Deep Learning for Computer Vision Quiz 2: 12 April 2026 (January 2026 term)

1. A

   The increased capacity leads to severe overfitting on the training set before convergence can be reached.

2. B

   Optimization becomes harder; skip connections make learning identity mappings easier and improve gradient flow.

3. C

   The receptive field becomes excessively large, causing the network to lose fine-grained spatial information.

4. D

   Saturated activations cause exploding gradients that cannot be mitigated by standard initialization techniques.

1. A

   3×3, 1×1, 3×3

2. B

   1×1, 3×3, 1×1

3. C

   1×1, 5×5, 1×1

4. D

   3×3, 3×3, 3×3

1. A

   Increase the receptive field of subsequent convolutional layers without downsampling.

2. B

   Reduce channel dimensionality (bottleneck) and add non-linearity

3. C

   Act as a differentiable substitute for max-pooling operations to preserve exact spatial hierarchies.

4. D

   Project the feature maps into a higher-dimensional space to separate entangled features.

1. A

   Grouped convolutions followed by spatial pyramid pooling

2. B

   Depthwise separable convolutions (depthwise + pointwise)

3. C

   Asymmetric convolutions (e.g., factorizing a 3×3 into 3×1 followed by 1×3)

4. D

   Low-rank matrix factorization of dense classification layers

1. A

   Scaling the network depth logarithmically while linearly increasing the input resolution

2. B

   Utilizing neural architecture search (NAS) to independently optimize depth, width, and resolution for each block

3. C

   Compound scaling of depth, width, and resolution using a single coefficient

4. D

   Adjusting the channel multiplier dynamically based on the target device's hardware constraints

1. A

   Fine-tune the entire network immediately using a very small learning rate to preserve pre-trained features

2. B

   Freeze backbone, train a new classifier head; optionally unfreeze later with smaller LR

3. C

   Train a completely new architecture from scratch using heavy data augmentation to compensate for the lack of data

4. D

   Unfreeze the backbone from the bottom layers upwards, gradually increasing the learning rate

1. A

   Applying a high dropout rate exclusively to the newly initialized classification head

2. B

   Use differential learning rates (small LR for the pre-trained backbone, larger LR for the new head)

3. C

   Re-initializing the weights of the last convolutional block before training

4. D

   Freezing the classification head and only fine-tuning the backbone features

1. A

   Guided Backpropagation

2. B

   Grad-CAM

3. C

   SmoothGrad

4. D

   Integrated Gradients

1. A

   SSD

2. B

   YOLO

3. C

   Faster R-CNN

4. D

   DETR

1. A

   Two-stage detectors optimize a single joint loss function for classification and localization, whereas single-stage uses decoupled losses.

2. B

   Single-stage detectors rely on selective search for region generation, whereas two-stage detectors use a dedicated sub-network.

3. C

   Two-stage detectors typically generate proposals then classify/refine; single- stage predicts boxes/classes densely.

4. D

   Two-stage detectors process the entire image in one pass, while single-stage detectors evaluate crops sequentially.

1. A

   A novel feature pyramid network (FPN) architecture that eliminates the need for anchor boxes

2. B

   Focal loss to address class imbalance

3. C

   The concept of Region of Interest (RoI) pooling to align features extracted from different scales

4. D

   A dynamic routing mechanism that directly assigns bounding box predictions to ground truth objects

1. A

   Panoptic segmentation

2. B

   Instance segmentation

3. C

   Semantic segmentation

4. D

   Image matting

1. A

   The continuous addition of bias terms at each timestep which suppresses the gradient signal

2. B

   Repeated multiplication by Jacobians whose spectral norm is often < 1

3. C

   The use of unbounded activation functions like ReLU that cause activations to decay over time

4. D

   The inability of the standard cross-entropy loss function to distinguish between short-term and long-term dependencies

1. A

   Transformers process sequences sequentially but utilize parallelized loss computation to speed up training.

2. B

   The attention mechanism strictly limits the context window size, whereas RNNs have a theoretically infinite context window in practice.

3. C

   Self-attention enables direct interactions between any pair of tokens within a layer.

4. D

   Transformers inherently encode positional information in their feed-forward weights, eliminating the need for explicit sequence tracking.

Select all that apply.

1. A

   Freezing BatchNorm running statistics can improve stability.

2. B

   If the new dataset distribution differs, re-estimating BN stats may help.

3. C

   Setting BatchNorm to training mode with a batch size of 1 provides an unbiased estimate of the population variance.

4. D

   Calling .eval() on a BatchNorm layer disables the application of the learned affine parameters (γ, β).

Select all that apply.

1. A

   Smooth L1 (Huber) is often used for box regression due to robustness to outliers.

2. B

   Focal loss primarily addresses foreground/background (class) imbalance.

3. C

   Standard Cross-Entropy loss scales the loss of easy examples to zero more aggressively than Focal Loss.

4. D

   In two-stage detectors, the RPN has its own objectness + box regression losses.

Select all that apply.

1. A

   Soft attention is differentiable and can be trained with backpropagation.

2. B

   Hard attention often needs sampling + REINFORCE (or similar) due to non- differentiability.

3. C

   Soft attention significantly reduces the computational complexity of the forward pass compared to Hard attention.

4. D

   Cross-attention can align decoder queries to encoder keys/values (e.g., words attending to image regions).