The IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) Quiz 2 paper sat on 16 Aug 2026, in the May 2026 term: 22 questions for 50 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.
Why do very deep plain CNNs (without skip connections) sometimes show higher training error than shallower CNNs?
They always overfit because parameter count increases.
Optimization becomes harder; skip connections make learning identity mappings easier and improve gradient flow.
BatchNorm guarantees this never happens.
Using sigmoid activations instead of ReLU always fixes it.
Correct answer
Optimization becomes harder; skip connections make learning identity mappings easier and improve gradient flow.
MobileNetV1 reduces computation primarily by using:
Dilated convolutions everywhere
Depthwise separable convolutions (depthwise + pointwise)
ROI pooling
Fully-connected layers instead of convolutions
Correct answer
Depthwise separable convolutions (depthwise + pointwise)
EfficientNet’s key scaling idea is:
Only scale depth while keeping width and resolution fixed
Only scale resolution while keeping depth and width fixed
Compound scaling of depth, width, and resolution using a single coefficient
Scale width but reduce depth to keep parameters constant
Correct answer
Compound scaling of depth, width, and resolution using a single coefficient
Which technique most directly helps reduce catastrophic forgetting early in fine-tuning?
Use very high weight decay on the backbone
Use differential learning rates (small LR for backbone, larger LR for head)
Replace ReLU with sigmoid
Remove skip connections
Correct answer
Use differential learning rates (small LR for backbone, larger LR for head)
Which method produces a heatmap by weighting convolutional feature maps using gradients of a target class score?
PCA
Grad-CAM
Histogram equalization
t-SNE
Correct answer
Grad-CAM
Which detector is a classic two-stage detector with an RPN + ROI feature extraction?
SSD
YOLO
Faster R-CNN
RetinaNet
Correct answer
Faster R-CNN
RetinaNet is notable primarily because it introduced:
ROI Align
Focal loss to address class imbalance
Depthwise separable convolutions
Self-attention
Correct answer
Focal loss to address class imbalance
A model that predicts a class label for every pixel without separating object instances is:
Object detection
Instance segmentation
Semantic segmentation
Captioning
Correct answer
Semantic segmentation
Which statement is true?
Transformers rely on recurrence to model sequences.
Transformers cannot handle variable-length sequences.
Self-attention enables direct interactions between any pair of tokens within a layer.
Transformers require optical flow for vision tasks.
Correct answer
Self-attention enables direct interactions between any pair of tokens within a layer.
In a standard ResNet-50 bottleneck block, the three convolutions are typically:
3×3, 3×3, 3×3
1×1, 3×3, 1×1
5×5, 3×3, 1×1
1×1, 1×1, 3×3
Correct answer
1×1, 3×3, 1×1
In Inception/GoogLeNet modules, the main purpose of 1×1 convolutions is to:
Increase spatial resolution
Reduce channel dimensionality (bottleneck) and add non-linearity
Replace pooling
Implement residual learning
Correct answer
Reduce channel dimensionality (bottleneck) and add non-linearity
Which statement is most accurate?
Two-stage detectors do not use CNN backbones.
Single-stage detectors cannot regress bounding boxes.
Two-stage detectors typically generate proposals then classify/refine
Single-stage detectors require ROI pooling.
Correct answer
Two-stage detectors typically generate proposals then classify/refine
Vanishing gradients in vanilla RNNs are largely caused by:
Too much data augmentation
Repeated multiplication by Jacobians whose spectral norm is often < 1
Using convolutions before the RNN
Using attention layers
Correct answer
Repeated multiplication by Jacobians whose spectral norm is often < 1
Which statements can be true in practice when fine-tuning with small batch sizes?
Freezing BatchNorm running statistics can improve stability.
If the new dataset distribution differs, re-estimating BN stats may help.
BN always improves performance even with batch size 1.
Setting BN to eval disables learned affine parameters (γ, β).
Correct answers
Freezing BatchNorm running statistics can improve stability.
If the new dataset distribution differs, re-estimating BN stats may help.
Which statements are true ?
Smooth L1 (Huber) is often used for box regression due to robustness to outliers.
Focal loss primarily addresses foreground/background (class) imbalance.
IoU loss cannot be used for box regression.
In two-stage detectors, the RPN has its own objectness + box regression losses.
Correct answers
Smooth L1 (Huber) is often used for box regression due to robustness to outliers.
Focal loss primarily addresses foreground/background (class) imbalance.
In two-stage detectors, the RPN has its own objectness + box regression losses.
Which statements are true?
Soft attention is differentiable and can be trained with backpropagation.
Hard attention often needs sampling + REINFORCE (or similar) due to non- differentiability.
Hard attention always has higher test-time compute than soft attention.
Cross-attention can align decoder queries to encoder keys/values (e.g., words attending to image regions).
Correct answers
Soft attention is differentiable and can be trained with backpropagation.
Hard attention often needs sampling + REINFORCE (or similar) due to non- differentiability.
Cross-attention can align decoder queries to encoder keys/values (e.g., words attending to image regions).
Input feature map: 14×14×128, output: 14×14×256, kernel: 3×3, stride 1, same padding, ignore bias.
Compute number of parameters for a standard convolution layer.
Correct answer: 294912
Input feature map: 14×14×128, output: 14×14×256, kernel: 3×3, stride 1, same padding, ignore bias, compute total MACs (multiply-accumulates).
Correct answer: 57802752
A ResNet bottleneck block takes 256 channels in and uses:
1×1 conv to 256 channels → 3×3 conv at 256 → 1×1 conv to 1024 channels. Ignore bias.
Compute total parameters in these three convolutions (exclude projection shortcut).
Correct answer: 917504
A detector uses three pyramid levels: P3 = 80×80, P4 = 40×40, P5 = 20×20. At each location, it uses 9 anchors (3 scales × 3 aspect ratios).
Compute total anchors per image.
Correct answer: 75600
Box A: top-left (2,2), bottom-right (10,12) Box B: top-left (6,5), bottom-right (14,15) Compute IoU rounded to 3 decimals.
Correct answer: 0.212 (accepted within ±0.001)
Let query q = [1,2]^T and keys:k1 = [1,0]^T, k2 = [0,1]^T, k3 = [1,1]^T. Scores si = q^T ki. Use softmax over scores. Use exp(1)=2.72, exp(2)=7.39, exp(3)=20.09. What is α3 (the attention weight for k3)? Provide a value in a reasonable range.
Correct answer: 0.645 (accepted within ±0.045)
