You are on page 1of 12

Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 1

Computer Vision
1. What is Computer Vision?
a. A type of artificial intelligence b. A hardware component in a computer c. A programming language d.
A computer monitor
Answer: a. A type of artificial intelligence
2. Which of the following is NOT a typical application of Computer Vision?
a. Object recognition b. Autonomous driving c. Speech recognition d. Facial recognition
Answer: c. Speech recognition
3. What is the primary goal of feature extraction in Computer Vision?
a. To detect objects b. To enhance image resolution c. To reduce image size d. To identify patterns and
key information in an image
Answer: d. To identify patterns and key information in an image
4. Which of the following is a commonly used deep learning framework for Computer Vision tasks?
a. TensorFlow b. Microsoft Word c. Adobe Photoshop d. Microsoft Excel
Answer: a. TensorFlow
5. What is image segmentation in Computer Vision?
a. Classifying images into categories b. Detecting and tracking objects in images c. Dividing an image into
meaningful regions d. Enhancing the brightness of an image
Answer: c. Dividing an image into meaningful regions
6. In Computer Vision, what does OCR stand for?
a. Optical Character Recognition b. Object Classification and Recognition c. Object Character Rendering d.
Optical Code Recognition
Answer: a. Optical Character Recognition
7. Which type of neural network architecture is commonly used for image classification in Computer Vision?
a. Convolutional Neural Network (CNN) b. Recurrent Neural Network (RNN) c. Multilayer Perceptron (MLP)
d. Decision Tree
Answer: a. Convolutional Neural Network (CNN)
8. What is the purpose of data augmentation in Computer Vision?
a. To reduce the size of image datasets b. To improve the performance of machine learning models c. To
remove noise from images d. To resize images to a standard resolution
Answer: b. To improve the performance of machine learning models
9. Which image file format is lossless and commonly used for storing images in Computer Vision
applications?
a. JPEG b. PNG c. GIF d. BMP
Answer: b. PNG
10. What is the "vanishing gradient problem" in deep learning, and how is it relevant to Computer Vision?

a. It's the problem of fading colors in images. b. It's the problem of gradients becoming too small during
backpropagation, which can hinder training deep neural networks, including those used in Computer
Vision. c. It's the problem of objects disappearing from images. d. It's the problem of images getting
distorted during image processing.

Answer: b. It's the problem of gradients becoming too small during backpropagation, which can
hinder training deep neural networks, including those used in Computer Vision.

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 2

CNN

1. What does CNN stand for in the context of deep learning?


a. Continuous Neural Network b. Convolutional Neural Network c. Centralized Neural Network d.
Computer Network Neural
Answer: b. Convolutional Neural Network
2. What is the primary purpose of a CNN in machine learning?
a. Natural language processing b. Image and video analysis c. Speech recognition d. Tabular data analysis
Answer: b. Image and video analysis
3. In CNNs, what is the main advantage of using convolutional layers?
a. Faster training times b. Improved image resolution c. Automatic labeling of images d. Feature extraction
and pattern recognition
Answer: d. Feature extraction and pattern recognition
4. Why are CNNs particularly well-suited for image classification tasks?
a. They can handle text and numeric data. b. They use fully connected layers for image analysis. c. They
leverage local spatial dependencies in images. d. They require large amounts of labeled data.
Answer: c. They leverage local spatial dependencies in images.
5. Which layer type in a CNN is responsible for reducing the spatial dimensions of the feature maps?
a. Convolutional layer b. Pooling layer c. Fully connected layer d. Activation layer
Answer: b. Pooling layer
6. What does the term "stride" refer to in a convolutional layer?
a. The amount of dropout applied to the layer b. The size of the filter used for convolution c. The step size
for sliding the filter over the input d. The learning rate during training
Answer: c. The step size for sliding the filter over the input
7. Why is the need for feature hierarchies in CNNs important for image analysis?
a. It simplifies the network architecture. b. It allows for more filters in each layer. c. It enables the detection
of complex features built upon simpler ones. d. It reduces the training time.
Answer: c. It enables the detection of complex features built upon simpler ones.
8. What role do activation functions play in CNNs?
a. They determine the learning rate. b. They regulate the number of convolutional layers. c. They introduce
non-linearity into the network. d. They control the filter size.
Answer: c. They introduce non-linearity into the network.
9. How do CNNs help address the vanishing gradient problem in deep learning?
a. By increasing the learning rate b. By using a different type of network architecture c. By introducing skip
connections d. By reducing the number of layers
Answer: c. By introducing skip connections
10. What is the primary need for CNNs in computer vision and image processing?
a. Speed up image loading times b. Handle unstructured image data c. Reduce the amount of labeled
training data required d. Improve text recognition in images

