MauryaHub PYQ Practice

da5006_2026T1_Q2_NA.pdf

Deep Learning for Computer Vision · Quiz 2 · Jan 2026

← Course papers · Start practice / exam

Questions and published explanations below are available without starting a test. Some questions may not have a published solution yet.

Question 2 MCQ · 1.0 marks

Why do very deep plain CNNs (without skip connections) sometimes show **higher training error** than shallower CNNs?
  1. The increased capacity leads to severe overfitting on the training set before convergence can be reached.
  2. Optimization becomes harder; skip connections make learning identity mappings easier and improve gradient flow.
  3. The receptive field becomes excessively large, causing the network to lose fine-grained spatial information.
  4. Saturated activations cause exploding gradients that cannot be mitigated by standard initialization techniques.

A published solution is not available for this question yet.

Question 3 MCQ · 1.0 marks

In a standard ResNet-50 bottleneck block, the three convolutions are typically:
  1. 3×3, 1×1, 3×3
  2. 1×1, 3×3, 1×1
  3. 1×1, 5×5, 1×1
  4. 3×3, 3×3, 3×3

A published solution is not available for this question yet.

Question 4 MCQ · 1.0 marks

In Inception/GoogLeNet modules, the main purpose of 1×1 convolutions is to:
  1. Increase the receptive field of subsequent convolutional layers without downsampling.
  2. Reduce channel dimensionality (bottleneck) and add non-linearity
  3. Act as a differentiable substitute for max-pooling operations to preserve exact spatial hierarchies.
  4. Project the feature maps into a higher-dimensional space to separate entangled features.

A published solution is not available for this question yet.

Question 5 MCQ · 1.0 marks

MobileNetV1 reduces computation primarily by using:
  1. Grouped convolutions followed by spatial pyramid pooling
  2. Depthwise separable convolutions (depthwise + pointwise)
  3. Asymmetric convolutions (e.g., factorizing a 3×3 into 3×1 followed by 1×3)
  4. Low-rank matrix factorization of dense classification layers

A published solution is not available for this question yet.

Question 6 MCQ · 1.0 marks

EfficientNet’s key scaling idea is:
  1. Scaling the network depth logarithmically while linearly increasing the input resolution
  2. Utilizing neural architecture search (NAS) to independently optimize depth, width, and resolution for each block
  3. Compound scaling of depth, width, and resolution using a single coefficient
  4. Adjusting the channel multiplier dynamically based on the target device's hardware constraints

A published solution is not available for this question yet.

Question 7 MCQ · 1.0 marks

You have a pretrained CNN backbone and only **500** labeled images for a new task. A strong first baseline is:
  1. Fine-tune the entire network immediately using a very small learning rate to preserve pre-trained features
  2. Freeze backbone, train a new classifier head; optionally unfreeze later with smaller LR
  3. Train a completely new architecture from scratch using heavy data augmentation to compensate for the lack of data
  4. Unfreeze the backbone from the bottom layers upwards, gradually increasing the learning rate

A published solution is not available for this question yet.

Question 8 MCQ · 1.0 marks

When fine-tuning a model, what is a common way to prevent the new training from "overwriting" or destroying the useful features already learned during pre-training?
  1. Applying a high dropout rate exclusively to the newly initialized classification head
  2. Use differential learning rates (small LR for the pre-trained backbone, larger LR for the new head)
  3. Re-initializing the weights of the last convolutional block before training
  4. Freezing the classification head and only fine-tuning the backbone features

A published solution is not available for this question yet.

Question 9 MCQ · 1.0 marks

Which method produces a heatmap by weighting convolutional feature maps using gradients of a target class score?
  1. Guided Backpropagation
  2. Grad-CAM
  3. SmoothGrad
  4. Integrated Gradients

A published solution is not available for this question yet.

Question 10 MCQ · 1.0 marks

Which detector is a classic **two-stage** detector with an RPN + ROI feature extraction?
  1. SSD
  2. YOLO
  3. Faster R-CNN
  4. DETR

A published solution is not available for this question yet.

