1. Question 1

Primary benefit of using a custom training loop:

• ❌ Faster training

• ❌ Less validation data

• ❌ Auto-handles training

• ✅ Greater control over the training process

Explanation:
Custom loops give you full control over forward pass, backward pass, metrics, and loss computation.

---

2. Question 2

Key component of a custom training loop:

• ❌ Model regularization

• ❌ Data augmentation

• ❌ Model callbacks

• ✅ Dataset

Explanation:
Custom loops manually iterate over tf.data.Dataset batches.

---

3. Question 3

Purpose of HyperParameters object in Keras Tuner:

• ❌ Store training data

• ✅ Define ranges/values for hyperparameters

• ❌ Compile model

• ❌ Save architecture

Explanation:
HyperParameters lets you define values like learning rate, layer size, dropout ranges, etc.

---

4. Question 4

Main benefit of hyperparameter tuning:

• ✅ Find best hyperparameters for highest performance

• ❌ Simplify architecture

• ❌ Reduce parameters

• ❌ Increase dataset size

Explanation:
Tuning maximizes accuracy/score by testing different configurations.

---

5. Question 5

Search algorithm provided in Keras Tuner:

• ❌ Simulated Annealing

• ❌ Particle Swarm Optimization

• ✅ Hyperband

• ❌ Genetic Algorithm

Explanation:
Built-in options include Hyperband, RandomSearch, BayesianOptimization.

---

6. Question 6

Best initialization for ReLU layers:

• ✅ He Initialization

• ❌ Zero

• ❌ Xavier

• ❌ Random

Explanation:
He initialization prevents vanishing/exploding activations in ReLU networks.

---

7. Question 7

Purpose of learning rate scheduling:

• ❌ Reduce dataset

• ✅ Adjust learning rate during training for better convergence

• ❌ Increase epochs

• ❌ Decrease model complexity

Explanation:
Schedulers gradually reduce or modify learning rate to stabilize training.

---

8. Question 8

Main benefit of batch normalization:

• ❌ Increases parameters

• ✅ Normalizes layer inputs to improve training stability

• ❌ Removes activation

• ❌ Eliminates dropout

Explanation:
Batch norm reduces internal covariate shift → smoother, faster training.

---

9. Question 9

Primary purpose of mixed precision training:

• ✅ Speed up training & reduce memory usage

• ❌ Remove validation

• ❌ Improve accuracy

• ❌ Simplify architecture

Explanation:
Mixed precision uses float16 + float32 for faster GPU compute.

---

10. Question 10

How quantization improves TensorFlow model performance:

• ❌ Removes optimizer

• ❌ Improves input quality

• ❌ Adds more layers

• ✅ Reduces model size + increases inference speed

Explanation:
Quantizing weights/ops → faster & smaller models, especially on edge devices.

---

🧾 Summary Table

---