Deep Learning for Computer Vision · End Term · Jan 2024 · AN

Multi-Query attention

Grouped Query attention

Paged attention

None of these       **DL(CV)** **Section Id :** 64065357859 **Section Number :** 8 **Section type :** Online **Mandatory or Optional :** Mandatory **Number of Questions :** 37 **Number of Questions to be attempted :** 37 **Section Marks :** 72 **Display Number Panel :** Yes **Section Negative Marks :** 0 **Group All Questions :** No **Enable Mark as Answered Mark for Review and** Yes **Clear Response :** **Maximum Instruction Time :** 0

The Gaussian blur helps to reduce noise in the image and smooth out pixel intensity variations. The standard deviation parameter controls the amount of blurring applied, with higher values resulting in more aggressive smoothing, which may lead to loss of edge detail.

The Gaussian blur enhances edge contrast in the image and amplifies pixel intensity gradients. The standard deviation parameter determines the width of the blur kernel, with larger values resulting in sharper edges and finer details preserved.

The Gaussian blur serves to highlight high-frequency components in the image and accentuate edge boundaries. The standard deviation parameter dictates the level of detail preserved, with smaller values preserving finer details but potentially amplifying noise.

The Gaussian blur functions to dilate edge regions in the image and expand the edge boundaries. The standard deviation parameter controls the extent of dilation, with higher values resulting in broader edge regions but also increasing the risk of false edge detections.

Scale space representation

Histogram of gradients

Keypoint detection

Orientation assignment

By enabling quick access to all regions in constant time.

By allowing for parallel computation of stability measures across regions.

By organizing regions hierarchically based on stability, facilitating efficient stability analysis.

By providing a compact representation of pixel intensities within each region, reducing computational overhead.

When the dataset contains a high percentage of outliers or erroneous data points.

When the dataset consists of noise-free and well-behaved data with minimal outliers.

When the computational resources are limited, and quick solutions are required.

When the dataset is small, and exact solutions are necessary without considering outliers.

VLAD captures spatial relationships between local descriptors, while BoW does not.

BoW captures spatial relationships between local descriptors, while VLAD does not.

VLAD and BoW both capture spatial relationships between local descriptors.

Neither VLAD nor BoW capture spatial relationships between local descriptors.

Pyramid matching matches images at multiple resolutions using multi- resolution representations, while SURF focuses on detecting and describing local features invariant to scale and rotation.

SURF relies on gradient-based descriptors for feature matching, whereas pyramid matching utilizes histogram-based representations.

Pyramid matching is computationally more efficient compared to SURF, which requires extensive feature extraction and matching procedures.

SURF is more suitable for matching images with complex patterns and cluttered backgrounds, while pyramid matching excels in scenarios where multi-scale matching is critical, such as image registration across different resolutions.

The increased number of parameters and layers in CNNs make backpropagation computationally intensive and prone to overfitting.

The presence of convolutional and pooling layers in CNNs requires the development of specialized backpropagation algorithms for efficient gradient computation.

The non-linear activation functions used in CNNs introduce discontinuities in the error surface, making it difficult to find the global minimum during backpropagation.

The spatial structure of CNNs results in weight sharing and local connectivity, requiring careful consideration of the error propagation process during backpropagation.

Depthwise Separable Convolutions reduce computational complexity and memory footprint by separating the spatial and channel-wise convolutions.

Depthwise Separable Convolutions introduce additional parameters to capture complex spatial relationships within feature maps.

Depthwise Separable Convolutions increase model capacity by performing convolutions across multiple layers of the network.

Depthwise Separable Convolutions enhance feature representation by incorporating multiscale convolutions within the same layer.

Increasing the depth and width of the neural network architecture to capture more complex patterns in the data.

Applying dropout regularization to randomly deactivate neurons during training to prevent co-adaptation of features.

Decreasing the learning rate during fine-tuning to ensure more stable convergence and smoother optimization.

Increasing the batch size during fine-tuning to improve the stability of gradient estimation and accelerate convergence.

It helps identify the specific input features that contribute most strongly to the model’s predictions.

It provides insights into the architectural design choices made in the neural network.

It quantifies the uncertainty associated with the model’s predictions.

It measures the computational resources required to activate the neurons.

The input layer.

The output layer.

The convolutional layers.

The fully connected layers.

The class activation map (CAM).

The global average pooling layer.

The guided backpropagation algorithm.

The gradient of the target class score with respect to the feature maps.

To compute the style loss by measuring the similarity between feature maps from the content and style images.

To generate high-resolution images by applying a series of convolutions and upsampling operations.

To regularize the optimization process by penalizing large gradients in the feature space.

To visualize the internal representations learned by the neural network during training.

They are computationally expensive and can significantly increase inference time.

They may fail to suppress redundant bounding boxes in cluttered scenes with multiple overlapping objects.

They require extensive hyperparameter tuning to achieve optimal performance across different datasets.

They are highly sensitive to variations in lighting conditions and viewpoints in the input images.

Focal loss introduces additional regularization terms to penalize large weights in the neural network.

Focal loss dynamically adjusts the loss weights based on the difficulty of each training example.

Focal loss computes the loss based on the Euclidean distance between predicted and groundtruth bounding boxes.

Focal loss prioritizes minimizing the overall loss across all classes equally.

Siamese networks require significantly larger datasets for training due to their complex architecture.

Siamese networks need to learn a similarity metric between pairs of input samples, which requires specialized loss functions and training procedures.

Siamese networks are computationally more intensive, making them less scalable for real-time applications.

Siamese networks tend to suffer from overfitting more frequently than traditional neural networks due to their increased model complexity.

Both Assertion and Reason are True; Reason is a correct explanation of the Assertion

Both Assertion and Reason are True; but Reason is not a correct explanation of the Assertion

Assertion is True; Reason is False

Assertion is False by itself

Reparametrization trick is used in training VAEs to overcome the problem of non-differentiability of the objective function across a sampling step.

In denoising autoencoder, reconstructed output is compared against noise- free ground truth data point.

While training a GAN, Generator and Discriminator weights are not updated concurrently but updated sequentially.

When a GAN is completely trained, Discriminator D outputs very high probability (approximately 1) for images from Generator G, that is, D(G(z)) ≈ 1

Continual learning

Zero-shot learning

Few-shot learning

Unsupervised learning

[1.1, 2.06, 1.08]

[0.25, 0.34, 0.18]

[0.2, 0.3, 0.1]

[1.2, 2.3, 1.1]

It allows for generating high-resolution images with unlimited variation.

It eliminates the need for using mini-batches during training.

It enables the discovery of large-scale structures before finer-scale details.

It ensures all layers in the networks remain fixed throughout the training process.

TRUE

FALSE

Maximum likelihood estimate

Minimum likelihood estimate

Maximizing mean square error

Minimizing mean square error

Maximizing negative log likelihood

Minimizing negative log likelihood