Back to Subject

Unit 3: Transfer Learning and Fine-Tuning for Vision Tasks - Practice Quiz

CSE471 — Deep Learning For Computer Vision 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is transfer learning in computer vision?

Transfer learning Easy
A. Removing all layers from a neural network
B. Training a model without any data
C. Using knowledge from one task for another task
D. Converting images into text files

2 Why is transfer learning commonly used for image classification?

Transfer learning Easy
A. It guarantees perfect classification accuracy
B. It prevents images from having different sizes
C. It removes the need for an output layer
D. It requires fewer labeled training images

3 What is the main feature of ResNet architectures?

ResNet Easy
A. Shortcut connections between layers
B. Training without labeled examples
C. A complete absence of pooling layers
D. Only one convolutional layer

4 What problem do residual connections in ResNet help address?

ResNet Easy
A. Lack of image file formats
B. Missing class names in datasets
C. Difficulty training very deep networks
D. Excessive image brightness

5 What is a key idea behind EfficientNet?

EfficientNet Easy
A. Training every model with one fixed image size
B. Using only fully connected layers
C. Scaling depth, width, and resolution together
D. Replacing images with numerical labels

6 EfficientNet models are designed mainly to improve which balance?

EfficientNet Easy
A. Accuracy and computational efficiency
B. Labels and file compression
C. Color and image brightness
D. Batch size and class names

7 What is MobileNet primarily designed for?

MobileNet Easy
A. Training models with unlimited memory
B. Replacing all image preprocessing steps
C. Efficient use on mobile and embedded devices
D. Processing only text-based documents

8 Which operation is commonly used by MobileNet to reduce computation?

MobileNet Easy
A. Repeated image enlargement
B. Unrestricted fully connected layers
C. Three-dimensional pooling only
D. Depthwise separable convolution

9 In feature extraction, what is usually done with the pretrained convolutional layers?

Feature extraction Easy
A. Their weights are kept fixed
B. Their layers are randomly removed
C. Their weights are deleted
D. Their inputs are replaced with labels

10 What do early convolutional layers often learn from images?

Feature extraction Easy
A. Database connection rules
B. Final class names only
C. Complete sentences
D. Edges and simple textures

11 When using a pretrained model as a feature extractor, which part is commonly replaced?

Feature extraction Easy
A. The training dataset format
B. The image file extension
C. The final classification layer
D. Every convolutional filter

12 What does fine-tuning mean in transfer learning?

Fine-tuning Easy
A. Using a model without any training
B. Freezing the entire neural network permanently
C. Changing image labels without changing model weights
D. Updating some pretrained weights on new data

13 Fine-tuning is often useful when the target dataset is:

Fine-tuning Easy
A. Related but different from the pretraining dataset
B. Made only of numerical formulas
C. Completely empty of images
D. Identical to the model architecture

14 What does it mean to freeze a neural network layer?

Managing freezing and unfreezing of layers Easy
A. Its weights are not updated during training
B. Its outputs are always set to zero
C. Its input images are permanently resized
D. Its filters are converted into labels

15 Why are pretrained layers often frozen at the beginning of training?

Managing freezing and unfreezing of layers Easy
A. To increase the number of target classes
B. To remove the need for a loss function
C. To make every layer randomly initialized
D. To preserve useful learned features

16 During gradual fine-tuning, what is commonly done after training the new classifier?

Managing freezing and unfreezing of layers Easy
A. The loss function is removed
B. Some earlier layers are unfrozen
C. All input images are discarded
D. The model output is changed to audio

17 What is a layer-wise learning rate strategy?

Applying layer-wise learning rates Easy
A. Assigning one label to every layer
B. Using one batch for every layer
C. Changing the image size after every layer
D. Using different learning rates for different layers

18 Which learning rate is commonly used for newly added classification layers?

Applying layer-wise learning rates Easy
A. A relatively larger learning rate
B. A rate that never changes model weights
C. A smaller rate than every frozen layer
D. A permanently zero learning rate

