1. Question 1

Which layer performs flattening and connects feature maps to outputs?

• ❌ Input layer

• ❌ Pooling layer

• ✅ Fully connected layer

• ❌ ReLU layer

Explanation:
Fully connected (Dense) layers require flattened input and map extracted features to class probabilities.

---

2. Question 2

What enables RNNs to retain sequential dependencies?

• ❌ Attention weights

• ❌ Parallel processing

• ✅ Hidden state propagation

• ❌ Convolutional filters

Explanation:
RNNs pass a hidden state from one timestep to the next to capture context.

---

3. Question 3

Key advantage of deep networks for image classification?

• ❌ Faster convergence

• ❌ Lower memory

• ❌ Reduced gradient issues

• ✅ Learn hierarchical features (edges → shapes → objects)

Explanation:
Deep layers allow building complex feature hierarchies, essential for vision.

---

4. Question 4

Primary transformer architectural components?

• ✅ Self-attention + positional encoding

• ❌ Convolution + pooling

• ❌ Recurrent layers

• ❌ Hidden + cell states

Explanation:
Transformers process sequences in parallel using attention + positional information.

---

5. Question 5

Principle enabling autoencoder-based anomaly detection?

• ✅ Learn normal patterns → high reconstruction error = anomaly

• ❌ Reinforcement learning

• ❌ Labeled data

• ❌ Generating synthetic anomalies

Explanation:
Autoencoders reconstruct only normal patterns well.

---

6. Question 6

Architecture optimized for visual recognition?

• ❌ Autoencoders

• ❌ RBMs

• ✅ Convolutional Neural Networks (CNNs)

• ❌ RNNs

Explanation:
CNNs excel at spatial pattern extraction.

---

7. Question 7 — Select all that apply

What benefits do pooling layers provide?

• ❌ Feature flattening

• ✅ Spatial resolution reduction → lower computation

• ❌ Negative value elimination

• ✅ Translation & scaling invariance

Explanation:
Pooling reduces dimensionality & makes features more robust to shifts.

---

8. Question 8

Optimal CNN layer configuration?

• ❌ GlobalAveragePooling2D → Dense → Activation

• ❌ Flatten → Dense → Dropout → Softmax

• ✅ Conv2D → BatchNormalization → ReLU → MaxPooling2D

• ❌ MaxPooling2D → Conv2D → Dense → ReLU

Explanation:
Correct ordering: Conv → Norm → Activation → Pool.

---

9. Question 9

Essential for transformers to understand word order?

• ❌ Convolution

• ❌ Pooling

• ✅ Positional encoding

• ❌ RNN layers

Explanation:
Transformers need explicit positional information.

---

10. Question 10

Key limitation of transfer learning?

• ❌ Pretrained models can’t be modified

• ❌ Requires larger datasets

• ❌ Always needs more compute

• ✅ Domain mismatch may limit performance

Explanation:
If pretrained domain ≠ target domain, performance may drop.

---

🧾 Summary Table

---