Deep Learning for Computer Vision End Term: 13 September 2026 (May 2026 term)

1. A

   36,864

2. B

   73,728

3. C

   147,456

4. D

   294,912

1. A

   [2,-1,3]

2. B

   [-2,1,-3]

3. C

   [3,-1,2]

4. D

   [3,1,2]

1. A

   Along the tangent direction of the edge

2. B

   Approximately normal to the edge

3. C

   Along the direction of maximum smoothing

4. D

   Along the local isophote

1. A

2. B

3. C

4. D

1. A

   Detecting candidate keypoints across multiple image scales

2. B

   Expressing local gradient orientations relative to an assigned keypoint orientation

3. C

   Normalizing the descriptor vector after histogram construction

4. D

   Building an image pyramid without assigning an orientation

1. A

   Taking the larger of the two branch gradients

2. B

   Summing the gradient contributions from the two branches

3. C

   Multiplying the two branch gradients

4. D

   Averaging the two branch gradients regardless of the graph

1. A

   Cancel updates in that direction

2. B

   Build velocity in that direction while smoothing short-term gradient fluctuations

3. C

   Make the gradient exactly unbiased

4. D

   Force every parameter to use the same accumulated gradient

1. A

   The model is necessarily underfitting

2. B

   Generalization is worsening even though optimization on the training set is improving

3. C

   The training objective is necessarily non-differentiable

4. D

   The model has reached the global optimum on unseen data

1. A

   Doubles

2. B

   Quadruples

3. C

   Remains unchanged

4. D

   Is halved

1. A

   A contribution from only the central output location

2. B

   Contributions accumulated from all output locations where that coefficient was used

3. C

   Only the gradient of the bias corresponding to that output channel

4. D

   A contribution from exactly one spatial location chosen during the forward pass

1. A

   It forces F(x) to be constant

2. B

   The block can approach an identity mapping when the residual F(x) approaches zero

3. C

   It prevents gradients from entering the residual branch

4. D

   It makes every convolution invertible

1. A

   Eliminates channel mixing from the layer

2. B

   Factorizes spatial filtering and channel mixing into cheaper operations

3. C

   Uses a larger spatial kernel with the same parameter count

4. D

   Forces the number of output channels to equal the number of input channels

1. A

   Reinitialize the entire network and use a very large learning rate

2. B

   Retain pretrained features and update them cautiously, often with a smaller learning rate than newly initialized layers

3. C

   Freeze the newly added prediction head and update only the old classifier

4. D

   Randomly permute pretrained channels before optimization

1. A

   Dense prediction of final detections in one stage with no proposal-processing step

2. B

   Generation of candidate regions followed by prediction on those candidate regions

3. C

   Pixel-wise class prediction without object localization

4. D

   Sequence-level prediction using recurrent hidden states

1. A

   One C-dimensional vector for the entire image only

2. B

   A spatial grid with C class scores at each output location

3. C

   One bounding box for each of the C classes

4. D

   A sequence with one recurrent state per class and no spatial layout

1. A

   Remove nonlinearities from the recurrent computation

2. B

   Regulate information retention and update across timesteps

3. C

   Make the model invariant to permutation of the sequence

4. D

   Replace recurrent state with a fixed convolution kernel

1. A

   The number of tokens in the sequence

2. B

   The attention weights used to combine the value vectors for that token

3. C

   The dimensionality of every value vector

4. D

   The number of transformer layers

1. A

   Treat the complete image batch as a single token

2. B

   Divide the image into patches and map the patches to token embeddings

3. C

   Replace each patch with its ground-truth class before attention

4. D

   Use only the global mean RGB value as the token sequence

1. A

   A GAN necessarily contains an explicit encoder that outputs a Gaussian posterior

2. B

   A VAE uses a latent-variable objective with a regularized approximate posterior, whereas a GAN uses an adversarial training objective

3. C

   A VAE is trained only through a discriminator

4. D

   GANs and VAEs have the same training objective but different optimizers

1. A

   Predictions from two external classifiers trained on different label sets

2. B

   Conditional and unconditional denoising predictions from the diffusion model

3. C

   The forward-noise sample and a segmentation mask with equal weights

4. D

   Two independently sampled latent codes without conditioning information

Select all that apply.

1. A

   A normalized averaging filter preserves a constant-valued region away from boundary effects.

2. B

   Increasing Gaussian scale suppresses progressively finer image structure.

3. C

   Image pyramids represent image information at multiple spatial scales.

4. D

   Cross-correlation always flips the filter before applying it.

Select all that apply.

1. A

   Backpropagation applies the chain rule efficiently through a computation graph.

2. B

   Regularization can change the learned solution even when the network architecture is unchanged.

3. C

   Gradient-descent variants can modify how current and past gradient information is used to update parameters.

4. D

   Improving training means validation performance must increase after every individual parameter update.

Select all that apply.

1. A

   Weight sharing allows the same convolution kernel parameters to be used at multiple spatial locations.

2. B

   Inception-style designs can use multiple processing branches.

3. C

   Residual connections create shortcut paths across learned transformations.

4. D

   Finetuning requires discarding all pretrained weights before optimization starts.

Select all that apply.

1. A

   Two-stage detectors separate candidate-region generation from later prediction on candidates.

2. B

   Single-stage detectors can produce final detection predictions without a separate proposal-processing stage.

3. C

   Segmentation is a dense prediction task with spatially resolved outputs.

4. D

   Object detection and segmentation necessarily produce identical output representations.

Select all that apply.

1. A

   Recurrent models can process ordered sequences of visual features.

2. B

   Backpropagation through a recurrent model propagates gradient information across unrolled timesteps.

3. C

   LSTMs and GRUs use gating mechanisms.

4. D

   Randomly permuting all video frames leaves every temporal model output unchanged by definition.

Select all that apply.

1. A

   Soft attention can form a weighted combination of visual features.

2. B

   Self-attention allows token representations to depend directly on other tokens.

3. C

   Vision Transformers adapt transformer-style token processing to visual inputs.

4. D

   Transformer-based segmentation is restricted to producing one class label for an entire image.

Select all that apply.

1. A

   GAN training involves adversarial objectives.

2. B

   VAEs introduce latent variables and regularize an approximate posterior.

3. C

   DDPMs use a noise-based forward process together with a learned reverse denoising process.

4. D

   Classifier-free guidance requires a separate external classifier at sampling time.

Select all that apply.

1. A

   SimCLR is a self-supervised contrastive representation-learning method.

2. B

   MoCo is a contrastive representation-learning method.

3. C

   CLIP learns relationships between visual and textual representations.

4. D

   In SimCLR, a positive pair is formed by two unrelated images solely because they receive the same predicted class.