Answer: b. Handle unstructured image data

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 3

Design space for Convolutional nets

1. What does the term "design space" refer to in the context of CNNs?
a. The physical space where CNNs are implemented b. The range of possible architectural choices and
hyperparameters for CNNs c. The space of image dimensions CNNs can handle d. The space of pre-
trained CNN models
Answer: b. The range of possible architectural choices and hyperparameters for CNNs
2. In the design of a CNN, what does the term "kernel size" refer to?
a. The size of the entire network b. The size of the dataset being used c. The size of the convolutional filter
used in the convolution layer d. The size of the output layer
Answer: c. The size of the convolutional filter used in the convolution layer
3. Which of the following is NOT a typical architectural decision in the design of a CNN?
a. Number of layers b. Learning rate c. Activation functions d. Pooling strategies
Answer: b. Learning rate
4. What is the purpose of using dropout in a CNN?
a. To increase the number of layers b. To remove low-intensity pixels from images c. To prevent overfitting
by randomly deactivating neurons during training d. To enhance the resolution of feature maps
Answer: c. To prevent overfitting by randomly deactivating neurons during training
5. Which hyperparameter defines the step size for moving the convolutional filter over the input in a CNN?
a. Filter size b. Stride c. Pooling size d. Padding
Answer: b. Stride
6. In the design of a CNN, what is the role of batch normalization?
a. Normalizing the batch size b. Standardizing the input data to have zero mean and unit variance c.
Reducing the number of layers d. Increasing the learning rate
Answer: b. Standardizing the input data to have zero mean and unit variance
7. Which architectural component of a CNN helps reduce the spatial dimensions of feature maps while
retaining important information?
a. Activation function b. Fully connected layer c. Pooling layer d. Convolutional layer
Answer: c. Pooling layer
8. What is the purpose of using skip connections or residual connections in the design of a CNN?
a. To create shortcut paths for gradients to flow, which helps mitigate the vanishing gradient problem b.
To increase the number of layers for more complex feature extraction c. To add more convolutional filters
to the network d. To reduce the computational complexity
Answer: a. To create shortcut paths for gradients to flow, which helps mitigate the vanishing
gradient problem
9. What is the concept of transfer learning in the context of CNN design?
a. Training a CNN from scratch with random weights b. Transferring weights from one CNN to another
with different architectures c. Freezing all layers of a pre-trained model d. Fine-tuning a pre-trained model
on a new task
Answer: d. Fine-tuning a pre-trained model on a new task
10. In the design space of CNNs, what does the term "hyperparameter tuning" refer to?
a. Adjusting the architecture of the CNN b. Tuning the learning rate to the optimal value c. Fine-tuning the
number of layers d. Optimizing the values of parameters like batch size, dropout rate, and filter sizes

