MauryaHub PYQ Practice

da5006_2025T3_ET_FN.pdf

Deep Learning for Computer Vision · End Term · Sep 2025 FN

← Course papers · Start practice / exam

Questions and published explanations below are available without starting a test. Some questions may not have a published solution yet.

Question 2 NAT · 1.0 marks

[[IMAGE:6f44f6e4b1f8fc76_2_2]]
Source diagram or notation

    A published solution is not available for this question yet.

    Question 3 NAT · 1.0 marks

    [[IMAGE:6f44f6e4b1f8fc76_2_3]]
    Source diagram or notation

      A published solution is not available for this question yet.

      Question 4 MCQ · 1.0 marks

      [[IMAGE:6f44f6e4b1f8fc76_3_4]]
      Source diagram or notation
      1. Passing patch embeddings through the Transformer encoder
      2. Linear projection of flattened image patches
      3. Adding a class token
      4. Layer normalization of patch embeddings

      A published solution is not available for this question yet.

      Question 5 MCQ · 1.0 marks

      [[IMAGE:6f44f6e4b1f8fc76_3_5]]
      Source diagram or notation
      1. Flattens the softmax distribution, reducing gradient magnitude
      2. sharpens the softmax distribution, increasing emphasis on hard negatives
      3. has no effect on the loss landscape
      4. makes image and text embeddings orthogonal

      A published solution is not available for this question yet.

      Question 6 MCQ · 1.0 marks

      In a Feature Pyramid Network (FPN), the primary purpose of the top-down pathway with lateral connections is to:
      1. increase the receptive field at the lowest-resolution level
      2. enrich high-resolution feature maps with strong semantics from deeper layers
      3. perform non-maximum suppression at multiple scales
      4. reduce the number of anchors required per level

      A published solution is not available for this question yet.

      Question 7 MCQ · 1.0 marks

      Consider the following pyTorch code: [[IMAGE:6f44f6e4b1f8fc76_4_6]]
      Source diagram or notation
      1. tensor([0.09, 0.24, 0.67])
      2. tensor([0.09, 0.24, 0.66])
      3. tensor([0.09, 0.23, 0.68])
      4. tensor([0.10, 0.27, 0.63])

      A published solution is not available for this question yet.

      Question 8 MCQ · 1.0 marks

      [[IMAGE:6f44f6e4b1f8fc76_4_7]]
      Source diagram or notation
      1. Use the test set to select the best hyperparameters
      2. Use k-fold cross-validation on the training data and keep the test set untouched for final re-porting
      3. Tune on the training loss only; validation is unnecessary
      4. Pick hyperparameters from a prior paper without checking performance

      A published solution is not available for this question yet.

      Question 9 MCQ · 1.0 marks

      How is disentangled representation evaluated?:
      1. Using Mutual Information Gap
      2. Using mean Average Precision
      3. KL divergence
      4. Sparsity in the latent one-hot codes

      A published solution is not available for this question yet.

      Question 10 MSQ · 1.0 marks

      Which statements about generative models are true? (Select all that apply.)
      1. VAEs optimize a lower bound on the data log-likelihood.
      2. GANs perform explicit maximum likelihood estimation.
      3. Diffusion models can trade off speed and quality via the number of sampling steps.
      4. Autoregressive models factorize the joint distribution into a product of conditionals.

      A published solution is not available for this question yet.

      Question 11 MSQ · 1.0 marks

      In classifier-free guidance for diffusion models, increasing the guidance scale γ tends to: (Select all that apply.)
      1. improve fidelity to the conditioning signal
      2. always reduce mode collapse to zero
      3. reduce sample diversity at very high values
      4. be equivalent to classifier guidance

      A published solution is not available for this question yet.

      Question 12 MSQ · 1.0 marks

      Which statements are correct? (Select all that apply.)
      1. A linear autoencoder with MSE and no activation learns a subspace equivalent to PCA (up to rotation) when the hidden dimension < input dimension.
      2. Global attention assigns alignment weights to all encoder time steps for each decoder step.
      3. Positional encodings convey order information that self-attention lacks.
      4. it is impossible to generate multiple, semantically similar captions with different styles for the same image.

      A published solution is not available for this question yet.

      Question 13 NAT · 1.0 marks

      [[IMAGE:6f44f6e4b1f8fc76_6_8]]
      Source diagram or notation

        A published solution is not available for this question yet.

        Question 14 NAT · 1.0 marks

        [[IMAGE:6f44f6e4b1f8fc76_6_9]]
        Source diagram or notation

          A published solution is not available for this question yet.

          Question 15 NAT · 1.0 marks

          [[IMAGE:6f44f6e4b1f8fc76_7_10]]
          Source diagram or notation

            A published solution is not available for this question yet.

            Question 16 MCQ · 1.0 marks

            Mixup data augmentation primarily:
            1. reduces label noise by hardening labels
            2. creates convex combinations of inputs and labels to encourage linear behavior between classes
            3. prunes redundant training samples
            4. increases input resolution without changing model capacity

            A published solution is not available for this question yet.

            Question 17 NAT · 1.0 marks

            [[IMAGE:6f44f6e4b1f8fc76_8_11]]
            Source diagram or notation

              A published solution is not available for this question yet.

              Question 18 MCQ · 1.0 marks

              During training, Batch Normalization uses:
              1. Running (population) statistics only
              2. batch statistics for normalization and updates running estimates for inference
              3. neither batch nor running statistics
              4. group-wise normalization identical to GroupNorm

              A published solution is not available for this question yet.

              Question 19 MSQ · 1.0 marks

              Which techniques help mitigate exploding/vanishing gradients in RNNs? (Select all that apply.)
              1. Gradient clipping
              2. Orthogonal or identity initialization of recurrent weights
              3. Using LSTM/GRU cells
              4. Randomly reversing input sequences as augmentation

              A published solution is not available for this question yet.

              Question 20 NAT · 1.0 marks

              [[IMAGE:6f44f6e4b1f8fc76_9_12]]
              Source diagram or notation

                A published solution is not available for this question yet.