the following statements about the (corrected) model are true?
Show answer
Correct answers
Question 5
Show answer
Correct answer: 30
Question 6
Show answer
Correct answer: 60
Question 7
Which layer is responsible for the decoder failing to reach the target 120×120 output, and why?
None of the layers — the decoder already produces a 120×120 output.
Show answer
Correct answer
Question 8
Show answer
Correct answer: 240
Question 9
Show answer
Correct answer: 18432
Question 10
Show answer
Correct answer: 2336
Question 11
What is the parameter compression ratio (standard ÷ depthwise-separable), rounded to 2 decimal places?
Show answer
Correct answer: 7.9 (accepted within ±0.3)
Question 12
The pointwise layer uses a 1×1 kernel to mix information across channels after the depthwise step.
Replacing a standard convolution with a depthwise-separable convolution always improves model accuracy.
The compression ratio achieved depends on both the kernel size and the number of output channels.
Show answer
Correct answers
The pointwise layer uses a 1×1 kernel to mix information across channels after the depthwise step.
The compression ratio achieved depends on both the kernel size and the number of output channels.
Question 13
Based on the above data, answer the given subquestions.
Which of the following statements correctly identify a real bug in the given code ?
Show answer
Correct answers
Question 14
Based on the above data, answer the given subquestions.
No effect — BatchNorm behaves identically in train and eval mode.
BatchNorm continues to use statistics from the current inference batch instead of the stored running statistics, so predictions can differ from proper evaluation-mode inference and can depend on batch composition.
The prediction will be correct but computed twice as slowly.
Show answer
Correct answer
BatchNorm continues to use statistics from the current inference batch instead of the stored running statistics, so predictions can differ from proper evaluation-mode inference and can depend on batch composition.
Question 15
Based on the above data, answer the given subquestions.
Show answer
Correct answer: 100
Question 16
A convolutional layer receives a feature map with 16 input channels and applies 32 filters, each of spatial size 5 × 5. Following the "generic level of CNN" view (a filter spans all input channels), calculate the total number of weights (excluding bias) in this layer.
Show answer
Correct answer: 12800
Question 17
A convolutional layer is applied to a 32 × 32 input using a 5 × 5 filter with zero-padding of 2 and stride 2. Using the standard convolution output-size formula, calculate the output width (one spatial dimension). The height is the same; enter one integer only.
Show answer
Correct answer: 16
Question 18
The final classification head applies a softmax over the pre-activation scores (logits). For a 3-class problem, the logits produced for a particular image are [3, 1, 0]. Compute the softmax probability assigned to the first class (round to 2 decimal places).
Show answer
Correct answer: 0.84 (accepted within ±0.01)
Question 19
Following the VGG design philosophy of replacing a single large-kernel convolution with a cascade of small 3 × 3, stride-1 convolutions, how many such 3 × 3 layers must be stacked so that the resulting stack has the same effective (one-dimensional) receptive field as a single 11 × 11 convolution (the kernel size used in AlexNet's first layer)?
Show answer
Correct answer: 5
Question 20
Show answer
Correct answer: 7840
Question 21
The decoder of a depth-estimation network upsamples feature maps using a transposed convolution ("up-convolution") with kernel size 4, stride 2, and padding 1. If the input feature map has spatial size 30 × 30, what is the output height? The width is the same; enter one integer only.
Show answer
Correct answer: 60
Question 22
In the original U-Net used for dense prediction (and adaptable to depth estimation), each encoder stage applies two successive 3×3 convolutions with no padding (stride 1), followed by a 2×2 max-pooling with stride 2. If the input feature map to one such encoder stage has spatial size 140 × 140, what is the output height of the feature map after both convolutions and the max-pooling of that stage? The width is the same; enter one integer only.
Show answer
Correct answer: 68
Question 23
The SRGAN discriminator is a stack of 8 convolutional layers, all using 3 × 3 kernels with padding 1, and strides alternating as s = 1, 2, 1, 2, 1, 2, 1, 2 across the eight layers (the stride-1 layers preserve spatial size; the stride-2 layers downsample). If the input image is 96 × 96, what is the output height of the feature map after the last (8th) convolutional layer, just before the dense layers? The width is the same; enter one integer only.
Show answer
Correct answer: 6
Question 24
Efficient restoration networks (e.g., LaKDNet) replace standard convolutions with depth-wise separable convolutions (a depth-wise convolution followed by a point-wise 1×1 convolution). An input feature map of size 32 × 32 × 16 (16 channels) is processed by a depth-wise separable convolution that uses 3 × 3 depth-wise kernels and produces 32 output channels, with stride 1 and padding 1 (so the spatial size is preserved). Calculate the total number of multiplications for the full depth-wise separable operation (depth-wise + point-wise). Give the answer in millions (e.g., if 1,234,567 write 1.23).
Show answer
Correct answer: 0.67 (accepted within ±0.01)
Question 25
Compared with the sigmoid activation, which of the following are correct reasons for preferring the ReLU activation in deep CNNs?
For positive inputs, ReLU does not saturate, so it avoids the gradient-killing that occurs in the flat tails of the sigmoid.
ReLU is preferred because it bounds every activation to the interval [0, 1], preventing activations from growing large.
ReLU is cheaper to evaluate — a simple thresholding at zero — whereas sigmoid/tanh require expensive exponentials.
The sigmoid's derivative peaks at only 0.25, so multiplying many such small factors across layers pushes gradients toward zero (vanishing gradient).