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 4

Pooling layer motivation in CNN

1. What is the primary motivation for including pooling layers in CNNs?


a. To increase the number of parameters in the network b. To speed up the training process c. To reduce
the spatial dimensions of feature maps while retaining important information d. To add non-linearity to
the network
Answer: c. To reduce the spatial dimensions of feature maps while retaining important information
2. Which of the following statements best describes the role of pooling layers in CNNs?
a. Pooling layers extract feature maps from input data. b. Pooling layers learn to recognize specific
features in images. c. Pooling layers downsample feature maps, reducing the computational burden. d.
Pooling layers introduce non-linearity into the network.
Answer: c. Pooling layers downsample feature maps, reducing the computational burden.
3. Why is downsampling important in the design of CNNs?
a. To increase the number of feature maps b. To increase the resolution of feature maps c. To decrease the
size of feature maps and improve computation efficiency d. To make the network deeper
Answer: c. To decrease the size of feature maps and improve computation efficiency
4. What is the advantage of max pooling over other pooling techniques like average pooling?
a. Max pooling retains the maximum values in each pooling region, preserving more salient features. b.
Max pooling is faster to compute. c. Average pooling introduces non-linearity. d. Average pooling reduces
the risk of overfitting.
Answer: a. Max pooling retains the maximum values in each pooling region, preserving more
salient features.
5. What is the typical size of the pooling window in a max pooling layer?
a. 3x3 b. 1x1 c. 2x2 d. 5x5
Answer: c. 2x2
6. How does the use of pooling layers contribute to translational invariance in CNNs?
a. Pooling layers introduce translation operations. b. Pooling layers average out the translation-related
variations. c. Pooling layers amplify the translation-related variations. d. Pooling layers have no effect on
translational invariance.
Answer: b. Pooling layers average out the translation-related variations.
7. What is the primary drawback of using too much pooling in a CNN?
a. Increased computational complexity b. Reduced ability to recognize important features c. Higher risk of
overfitting d. Improved feature representation
Answer: b. Reduced ability to recognize important features
8. In which layers of a CNN are pooling layers typically inserted?
a. At the beginning of the network b. After every convolutional layer c. After fully connected layers d.
Periodically throughout the network
Answer: b. After every convolutional layer
9. How does the choice of pooling size (e.g., 2x2 or 3x3) affect the information retained in feature maps?
a. Larger pooling sizes retain more information. b. Smaller pooling sizes retain more information. c.
Pooling size does not impact information retention. d. Information retention is solely determined by the
type of pooling used.
Answer: a. Larger pooling sizes retain more information.
10. Which of the following is NOT a reason for using pooling layers in CNNs?
a. Reducing computational complexity b. Maintaining spatial resolution c. Improving computational
efficiency d. Promoting translational invariance

Answer: b. Maintaining spatial resolution

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 5

Design a convolution layered application

1. What is the primary purpose of the convolutional layers in a CNN-based application?


a. Feature extraction b. Final decision making c. Data preprocessing d. Data visualization
Answer: a. Feature extraction
2. In the design of a CNN, which layer type is responsible for down-sampling feature maps to reduce
computational load?
a. Convolutional layer b. Activation layer c. Pooling layer d. Fully connected layer
Answer: c. Pooling layer
3. What is the typical role of activation functions in a convolutional layered application?
a. Reducing the number of layers in the network b. Introducing non-linearity into the model c. Controlling
the learning rate d. Increasing the size of convolutional filters
Answer: b. Introducing non-linearity into the model
4. Which of the following is a hyperparameter often adjusted during the design of convolutional layers in a
CNN?
a. Learning rate b. Filter size c. Batch size d. Pooling strategy
Answer: b. Filter size
5. Why is it important to add dropout layers in a CNN-based application?
a. To increase the number of layers b. To enhance image resolution c. To prevent overfitting by
deactivating neurons during training d. To improve feature extraction
Answer: c. To prevent overfitting by deactivating neurons during training
6. What is the main motivation behind using batch normalization in CNNs?
a. Increasing computational complexity b. Standardizing input data to improve training stability c.
Reducing the number of layers in the network d. Adding non-linearity to the model
Answer: b. Standardizing input data to improve training stability
7. Which layer type in a CNN typically connects to fully connected layers for making the final decisions?
a. Convolutional layer b. Activation layer c. Pooling layer d. Flatten layer
Answer: d. Flatten layer
8. In the design of a CNN, what is the purpose of a stride in the convolutional layers?
a. To control the learning rate b. To increase the size of the feature maps c. To specify the filter size d. To
determine how the filter moves across the input
Answer: d. To determine how the filter moves across the input
9. What is the primary goal of transfer learning in a CNN-based application?
a. Training a CNN from scratch with random weights b. Transferring weights from one CNN to another
with different architectures c. Freezing all layers of a pre-trained model d. Fine-tuning a pre-trained model
for a new task
Answer: d. Fine-tuning a pre-trained model for a new task
10. What is the primary objective of hyperparameter tuning when designing convolutional layers in a CNN?
a. Increasing the number of layers in the network b. Maximizing the size of convolutional filters c.
Optimizing the values of hyperparameters like learning rate, batch size, and dropout rate d. Reducing the
non-linearity in the network
Answer: c. Optimizing the values of hyperparameters like learning rate, batch size, and dropout
rate

