StudyFox
कीमतेंब्लॉगएक्सटेंशनमुफ़्त टूल
  • कीमतें
  • ब्लॉग
  • एक्सटेंशन
  • मुफ़्त टूल
  • शुरू करेंसाइन इन
StudyFox

अपने नोट्स, PDF और स्लाइड्स को फ़्लैशकार्ड, क्विज़, नोट्स और माइंड मैप में बदलें।

StudyFox खोलें

स्टडी टूल्स

  • होमवर्क हेल्पर
  • नोट मेकर
  • फ़्लैशकार्ड मेकर
  • क्विज़ मेकर
  • माइंड मैप मेकर

प्रोडक्ट

  • कीमतें
  • Chrome एक्सटेंशन
  • मुफ़्त एक्सटेंशन
  • स्टडी सेट देखें
  • ब्लॉग

सहायता

  • सहायता केंद्र
  • संपर्क करें
  • गोपनीयता नीति
  • सेवा की शर्तें

तुलना करें

  • Chegg का विकल्प
  • Course Hero का विकल्प
  • Coursology का विकल्प
  • Photomath का विकल्प
  • Gauthmath का विकल्प
  • PhotoSolve का विकल्प
  • CollegeTools का विकल्प
  • Transcript का विकल्प
  • SmartSolve का विकल्प
  • Homeworkify का विकल्प
  • AnswersAi का विकल्प
  • StudyX का विकल्प
  • Quizlet का विकल्प
  • Anki का विकल्प
  • Knowt का विकल्प
  • Brainscape का विकल्प
  • सभी तुलनाएँ

© 2026 StudyFox AI. सर्वाधिकार सुरक्षित।

  • X
  • Discord
  • Instagram
ExploreEngineering & TechData Science & AIDeep Learning Calculations and Training Problems: Hard Applied Quiz

Quiz · Hard

Deep Learning Calculations and Training Problems: Hard Applied Quiz

Parameter counts, convolution sizes, softmax and cross-entropy numbers, optimizer behavior, attention scaling and diagnosing training problems.

by Kill h.15 questionsCreated Nov 20, 2026, 08:26 PM
Data Science & AI· Deep LearningTags#deep-learning#calculations#exam-prep#training

15 questions

  1. 1

    A network has layers 20 -> 64 -> 32 -> 1, fully connected with biases. How many trainable parameters?

    • A

      3,393

    • 3,457

    • C

      3,521

    • D

      3,200

    Layer 1: 20x64+64 = 1,344. Layer 2: 64x32+32 = 2,080. Layer 3: 32x1+1 = 33. Total = 1,344 + 2,080 + 33 = 3,457.

  2. 2

    A CNN layer has 32 filters of size 5x5 on an input with 16 channels (with biases). How many parameters?

    • A

      400

    • 12,832

    • C

      12,800

    • D

      12,544

    Each filter: 5x5x16 = 400 weights + 1 bias = 401. 401 x 32 = 12,832. Forgetting the biases gives 12,800.

  3. 3

    An input of 64x64 passes through conv(3x3, stride 1, padding 0), then max-pool (2x2, stride 2), then conv(3x3, stride 1, padding 0). What is the final spatial size?

    • 29x29

    • B

      31x31

    • C

      60x60

    • D

      30x30

    Conv: 64 - 3 + 1 = 62. Pool: 62 / 2 = 31. Conv: 31 - 3 + 1 = 29.

  4. 4

    Softmax of logits (1, 1, 1, 1) gives what for the first class?

    • A

      0.5

    • 0.25

    • C

      1.0

    • D

      0.0

    Equal logits give equal probabilities: e^1 / (4e^1) = 0.25. Softmax depends only on differences between logits.

  5. 5

    A one-hot target is (0, 1, 0) and softmax outputs (0.2, 0.7, 0.1). What is the cross-entropy loss (natural log)?

    • 0.36

    • B

      1.61

    • C

      0.70

    • D

      0.22

    Only the true class term counts: -ln(0.7) = 0.357, about 0.36. Using the wrong class, e.g. -ln(0.2), would give 1.61.

