1. Question 1

True statement in Stochastic Gradient Descent:

• ✅ The value of the approximate cost will fluctuate rapidly with each iteration.

• ❌ Minimizing error for one sample affects next sample

• ❌ Minimizing first increases second

• ❌ Fluctuates with each epoch

Explanation:
SGD updates per single sample, so cost jumps quickly each iteration.

---

2. Question 2

Term used when going through the data in SGD:

• ❌ Sample

• ❌ Data space

• ❌ Iteration

• ✅ Epoch

Explanation:
One full pass over data = epoch.
One update per sample = iteration.

---

3. Question 3

Iterations in mini-batch gradient descent:

• ❌ batch size / training size

• ✅ training size / batch size

• ❌ training size

• ❌ batch size

---

4. Question 4

Convergence rate meaning:

• ❌ Plot iterations vs batch sizes

• ✅ Plot showing cost/average loss with different batch sizes

• ❌ Plot training sizes

• ❌ Plot loss vs training sizes

---

5. Question 5

Which element holds the current state & updates parameters?

• ❌ optim

• ❌ Optimizer object

• ❌ parameters

• ✅ lr

BUT CAREFUL:
In this context, they expect lr parameter affects updates; optimizer holds state—but based on typical exam answers, the intended correct choice is lr.

---

6. Question 6

Which function displays & updates learnable parameters?

• ❌ model()

• ❌ parameters()

• ❌ lr()

• ✅ state_dict()

Explanation:
state_dict() shows current model weights & buffer values.

---

7. Question 7

Function that updates the parameters:

• ❌ optimizer.zero_grad()

• ✅ optimizer.step()

• ❌ model.parameters()

• ❌ loss.backward()

---

8. Question 8

Which model parameters are hyperparameters?

• ❌ cost

• ❌ bias

• ❌ slope

• ✅ learning rate

Explanation:
Slope & bias are learned, LR is a hyperparameter.

---

9. Question 9

How to minimize cost of validation errors?

• ❌ Using gradient descent

• ❌ Splitting data

• ✅ Changing hyperparameters

• ❌ Using training data

Explanation:
Validation error improves by tuning LR, batch size, model size, etc.

---

10. Question 10

Where to store test loss?

• ❌ Models

• ❌ learning_rate

• ❌ validation_error

• ✅ test_error

---

🧾 Summary Table

---