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 6

Understanding and visualizing a CNN

1. What is the primary purpose of visualization techniques in understanding a CNN?


a. To create visually appealing models b. To uncover the inner workings of the network and interpret its
decisions c. To generate synthetic images for training data d. To enhance the performance of the network
Answer: b. To uncover the inner workings of the network and interpret its decisions
2. What does a CNN learn in its early convolutional layers?
a. High-level concepts and complex features b. Low-level features like edges, corners, and textures c. Fully
connected layers d. Pooling strategies
Answer: b. Low-level features like edges, corners, and textures
3. What is the main purpose of feature visualization in CNNs?
a. To visualize the entire dataset b. To explore the network architecture c. To generate synthetic images d.
To understand what activates specific neurons or filters
Answer: d. To understand what activates specific neurons or filters
4. What is the term used to describe the process of calculating the gradient of the loss with respect to the
input image?
a. Feature extraction b. Backpropagation c. Activation mapping d. Gradient ascent
Answer: d. Gradient ascent
5. How is a heat map generated in the context of CNN visualization?
a. By averaging the feature maps of all layers b. By plotting the color distribution of an image c. By
displaying areas of an image that most strongly activate a specific neuron d. By using grayscale images
Answer: c. By displaying areas of an image that most strongly activate a specific neuron
6. In a CNN visualization, what does it mean when a neuron in a specific layer activates strongly in response
to a certain feature in an image?
a. The feature is irrelevant to the network's task. b. The feature is important for the network's task. c. The
network is malfunctioning. d. The feature is related to the filter size.
Answer: b. The feature is important for the network's task.
7. How can saliency maps be used in CNN visualization?
a. To identify the least important features in an image b. To highlight the most visually appealing regions
in an image c. To show which parts of an image contributed most to the network's decision d. To visualize
the entire dataset
Answer: c. To show which parts of an image contributed most to the network's decision
8. What is the primary objective of deconvolutional networks (also known as upsampling or transposed
convolution) in CNN visualization?
a. To create blurry images b. To generate synthetic data c. To visualize the original input image d. To map
activations back to the input space
Answer: d. To map activations back to the input space
9. When visualizing CNNs, which technique can be used to generate synthetic images that maximally
activate a specific neuron?
a. Backpropagation b. Gradient ascent c. Activation mapping d. Heat mapping
Answer: b. Gradient ascent
10. In CNN visualization, what is the primary benefit of techniques like t-SNE and PCA (Principal Component
Analysis)?
a. They enhance image resolution. b. They reduce the number of layers in the network. c. They help
visualize high-dimensional feature representations in lower dimensions. d. They improve the performance
of CNNs.
Answer: c. They help visualize high-dimensional feature representations in lower dimensions.

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 7

Transfer Learning

1. What is Transfer Learning in the context of machine learning?


a. Transferring data between different devices b. Using knowledge gained from one task to improve
performance on a related task c. Migrating machine learning models to a different framework d. Sharing
model weights between competing companies
Answer: b. Using knowledge gained from one task to improve performance on a related task
2. In Transfer Learning, what is the source domain?
a. The domain of the target task b. The domain of the model architecture c. The domain of the model's
training data d. The domain of the pre-trained model
Answer: d. The domain of the pre-trained model
3. Which type of layer is often modified or added when applying Transfer Learning to a pre-trained model?
a. Pooling layer b. Fully connected layer c. Convolutional layer d. Dropout layer
Answer: b. Fully connected layer
4. What is fine-tuning in Transfer Learning?
a. Adapting a pre-trained model to a completely unrelated task b. Keeping all pre-trained layers fixed and
training only the top layers for a new task c. Using the pre-trained model as is, without any adjustments d.
Rewriting the entire model from scratch
Answer: b. Keeping all pre-trained layers fixed and training only the top layers for a new task
5. In Transfer Learning, what is domain adaptation?
a. Adapting a model to a different hardware domain b. Adapting a model to a different model
architecture c. Adapting a model to a different dataset distribution d. Adapting a model to a different
programming language
Answer: c. Adapting a model to a different dataset distribution
6. Which famous deep learning model pre-trained on a large dataset is often used for Transfer Learning in
various applications?
a. LeNet b. AlexNet c. VGGNet d. ImageNet
Answer: d. ImageNet
7. What is the primary advantage of Transfer Learning in terms of model training?
a. It always leads to a better-performing model. b. It speeds up the training process. c. It requires a larger
dataset for the target task. d. It doesn't require any additional training.
Answer: b. It speeds up the training process.
8. When should you consider using Transfer Learning?
a. Only when you have a massive amount of labeled data for the target task b. When you have a related
task with a limited dataset, and a pre-trained model is available c. When you want to avoid any additional
training for your model d. Only when the source domain is exactly the same as the target domain
Answer: b. When you have a related task with a limited dataset, and a pre-trained model is
available
9. Which transfer learning strategy is suitable when the source and target tasks are related but not exactly
the same?
a. Fine-tuning b. Domain adaptation c. Feature extraction d. Rewriting
Answer: a. Fine-tuning
10. What is the potential drawback of using Transfer Learning with a pre-trained model?
a. It always results in a worse performance compared to training from scratch. b. It may not be compatible
with the target task, leading to suboptimal results. c. It significantly increases the required training data. d.
It reduces the model's capacity for new tasks.
Answer: b. It may not be compatible with the target task, leading to suboptimal results.