  6. 6

    With softmax + cross-entropy, the gradient of the loss with respect to the logits is (0.2, -0.3, 0.1) for a 3-class example. What are the predicted probabilities if the true class is the second?

    • A

      (0.2, 0.3, 0.1)

    • B

      (0.8, 0.7, 0.9)

    • C

      (0.2, -0.3, 0.1)

    • (0.2, 0.7, 0.1)

    The gradient is p - y. For the true class (index 2): p2 - 1 = -0.3, so p2 = 0.7. The others have p = gradient: 0.2 and 0.1. They sum to 1.

  7. 7

    Adam (lr 0.001) vs SGD: which statement is true?

    • A

      Adam never overfits

    • Adam keeps a separate adaptive step size per parameter

    • C

      Adam requires no learning rate

    • D

      SGD with momentum cannot be used for deep networks

    Adam keeps running averages of gradients and squared gradients to scale each parameter's step. It still needs a learning rate and can overfit. SGD with momentum trains deep CNNs well.

  8. 8

    A deep sigmoid network trains very slowly in its early layers. Which change is the LEAST likely to help?

    • A

      Switch hidden activations to ReLU

    • B

      Add residual connections

    • Make the sigmoid derivative smaller by scaling the inputs up

    • D

      Use batch normalization

    Scaling inputs up pushes sigmoids further into saturation, shrinking derivatives even more. ReLU, batch norm and skip connections all improve gradient flow.

  9. 9

    Training accuracy is 99.5% and validation accuracy 82%. Which is the best first step?

    • A

      Remove normalization layers

    • Add dropout/weight decay or more data augmentation

    • C

      Increase the number of layers

    • D

      Train for more epochs

    A big train/validation gap means overfitting, so regularize or get more varied data. More epochs or more layers would widen the gap.

  10. 10

    In a transformer with sequence length n and model width d, why does self-attention scale poorly for long sequences?

    • The attention score matrix has n x n entries, so compute and memory grow quadratically in n

    • B

      The embedding dimension d grows with n

    • C

      The softmax cannot be computed in parallel

    • D

      Positional encodings must be recomputed for every token

    Every token attends to every other token, producing an n-by-n score matrix. This O(n^2) cost motivates efficient attention variants.

  11. 11

    A model has d_k = 64. By what factor are the attention scores divided before the softmax?

    • A

      16

    • B

      64

    • 8

    • D

      4096

    The scale is sqrt(d_k) = sqrt(64) = 8. Dividing by d_k itself (64) is a common mix-up.

  12. 12

    Batch norm vs layer norm: which statement is correct?

    • Batch norm normalizes over the batch for each feature; layer norm normalizes over the features of each example

    • B

      Both normalize over the same axis, they only differ in name

    • C

      Batch norm is used in transformers; layer norm is used in CNNs

    • D

      Layer norm needs large batches; batch norm does not

    Layer norm is independent of batch size, so it suits transformers and sequence models. Batch norm is common in CNNs and depends on batch statistics.

  13. 13

    A variational autoencoder differs from a plain autoencoder mainly because it:

    • A

      Has more layers

    • Encodes each input as a distribution and adds a KL term to keep the latent space smooth

    • C

      Uses no decoder

    • D

      Is trained with labels

    That structure lets you sample new points from the latent space and decode them into plausible new data. A plain autoencoder maps to single points and can have gaps.

  14. 14

    Why might a GAN generator produce only a handful of distinct outputs?

    • Mode collapse: it finds a few samples that fool the discriminator and keeps repeating them

    • B

      Because batch size is too large

    • C

      Because the discriminator is too weak to train

    • D

      Because noise vectors are not random

    The generator is rewarded for fooling the discriminator, not for diversity, so it can settle on a few convincing outputs. This is mode collapse.

  15. 15

    Fine-tuning a large pre-trained model on a small data set: which learning rate strategy is typical?

    • A

      The largest learning rate that does not give NaN

    • B

      A learning rate of exactly 1

    • A small learning rate (often smaller than when training from scratch)

    • D

      Zero learning rate for the whole network

    Large steps would destroy the useful pre-trained features. Fine-tuning usually uses small learning rates, sometimes smaller for early layers.