1. Question 1

Architecture that uses small 3×3 filters and increases network depth:

• ❌ GRU

• ❌ RNN

• ❌ LSTM

• ✅ VGG

Explanation:
VGG16/VGG19 use repeat blocks of 3×3 convolutions to form deep CNNs.

---

2. Question 2

Purpose of MaxPooling2D((2,2))?

• ✅ Reduces dimensionality

• ❌ Flattens feature maps

• ❌ Final classification

• ❌ Extracts features

Explanation:
MaxPooling reduces spatial size, computation, and helps retain dominant features.

---

3. Question 3

How ImageDataGenerator helps in augmentation?

• ❌ Reducing resolution

• ❌ Cropping only

• ❌ Adding noise

• ✅ Rotating, shifting, flipping images

Explanation:
ImageDataGenerator creates varied images to reduce overfitting.

---

4. Question 4

Purpose of featurewise_center?

• ❌ Normalize each sample individually

• ✅ Set the dataset’s mean to 0

• ❌ Rotate images

• ❌ Add noise

Explanation:
featurewise_center subtracts the global dataset mean from all images.

---

5. Question 5

Common ImageNet pre-trained model for transfer learning?

• ❌ RNN

• ❌ GRU

• ✅ VGG16

• ❌ LSTM

Explanation:
VGG16 is a classic CNN widely used for transfer learning.

---

6. Question 6

Meaning of include_top=False?

• ❌ Exclude batch norm

• ❌ Exclude pooling layers

• ❌ Exclude convolution layers

• ✅ Exclude fully connected (classification) layers

Explanation:
This keeps the convolutional base but removes the dense classifier.

---

7. Question 7

Why freeze pre-trained model layers initially?

• ❌ Speed up training (secondary effect)

• ✅ Keep pre-trained weights unchanged

• ❌ Reduce memory

• ❌ Prevent overfitting

Explanation:
Freezing preserves learned features from ImageNet.

---

8. Question 8

Fine-tuning a pre-trained model means:

• ❌ Freeze all layers

• ❌ Train only top layers

• ❌ Change architecture

• ✅ Unfreeze and retrain some deeper layers

Explanation:
You refine higher-level feature representations.

---

9. Question 9

Role of flow_from_directory?

• ✅ Load images + apply augmentation from directory

• ❌ Generate synthetic images

• ❌ Convert to grayscale

• ❌ Compile model

Explanation:
It loads batches of images directly from folders.

---

10. Question 10

How does transpose convolution upsample images?

• ❌ Compile with Adam

• ❌ Downsampling

• ❌ Simplify architecture

• ✅ Inserting zeros between elements of the feature map

Explanation:
Transpose convolution performs learnable upsampling via zero-insertion + convolution.

---

🧾 Summary Table

---