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 8

Fine-tunning in CNN

1. What is fine-tuning in the context of CNNs?


a. Refining the details of CNN architectures b. Training a CNN from scratch with random weights c.
Adjusting the learning rate during training d. Adapting a pre-trained CNN for a new task
Answer: d. Adapting a pre-trained CNN for a new task
2. What is the primary advantage of fine-tuning a pre-trained CNN?
a. Faster convergence and reduced training time b. Increased model complexity c. Improved
generalization to new tasks d. Lower computational requirements
Answer: a. Faster convergence and reduced training time
3. In fine-tuning, which layers of a pre-trained CNN are typically modified or retrained for the new task?
a. All layers b. None of the layers c. Only the convolutional layers d. Only the fully connected layers
Answer: d. Only the fully connected layers
4. What is the general approach for fine-tuning a pre-trained CNN?
a. Train only the new layers added for the specific task. b. Train all layers from scratch. c. Use a higher
learning rate for the pre-trained layers. d. Remove the pre-trained layers and replace them with new ones.
Answer: a. Train only the new layers added for the specific task.
5. Why is fine-tuning recommended over training a CNN from scratch for most tasks?
a. Fine-tuning is less computationally intensive. b. Fine-tuning usually results in higher accuracy. c.
Training from scratch doesn't require labeled data. d. Fine-tuning doesn't require a pre-trained model.
Answer: b. Fine-tuning usually results in higher accuracy.
6. When fine-tuning a pre-trained CNN, what should be done with the pre-trained layers?
a. Freeze them (keep them fixed) b. Remove them completely c. Increase their learning rate d. Train them
from scratch
Answer: a. Freeze them (keep them fixed)
7. What is the primary consideration when selecting a pre-trained CNN model for fine-tuning?
a. The age of the model b. The size of the model c. The source domain of the pre-trained model d. The
number of layers in the model
Answer: c. The source domain of the pre-trained model
8. In fine-tuning, what is the primary challenge when the source and target tasks are significantly different?
a. The challenge of training the convolutional layers from scratch b. The challenge of adapting the fully
connected layers c. The challenge of finding labeled data for the source task d. The challenge of
transferring knowledge effectively
Answer: d. The challenge of transferring knowledge effectively
9. Which technique can be used to avoid overfitting when fine-tuning a pre-trained CNN?
a. Removing the pre-trained layers b. Reducing the batch size during training c. Adding dropout layers d.
Increasing the learning rate
Answer: c. Adding dropout layers
10. What is the recommended approach when fine-tuning a pre-trained CNN for a new task with limited
labeled data?
a. Increase the model complexity by adding more layers. b. Train the entire model from scratch. c. Use
data augmentation techniques to artificially increase the dataset size. d. Reduce the number of epochs
during training.
Answer: c. Use data augmentation techniques to artificially increase the dataset size.

