1. Which metric should Jane use to measure how well the model distinguishes between classes?

❌ Mean Squared Error
✅ AU-ROC (Area Under ROC Curve)
❌ Model accuracy
❌ Cross-entropy loss

Explanation:
AU-ROC evaluates how well the model separates classes across all thresholds, making it ideal for classification performance assessment.

---

2. What key component must Liam include in a CNN model for image classification?

❌ Dropout layers to increase model capacity
❌ Recurrent layers to capture temporal sequences
✅ Convolutional layers to extract spatial features
❌ Batch normalization layers to reduce model size

Explanation:
Convolutional layers are the core of CNNs, extracting spatial patterns such as edges, textures, and shapes in images.

---

3. What should the AgriTech team do first when building a CNN classifier in Keras?

❌ Compile the CNN model
✅ Preprocess the dataset to ensure all images are the same size
❌ Initialize the CNN model architecture
❌ Train the CNN model

Explanation:
All images must be resized and normalized before feeding them into the model; preprocessing is always the first step.

---

4. Why is HeUniform() used as a kernel initializer?

❌ It regularizes the model with L2 norms
✅ It initializes weights suited to ReLU-based activations
❌ It schedules the learning rate automatically
❌ It freezes pretrained ImageNet weights

Explanation:
He initialization prevents vanishing/exploding gradients and works best with ReLU or variants.

---

5. Why is torch.no_grad() used during validation?

✅ To disable gradient calculation, saving memory and computation
❌ To randomize batch order
❌ To reset model weights
❌ To enable GPU acceleration

Explanation:
torch.no_grad() ensures no gradients are tracked, making evaluation faster and more memory-efficient.

---

6. Why is F1 Score preferred for highly imbalanced datasets?

❌ It considers only true positives and true negatives
❌ It measures area under the precision-recall curve
✅ It balances precision and recall, focusing on minority-class performance
❌ It is threshold-independent like ROC-AUC

Explanation:
F1 is the harmonic mean of precision and recall, giving a fairer performance measure when classes are imbalanced.

---

7. How can you generate a detailed summary with per-class precision, recall, and F1?

❌ accuracy_score
✅ classification_report
❌ roc_curve
❌ confusion_matrix

Explanation:
classification_report outputs precision, recall, F1-score, and support for each class.

---

🧾 Summary Table

---