19 What is domain adaptation used for?

Domain adaptation and dataset Easy
A. Adapting a model to a different data distribution
B. Converting a classification task into file storage
C. Increasing the number of network parameters automatically
D. Removing all differences between image classes

20 What is the target dataset in transfer learning?

Domain adaptation and dataset Easy
A. The dataset used only to initialize random numbers
B. A dataset that has no relationship to the task
C. The dataset used for the new task
D. A dataset containing model source code

21 A team has 2,000 labeled images of industrial defects and a model pretrained on ImageNet. What is the most appropriate initial transfer-learning strategy?

Transfer learning Medium
A. Freeze the classifier and retrain only the earliest convolution
B. Replace all convolutional layers with fully connected layers
C. Train the entire network from randomly initialized weights
D. Freeze the backbone and train a new task-specific classifier

22 Transfer learning from ImageNet is likely to provide the smallest benefit for which target task?

Transfer learning Medium
A. Classifying patterns in radar measurements that differ substantially from natural images
B. Classifying common household objects
C. Recognizing vehicles in street images
D. Identifying species in wildlife photographs

23 A 50-layer plain CNN trains poorly, while a ResNet-50 trains successfully on the same dataset. Which ResNet mechanism most directly explains the improvement?

ResNet Medium
A. Dropout is automatically added after each convolution
B. Residual blocks remove the need for nonlinear activations
C. Max pooling prevents every vanishing-gradient problem
D. Residual connections provide shorter paths for gradient flow

24 In a ResNet block, the residual branch produces and the shortcut carries . When their shapes match, what is the block output before the final activation?

ResNet Medium
A.
B.
C.
D.

25 A practitioner wants to scale an EfficientNet model while preserving its core scaling principle. Which change best follows compound scaling?

EfficientNet Medium
A. Increase width and remove all resolution scaling to keep the model's spatial dimensions unchanged
B. Increase depth, width, and input resolution in a balanced way
C. Increase only input resolution until GPU memory is exhausted
D. Increase only depth while keeping width and resolution fixed

26 When replacing EfficientNet-B0 with a larger EfficientNet variant for transfer learning, what trade-off should be expected?

EfficientNet Medium
A. Fewer parameters because every larger variant uses stronger pooling
B. No change in inference latency because compound scaling affects training only
C. Identical accuracy with lower computation and memory use
D. Potentially better accuracy with greater computation and memory use

27 Why is MobileNet commonly selected for an image-classification application running on a mobile device?

MobileNet Medium
A. Depthwise separable convolutions reduce computation and parameters
B. Residual additions eliminate the need to store feature maps
C. Every convolution operates only on one-pixel input images
D. Large fully connected layers improve low-power inference

28 A MobileNet model uses a width multiplier . What is the primary effect of decreasing ?

MobileNet Medium
A. It converts depthwise convolutions into standard convolutions with additional trainable kernels
B. It increases image resolution and receptive-field size
C. It freezes a larger fraction of pretrained layers
D. It reduces channel counts and computational cost

29 A pretrained CNN is used only to compute embeddings, which are then supplied to a newly trained linear classifier. Which layer output is usually the most suitable embedding?

Feature extraction Medium
A. The original classifier's class-probability vector
B. The final pooled feature vector before the original classifier
C. The loss value computed after each training batch
D. The raw RGB image before normalization

30 A frozen pretrained backbone is used for feature extraction, but the new classifier overfits. Which intervention most directly addresses the problem without modifying backbone weights?

Feature extraction Medium
A. Remove the validation set to provide more classifier training samples
B. Unfreeze every backbone layer and greatly increase its learning rate
C. Cache each training feature multiple times and treat the copies as independently labeled examples
D. Add regularization and augmentation while training the classifier

31 After training a new classification head, a practitioner unfreezes part of the pretrained backbone. Which learning-rate change is generally appropriate?

