WorksheetsGLIM - PGPM - CV Quiz3_Transfer Learning models (DL)
Total questions: 11
Worksheet time: 6mins
What is the primary idea behind transfer learning in computer vision?
Training a CNN from scratch on a small dataset
Reusing knowledge from a model trained on a large dataset for a new task
Increasing the number of hidden layers in a CNN
Compressing a deep learning model for deployment
Which dataset is most commonly used to pretrain models such as VGG, ResNet, and Inception?
CIFAR-10
MNIST
ImageNet
COCO
Which part of a pretrained CNN is usually replaced for a new classification task?
Convolution layers
Pooling layers
Final fully connected (classification) layer
Input layer
A major limitation of VGG compared to modern CNNs is:
Poor accuracy
Very small number of parameters
High memory and computational requirements
Inability to use transfer learning
ResNet introduced which key concept to solve the vanishing gradient problem?
Dropout
Batch normalization
Skip (residual) connections
Max pooling
YOLO is primarily designed for:
Image classification only
Semantic segmentation
Real-time object detection
Image compression
A key difference between YOLO and CNN classifiers (VGG/ResNet) is that YOLO:
Uses grayscale images only
Performs detection and classification in a single forward pass
Requires image patches as input
Cannot be trained using transfer learning
Which metric is most important when missing a critical class (e.g., large potholes or safety hazards)?
Accuracy
Precision
Recall
Loss value
F1 score is preferred over accuracy when:
Dataset is perfectly balanced
Training time is high
Model size is large
Classes are imbalanced
From a managerial perspective, the best model is the one that:
Has the highest accuracy only
Is the deepest network available
Balances performance, speed, and deployment cost
Uses the latest architecture
Which model would be most suitable for deployment on a low-memory mobile device?
MobileNet
EfficientNet
ResNet50
VGG16
