Quiz Space

Deep Learning for Computer Vision · Quiz 2 · 12 Apr 2026 · January 2026 term

Question 15: Which statements can be true in practice when fine-tunin…

Question 15

+1 markOne or more correct options

Which statements can be true in practice when fine-tuning with small batch sizes?

Select all that apply.

  1. A

    Freezing BatchNorm running statistics can improve stability.

  2. B

    If the new dataset distribution differs, re-estimating BN stats may help.

  3. C

    Setting BatchNorm to training mode with a batch size of 1 provides an unbiased estimate of the population variance.

  4. D

    Calling .eval() on a BatchNorm layer disables the application of the learned affine parameters (γ, β).

Show answer

Correct answers

  • A

    Freezing BatchNorm running statistics can improve stability.

  • B

    If the new dataset distribution differs, re-estimating BN stats may help.

Video solution

The video solution to this question is being recorded and will play here. Until then, the answer key is above.

See the whole paperTake it as a mock test

Question 15 of 23 in the IIT Madras BS Deep Learning for Computer Vision (Deep Learning for Computer Vision) Quiz 2 paper sat on 12 Apr 2026, in the January 2026 term (Deep Learning For Computer Vision 07 Apr 26). It carries 1 mark.

More questions from this paper

  1. Q1Why do very deep plain CNNs (without skip connections) sometimes show higher training error than shallower CNNs?
  2. Q2In a standard ResNet-50 bottleneck block, the three convolutions are typically:
  3. Q3In Inception/GoogLeNet modules, the main purpose of 1×1 convolutions is to:
  4. Q4MobileNetV1 reduces computation primarily by using:
  5. Q5EfficientNet’s key scaling idea is:
  6. Q6You have a pretrained CNN backbone and only 500 labeled images for a new task. A strong first baseline is:
  7. Q7When fine-tuning a model, what is a common way to prevent the new training from "overwriting" or destroying the useful …
  8. Q8Which method produces a heatmap by weighting convolutional feature maps using gradients of a target class score?
  9. Q9Which detector is a classic two-stage detector with an RPN + ROI feature extraction?
  10. Q10Which statement is most accurate?
  11. Q11RetinaNet is notable primarily because it introduced:
  12. Q12A model that predicts a class label for every pixel without separating object instances is:
  13. Q13Vanishing gradients in vanilla RNNs are largely caused by:
  14. Q14Which statement is true?
  15. Q16Which statements are true?
  16. Q17Which statements are true ?
  17. Q18Input feature map: 14×14×128, output: 14×14×256, kernel: 3×3, stride 1, same padding, ignore bias. Compute number of pa…
  18. Q19Using the same setup (Input feature map: 14×14×128, output: 14×14×256, kernel: 3×3, stride 1, same padding, ignore bias…
  19. Q20A ResNet bottleneck block takes 256 channels in and uses: 1×1 conv to 256 channels → 3×3 conv at 256 → 1×1 conv to 1024…
  20. Q21A detector uses three pyramid levels: P3 = 80×80, P4 = 40×40, P5 = 20×20. At each location, it uses 9 anchors (3 scales…
  21. Q22Box A: top-left (2,2), bottom-right (10,12) Box B: top-left (6,5), bottom-right (14,15) Compute IoU rounded to 3 decima…
  22. Q23Let query q = [1,2]T and keys: k1 = [1,0]T, k2 = [0,1]T, k3 = [1,1]T. Scores si = qT ki. Use softmax over scores. Use e…