Show answer
Correct answers
For positive inputs, ReLU does not saturate, so it avoids the gradient-killing that occurs in the flat tails of the sigmoid.
ReLU is cheaper to evaluate — a simple thresholding at zero — whereas sigmoid/tanh require expensive exponentials.
The sigmoid's derivative peaks at only 0.25, so multiplying many such small factors across layers pushes gradients toward zero (vanishing gradient).
Question 26
VGGNet replaces large convolution kernels with stacks of small 3 × 3 convolutions. Which of the following statements correctly describe the advantages of this design choice?
A stack of three 3 × 3 conv layers (stride 1) has the same effective receptive field as one 7 × 7 conv layer.
For C channels per layer, three stacked 3 × 3 layers use fewer parameters than a single 7 × 7 layer.
A single 3 × 3 convolution by itself already spans the same receptive field as a 7 × 7 convolution, so stacking is unnecessary.
Inserting a ReLU non-linearity between successive 3 × 3 layers makes the composite function more discriminative than a single linear projection over the same region.
Show answer
Correct answers
A stack of three 3 × 3 conv layers (stride 1) has the same effective receptive field as one 7 × 7 conv layer.
For C channels per layer, three stacked 3 × 3 layers use fewer parameters than a single 7 × 7 layer.
Inserting a ReLU non-linearity between successive 3 × 3 layers makes the composite function more discriminative than a single linear projection over the same region.
Question 27
Which of the following statements correctly describe the Region Proposal Network (RPN) in Faster R-CNN?
The RPN is fully convolutional and shares its convolutional feature maps with the downstream detection network.
The RPN relies on an external Selective Search module to generate its region proposals.
At each sliding-window location, the RPN predicts objectness scores and box refinements relative to k predefined anchor boxes.
Anchors of multiple scales and aspect ratios allow the RPN to propose regions for objects of very different sizes and shapes.
Show answer
Correct answers
The RPN is fully convolutional and shares its convolutional feature maps with the downstream detection network.
At each sliding-window location, the RPN predicts objectness scores and box refinements relative to k predefined anchor boxes.
Anchors of multiple scales and aspect ratios allow the RPN to propose regions for objects of very different sizes and shapes.
Question 28
Consider the multi-task loss used to train the original YOLOv1 model. Which of the following statements about it are correct?
The loss predicts the square roots of width and height so that a fixed absolute error is penalized more for small boxes than for large boxes.
The hyperparameters are set to λ_coord = 5 (upweighting localization error) and λ_noobj = 0.5 (downweighting confidence error for cells with no object).
The classification term for a cell is only penalized when an object is actually present in that grid cell.
For the predictor "responsible" for an object, the target confidence value is fixed at 1, independent of the box's IoU with the ground truth.
Show answer
Correct answers
The loss predicts the square roots of width and height so that a fixed absolute error is penalized more for small boxes than for large boxes.
The hyperparameters are set to λ_coord = 5 (upweighting localization error) and λ_noobj = 0.5 (downweighting confidence error for cells with no object).
The classification term for a cell is only penalized when an object is actually present in that grid cell.
Question 29
Regarding the channel-attention module in CBAM, which of the following are correct?
It exploits the inter-channel relationship of features, learning "what" is meaningful by re-weighting feature channels.
It produces a spatial attention map of size H × W × 1 that highlights where to focus, discarding all channel information.
The spatial dimensions of the input feature map are squeezed using both average-pooling and max-pooling, producing two channel descriptors.
The pooled descriptors are passed through a shared MLP, combined, and a sigmoid produces the channel attention weights that rescale the input feature map.
Show answer
Correct answers
It exploits the inter-channel relationship of features, learning "what" is meaningful by re-weighting feature channels.
The spatial dimensions of the input feature map are squeezed using both average-pooling and max-pooling, producing two channel descriptors.
The pooled descriptors are passed through a shared MLP, combined, and a sigmoid produces the channel attention weights that rescale the input feature map.
Question 30
SRGAN performs photo-realistic super-resolution using a GAN. Which of the following statements are correct?
Its perceptual loss combines a content loss and an adversarial loss, rather than relying on pixel-wise MSE alone.
The content loss is computed on feature maps of a pre-trained VGG network (perceptual similarity), not purely in pixel space.
Purely MSE-optimized solutions tend to be overly smooth because they approximate a pixel-wise average of many plausible HR solutions, losing high-frequency texture.
The adversarial term makes SRGAN maximize PSNR, which is why it always achieves a higher PSNR than the MSE-based SRResNet.
Show answer
Correct answers
Its perceptual loss combines a content loss and an adversarial loss, rather than relying on pixel-wise MSE alone.
The content loss is computed on feature maps of a pre-trained VGG network (perceptual similarity), not purely in pixel space.
Purely MSE-optimized solutions tend to be overly smooth because they approximate a pixel-wise average of many plausible HR solutions, losing high-frequency texture.
Question 31
In the multi-scale deep network for single-image depth estimation (Eigen et al.), the architecture is split into a coarse network and a fine network. Which statement best explains their respective roles?
The coarse network uses a large receptive field (via deeper layers and fully- connected layers) to predict the global scene depth, while the fine network refines this prediction locally using input-image details.
The fine network predicts the global scene structure, and the coarse network only sharpens edges.
Both networks are architecturally identical and their outputs are simply averaged (ensembled).
The coarse network processes each pixel independently and therefore ignores global scene context.
Show answer
Correct answer
The coarse network uses a large receptive field (via deeper layers and fully- connected layers) to predict the global scene depth, while the fine network refines this prediction locally using input-image details.