Virtual Labs

Transfer Learning with Deep CNNs

1. What is the main purpose of unfreezing the last layers in Transfer Learning?

a: To allow the model to adapt high-level features to the target dataset Explanation

Explanation

b: To permanently delete the pretrained backbone Explanation

Explanation

c: To stop training in the network Explanation

Explanation

d: To increase the dataset size automatically Explanation

Explanation

2. If training accuracy is very high but validation accuracy is low, it indicates:

a: Perfect model performance Explanation

Explanation

b: Underfitting Explanation

Explanation

c: Overfitting Explanation

Explanation

d: Data preprocessing is not required Explanation

Explanation

3. Why is a smaller learning rate preferred during fine-tuning?

a: To increase memory usage Explanation

Explanation

b: To avoid large weight updates that can damage pretrained features Explanation

Explanation

c: To make the dataset larger Explanation

Explanation

d: To remove the need for backpropagation Explanation

Explanation

4. In MobileNetV2, base freeze means:

a: Training the entire model from scratch Explanation

Explanation

b: Freezing all backbone layers and training only the added classifier layers Explanation

Explanation

c: Removing the classifier layers Explanation

Explanation

d: Increasing the image size automatically Explanation

Explanation

5. Unfreezing 20% of the last layers mainly helps in:

a: Learning only basic edges and textures Explanation

Explanation

b: Increasing batch size automatically Explanation

Explanation

c: Learning task-specific patterns for the target dataset Explanation

Explanation

d: Converting the CNN into an SVM model Explanation

Explanation

6. Which pretrained model is generally more computationally heavy?

a: VGG19 Explanation

Explanation

b: MobileNetV2 Explanation

Explanation

c: KNN Explanation

Explanation

d: Naive Bayes Explanation

Explanation

7. Which plot best helps in checking training performance and overfitting?

a: Histogram of file names Explanation

Explanation

b: Scatter plot of random points Explanation

Explanation

c: Confusion matrix before training Explanation

Explanation

d: Accuracy and loss curves Explanation

Explanation

8. If validation loss increases while training loss decreases, it usually means:

a: The model is underfitting Explanation

Explanation

b: The model is generalizing perfectly Explanation

Explanation

c: The model is overfitting Explanation

Explanation

d: The dataset has no images Explanation

Explanation

9. What is the purpose of the Softmax layer in CNN classification?

a: To reduce image size Explanation

Explanation

b: To normalize dataset images Explanation

Explanation

c: To detect edges in images Explanation

Explanation

d: To convert outputs into probabilities over classes Explanation

Explanation

10. Negative Transfer occurs when:

a: Transfer improves accuracy strongly Explanation

Explanation

b: Only classifier layer is trained Explanation

Explanation

c: Learning rate is small Explanation

Explanation

d: Transfer from an unrelated source reduces target task performance Explanation

Explanation