Question 11 MCQ · 1.0 marks

Which statement is most accurate?
  1. Two-stage detectors optimize a single joint loss function for classification and localization, whereas single-stage uses decoupled losses.
  2. Single-stage detectors rely on selective search for region generation, whereas two-stage detectors use a dedicated sub-network.
  3. Two-stage detectors typically generate proposals then classify/refine; single- stage predicts boxes/classes densely.
  4. Two-stage detectors process the entire image in one pass, while single-stage detectors evaluate crops sequentially.

A published solution is not available for this question yet.

Question 12 MCQ · 1.0 marks

RetinaNet is notable primarily because it introduced:
  1. A novel feature pyramid network (FPN) architecture that eliminates the need for anchor boxes
  2. Focal loss to address class imbalance
  3. The concept of Region of Interest (RoI) pooling to align features extracted from different scales
  4. A dynamic routing mechanism that directly assigns bounding box predictions to ground truth objects

A published solution is not available for this question yet.

Question 13 MCQ · 1.0 marks

A model that predicts a class label for **every pixel** without separating object instances is:
  1. Panoptic segmentation
  2. Instance segmentation
  3. Semantic segmentation
  4. Image matting

A published solution is not available for this question yet.

Question 14 MCQ · 1.0 marks

Vanishing gradients in vanilla RNNs are largely caused by:
  1. The continuous addition of bias terms at each timestep which suppresses the gradient signal
  2. Repeated multiplication by Jacobians whose spectral norm is often < 1
  3. The use of unbounded activation functions like ReLU that cause activations to decay over time
  4. The inability of the standard cross-entropy loss function to distinguish between short-term and long-term dependencies

A published solution is not available for this question yet.

Question 15 MCQ · 1.0 marks

Which statement is true?
  1. Transformers process sequences sequentially but utilize parallelized loss computation to speed up training.
  2. The attention mechanism strictly limits the context window size, whereas RNNs have a theoretically infinite context window in practice.
  3. Self-attention enables direct interactions between any pair of tokens within a layer.
  4. Transformers inherently encode positional information in their feed-forward weights, eliminating the need for explicit sequence tracking.

A published solution is not available for this question yet.

Question 16 MSQ · 1.0 marks

Which statements can be true in practice when fine-tuning with small batch sizes?
  1. Freezing BatchNorm running statistics can improve stability.
  2. If the new dataset distribution differs, re-estimating BN stats may help.
  3. Setting BatchNorm to training mode with a batch size of 1 provides an unbiased estimate of the population variance.
  4. Calling .eval() on a BatchNorm layer disables the application of the learned affine parameters (γ, β).

A published solution is not available for this question yet.

Question 17 MSQ · 1.0 marks

Which statements are true?
  1. Smooth L1 (Huber) is often used for box regression due to robustness to outliers.
  2. Focal loss primarily addresses foreground/background (class) imbalance.
  3. Standard Cross-Entropy loss scales the loss of easy examples to zero more aggressively than Focal Loss.
  4. In two-stage detectors, the RPN has its own objectness + box regression losses.

A published solution is not available for this question yet.

Question 18 MSQ · 1.0 marks

Which statements are true ?
  1. Soft attention is differentiable and can be trained with backpropagation.
  2. Hard attention often needs sampling + REINFORCE (or similar) due to non- differentiability.
  3. Soft attention significantly reduces the computational complexity of the forward pass compared to Hard attention.
  4. Cross-attention can align decoder queries to encoder keys/values (e.g., words attending to image regions).

A published solution is not available for this question yet.

Question 19 NAT · 1.0 marks

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.

    A published solution is not available for this question yet.

    Question 20 NAT · 1.0 marks

    Using the same setup (Input feature map: 14×14×128, output: 14×14×256, kernel: 3×3, stride 1, same padding, ignore bias), compute total **MACs operation** (multiply-accumulates). A single MAC operation multiplies two numbers and adds the result to an accumulator.

      A published solution is not available for this question yet.

      Question 21 NAT · 1.0 marks

      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).

        A published solution is not available for this question yet.

        Question 22 NAT · 1.0 marks

        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.

          A published solution is not available for this question yet.

          Question 23 NAT · 1.0 marks

          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.

            A published solution is not available for this question yet.

            Question 24 NAT · 1.0 marks

            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.

              A published solution is not available for this question yet.