Fine-tuning Medium
A. Use the largest stable learning rate for all layers
B. Set the backbone learning rate permanently to zero
C. Increase the learning rate whenever validation loss rises
D. Use a smaller learning rate for the unfrozen backbone

32 During fine-tuning, training accuracy rises while validation accuracy falls after several epochs. What is the most appropriate response?

Fine-tuning Medium
A. Increase model size without changing the training setup
B. Evaluate only on the training set because pretrained models do not require validation
C. Continue training until both accuracies reach 100%
D. Apply early stopping and strengthen regularization

33 For a small target dataset similar to ImageNet, which staged training procedure is generally most appropriate?

Managing freezing and unfreezing of layers Medium
A. Alternate frozen and unfrozen states after every individual batch throughout training
B. Unfreeze the entire network before training the new head
C. Train the new head, then unfreeze later blocks gradually
D. Train only the earliest convolution and discard later blocks

34 Why are the final backbone blocks usually unfrozen before the earliest blocks?

Managing freezing and unfreezing of layers Medium
A. Later blocks always require less GPU memory
B. Earlier blocks produce class probabilities rather than visual features
C. Later blocks contain more task-specific semantic features
D. Earlier blocks contain no trainable parameters

35 A frozen backbone contains Batch Normalization layers. The validation performance becomes unstable when these layers continue updating their running statistics on very small batches. What should be tried?

Managing freezing and unfreezing of layers Medium
A. Increase the classifier learning rate after every batch
B. Keep the Batch Normalization layers in evaluation mode
C. Delete all normalization layers from the pretrained network
D. Replace running means with randomly sampled values during each validation pass

36 A model is divided into an early backbone, a late backbone, and a new classifier. Which learning-rate assignment best matches discriminative fine-tuning?

Applying layer-wise learning rates Medium
A. , , and , respectively
B. , , and , respectively
C. , , and , respectively
D. , , and , respectively

37 Why can layer-wise learning rates reduce catastrophic forgetting during transfer learning?

Applying layer-wise learning rates Medium
A. They force all layers to converge to identical parameter values after each optimization step
B. They prevent the classifier from receiving any gradients
C. They guarantee that the target dataset has balanced classes
D. They limit changes to general features in early layers

38 A classifier is trained on studio product photos but deployed on dim warehouse-camera images. Which change most directly reduces this domain shift?

Domain adaptation and dataset Medium
A. Shuffle the studio training set using a different random seed
B. Fine-tune using labeled warehouse-camera images
C. Increase the number of source-domain output classes
D. Add more studio photos with the same lighting

39 Only unlabeled target-domain images are available. Which technique can still help align source and target representations?

Domain adaptation and dataset Medium
A. Assigning every target image to the most frequent source class and treating those assignments as verified ground truth
B. Training only on source labels without using target images
C. Supervised fine-tuning using invented target labels
D. Adversarial domain adaptation with a domain discriminator

40 A medical-image dataset contains multiple scans from each patient. Which split best avoids leakage when evaluating a transfer-learned model?

Domain adaptation and dataset Medium
A. Place augmented scans in validation and originals in training
B. Randomly split individual scans across all dataset splits
C. Assign all scans from a patient to only one split
D. Use the same patient scans for training and testing but apply different image normalization methods

41 A CNN pretrained on ImageNet is transferred to a medical-image dataset with only 800 labeled examples. The target images are grayscale, but their low-level edge statistics resemble natural images. Which initialization strategy is most defensible?

Transfer learning Hard
A. Initialize the first layer randomly and retain only the pretrained classifier weights
B. Replicate the grayscale channel three times and retain pretrained convolutional weights
C. Convert filters to grayscale and train all layers with the same large learning rate
D. Replace every convolutional filter with random weights before classifier training

42 In a pretrained ResNet block, the identity shortcut has dimension while the residual branch outputs channels and uses stride 2. Which modification preserves the residual-addition operation?

