Indian Institute of Technology Madras
BS Degree Programme · Examination & Study Paper
Mathematical Foundations of Generative AI
DDPM: forward & reverse process
Official Paper
Examination Instructions & Guidelines
- Solve on notebook/paper first before checking solutions.
- MCQ: Exactly 1 option is correct. Negative marking applies.
- NAT: Numerical response. No negative marking.
- Scientific calculator allowed for all calculations.
- Official answer key is sealed at the end of this sheet.
End Term · 31 Aug 2025 (FN)End Term
- Q1MCQ5 Marks·Choose one option

- A1.797
- B1.910
- C2.000
- D1.620
- Q2MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
- Q3MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
- Q4MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
End Term · 31 Aug 2025 (AN)End Term
- Q5MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
- Q6MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
- Q7MCQ5 Marks·Choose one option

- A

- B

- C

- D

- A
- Q8MCQ5 Marks·Choose one option

- A0.28
- B0.32
- C0.20
- D0.18
Quiz 2 · 3 Aug 2025 (AN)Quiz 2
- Q9NAT3 Marks·Write the number
Your answer: - Q10NAT4 Marks·Write the number
Your answer: - Q11MCQ3 Marks·Choose one option

- AAt t = 0, we assume perfect reconstruction
- BAt t = 0, adding noise will corrupt the final output
- CAt t = 0, the posterior is deterministic
- DAll of these
- Q12MCQ3 Marks·Choose one option

- A

- B

- C

- D

- A
- Q13MCQ3 Marks·Choose one option
In a U-Net for DDPMs, the time embedding is usually added to the feature maps after a convolutional layer. If a feature map has shape ‘[B, C, H, W]’ and the time embedding has shape ‘[B, C]’, how is the embedding typically broadcasted and combined?
- AIt is reshaped to ‘[B, C, 1, 1]’ and then added.
- BIt is tiled to ‘[B, C, H, W]’ and then multiplied.
- CIt is passed through a separate linear layer for each spatial location.
- DIt is concatenated along the channel dimension.
- Q14MCQ3 Marks·Choose one option

- A[1]
- B[1, 1]
- C[1, 64]
- D[64, 64]