ResNet Hard
A. Average the residual channels into channels before addition
B. Remove the shortcut and double the residual branch activation width
C. Use a projection with stride 2 on the identity branch
D. Add zero padding to the identity tensor without changing its spatial size

43 EfficientNet scales network depth, width, and input resolution using a compound coefficient. If compute is approximately proportional to , which scaling choice best preserves balanced growth when increasing model capacity?

EfficientNet Hard
A. Increase depth, width, and resolution by unrelated manually selected factors
B. Increase resolution first and keep depth and width fixed for stability
C. Increase all three dimensions under a shared compound scaling constraint
D. Increase only width because channel expansion dominates representational capacity

44 For an input feature map of size and a depthwise-separable convolution producing output channels with kernel size , what is the approximate parameter reduction relative to a standard convolution?

MobileNet Hard
A. From to
B. From to
C. From to
D. From to

45 A frozen backbone produces features that separate training classes well, but target validation accuracy remains poor after replacing the classifier. Which observation most strongly suggests that the fixed representation is unsuitable for the target task?

Feature extraction Hard
A. The backbone contains more parameters than the newly initialized classifier
B. The optimizer reports a lower learning rate for the classifier than for the backbone
C. The classifier loss decreases while validation errors remain structured by target subdomain
D. The target classifier uses fewer output units than the source classifier

46 A pretrained model reaches high training accuracy immediately after unfreezing all layers, but target validation accuracy falls and feature norms grow rapidly. Which intervention is most appropriate first?

Fine-tuning Hard
A. Reinitialize the backbone and train it from scratch on the target set
B. Use a smaller backbone learning rate with stronger regularization
C. Increase the learning rate to escape the pretrained parameter basin
D. Freeze only the classifier and increase augmentation strength afterward

47 During staged fine-tuning, the backbone is frozen but contains BatchNorm layers. The target dataset has a different intensity distribution and is small. Which configuration is generally safest?

Managing freezing and unfreezing of layers Hard
A. Update BatchNorm statistics on every target batch regardless of batch size
B. Keep BatchNorm statistics frozen while training only the new classifier
C. Replace BatchNorm with dropout and unfreeze every convolutional layer
D. Remove all BatchNorm layers before transferring the pretrained backbone

48 A three-stage backbone uses learning rates , , and from the earliest to latest stage, while the new classifier uses . What principle does this schedule encode?

Applying layer-wise learning rates Hard
A. Earlier features receive larger updates because they have higher generality
B. Every stage receives the same effective update after normalization
C. Later features and the classifier adapt more rapidly to target semantics
D. The classifier is protected from overfitting by using the largest learning rate

49 A source model classifies daytime street images, while the target dataset contains nighttime images with identical labels. Unlabeled target data are abundant. Which method directly addresses the domain shift without requiring target labels?

Domain adaptation and dataset Hard
A. Discard nighttime images because their labels cannot be verified
B. Use random labels for target images to balance the training batches
C. Align source and target feature distributions with an unsupervised adaptation loss
D. Train a larger classifier using only the labeled source images

50 A target dataset has the same visual domain as the pretraining data but a highly different label taxonomy. Which transfer strategy best reduces harmful source-task bias?

Transfer learning Hard
A. Reuse the entire source classifier and freeze the complete network
B. Copy source logits directly into the target labels without recalibration
C. Reuse the backbone, replace the classifier, and tune upper layers cautiously
D. Train only the source classifier until its accuracy stops improving

51 Why can a ResNet often be fine-tuned more deeply than a plain network of comparable depth, even when the target dataset is small?

ResNet Hard
A. Identity paths improve gradient propagation and preserve pretrained representations
B. Residual branches eliminate the need for normalization during adaptation
C. Shortcut connections guarantee that target-domain features are linearly separable
D. Residual shortcuts make every layer independent of the loss function

52 When transferring an EfficientNet model to a target dataset whose objects occupy much smaller image regions than in pretraining, which change is most likely to help before substantially increasing model width?

EfficientNet Hard
A. Reduce input resolution so the object occupies fewer feature-map cells
B. Increase input resolution while preserving the architecture's scaling assumptions
C. Remove squeeze-and-excitation modules to prevent channel recalibration
D. Replace depthwise convolutions with fully connected layers at every stage

53 A MobileNet-based classifier must run under a strict latency budget. Validation accuracy improves after increasing width, but measured latency exceeds the budget. Which adjustment is most principled?

MobileNet Hard
A. Replace depthwise convolutions with standard convolutions for better hardware use
B. Increase width further and compensate with heavier quantization later
C. Add dense layers because parameter count alone determines inference latency
D. Reduce the width multiplier and evaluate accuracy at the deployment resolution

54 A frozen backbone is used with global average pooling. The target classes differ mainly by precise spatial arrangement rather than overall texture. What limitation should be expected?

Feature extraction Hard
A. Global pooling always increases sensitivity to object translation errors
B. Global pooling makes the feature dimension dependent on image width
C. Global pooling prevents gradients from reaching the classifier parameters
D. Global pooling may discard spatial information needed to distinguish arrangements

55 A target dataset contains severe class imbalance. During fine-tuning, overall accuracy rises while minority-class recall collapses. Which evaluation and optimization change is most appropriate?

Fine-tuning Hard
A. Freeze the classifier and optimize only the backbone with unweighted loss
B. Use balanced metrics and class-aware loss or sampling during adaptation
C. Use accuracy alone because it reflects the deployment class frequencies
D. Reduce augmentation because minority examples should remain visually identical

56 After unfreezing a new block, its gradients are nonzero but its weights appear unchanged during training. Which implementation issue is most likely?

Managing freezing and unfreezing of layers Hard
A. The residual block contains an identity shortcut
B. The classifier has a larger output dimension than the backbone
C. The block is absent from the optimizer parameter groups
D. The input images were normalized using source statistics

57 Two layers have gradients with similar magnitudes, but the earlier layer has much larger pretrained weights and is more transferable. Why can assigning it a smaller learning rate still be justified?

Applying layer-wise learning rates Hard
A. Smaller rates guarantee identical updates across layers with different norms
B. Learning rates determine only gradient direction, not parameter displacement
C. Layer-wise rates remove the need to monitor validation performance
D. Smaller rates limit relative parameter drift and protect established features

58 A target validation set was collected from the same camera and locations as the training set, while deployment uses unseen cameras. Fine-tuning improves validation accuracy but deployment performance decreases. What is the primary diagnosis?

Domain adaptation and dataset Hard
A. The source dataset must be discarded because deployment has unseen cameras
B. The target classifier requires more output classes than the deployment system
C. The model has overfit the validation distribution and the evaluation is not representative
D. The backbone has become too shallow to represent camera-specific features

59 A target dataset is larger than the source pretraining dataset but differs substantially in sensor modality and image formation. Which conclusion is most defensible?

Transfer learning Hard
A. Source pretraining is useless whenever the target dataset has more examples
B. Transfer may still help, but modality mismatch should be tested against strong baselines
C. The target dataset size alone guarantees that transfer will improve accuracy
D. The pretrained model should always be frozen because modality changes are irreversible

60 A fine-tuned model performs well on clean target images but fails under a mild illumination change that is common in deployment. Which response best targets the failure mode?

Fine-tuning Hard
A. Use augmentation or domain-relevant normalization that represents illumination variation
B. Freeze all layers permanently so illumination cannot affect learned features
C. Increase classifier depth without changing the training distribution
D. Remove validation images with illumination changes to stabilize the metric
Previous MCQ
Unit 2
Next MCQ
Unit 4