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 9

Image Classification

1. What is the primary task of image classification using CNNs?


a. Detecting objects in images b. Labeling images into predefined categories or classes c. Generating
realistic images from textual descriptions d. Enhancing image resolution
Answer: b. Labeling images into predefined categories or classes
2. What is the role of convolutional layers in a CNN for image classification?
a. They apply image transformations. b. They classify images. c. They extract features from the input
images. d. They reduce image resolution.
Answer: c. They extract features from the input images.
3. In image classification using CNNs, what is the purpose of pooling layers?
a. To increase image resolution b. To learn high-level features c. To reduce the size of feature maps d. To
apply activation functions
Answer: c. To reduce the size of feature maps
4. What does the softmax activation function do in the final layer of an image classification CNN?
a. It reduces overfitting. b. It converts raw scores into class probabilities. c. It applies batch normalization.
d. It enhances feature extraction.
Answer: b. It converts raw scores into class probabilities.
5. Which loss function is commonly used for image classification tasks in CNNs?
a. Mean Absolute Error (MAE) b. Mean Squared Error (MSE) c. Cross-Entropy Loss d. Hinge Loss
Answer: c. Cross-Entropy Loss
6. What is the term for a technique that helps prevent overfitting in CNNs during training for image
classification?
a. Data augmentation b. Pooling c. Batch normalization d. Dropout
Answer: d. Dropout
7. What is the role of the fully connected layers at the end of a CNN for image classification?
a. They perform convolution operations. b. They apply the softmax activation function. c. They reduce
image resolution. d. They make the final class predictions.
Answer: d. They make the final class predictions.
8. What is the purpose of data augmentation in image classification using CNNs?
a. To decrease the amount of training data needed b. To reduce computational complexity c. To create
synthetic images for the dataset d. To remove noise from images
Answer: a. To decrease the amount of training data needed
9. In transfer learning for image classification, what is typically transferred from a pre-trained model to the
new model?
a. Entire model architecture b. Learning rate c. Weights and learned features d. Image resolution
Answer: c. Weights and learned features
10. What is the advantage of using deep CNN architectures, like ResNet or VGG, for image classification?
a. Faster training times b. Simpler model designs c. Better handling of small image datasets d. Improved
feature extraction and classification accuracy
Answer: d. Improved feature extraction and classification accuracy

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 10

Text Classification

1. What is the primary goal of text classification?


a. Summarizing text documents b. Identifying the language of a text c. Categorizing text into predefined
classes or categories d. Extracting named entities from text
Answer: c. Categorizing text into predefined classes or categories
2. In text classification, what is the input data usually represented as?
a. Images b. Vectors c. Audio files d. Graphs
Answer: b. Vectors
3. Which machine learning algorithm is commonly used for text classification?
a. K-Means Clustering b. Linear Regression c. Naive Bayes d. Decision Trees
Answer: c. Naive Bayes
4. In text classification, what is a common preprocessing step to convert text data into numerical features?
a. Tokenization b. Stemming c. Sentiment analysis d. Image embedding
Answer: a. Tokenization
5. What is the term for a dataset used for training and evaluating text classification models, where each text
document is labeled with a category?
a. Corpus b. Lexicon c. Embedding d. Training set
Answer: d. Training set
6. Which evaluation metric is often used to assess the performance of text classification models?
a. Mean Absolute Error (MAE) b. Area Under the Curve (AUC) c. F1-score d. R-squared
Answer: c. F1-score
7. What is the primary challenge in text classification when dealing with unstructured text data?
a. Limited computational resources b. Ambiguity and context dependence c. High availability of labeled
data d. Lack of text preprocessing techniques
Answer: b. Ambiguity and context dependence
8. In deep learning-based text classification, which type of neural network architecture is often used to
process sequences of words?
a. Convolutional Neural Network (CNN) b. Recurrent Neural Network (RNN) c. Decision Tree d. Support
Vector Machine (SVM)
Answer: b. Recurrent Neural Network (RNN)
9. What is the purpose of stop word removal in text classification?
a. To remove all text documents from the dataset b. To remove irrelevant or common words that don't
carry meaningful information c. To correct spelling and grammatical errors in text d. To add important
keywords to the text
Answer: b. To remove irrelevant or common words that don't carry meaningful information
10. Which application area does sentiment analysis fall under within the context of text classification?
a. Part-of-speech tagging b. Language detection c. Emotion recognition d. Assessing the sentiment or
polarity of text
Answer: d. Assessing the sentiment or polarity of text

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 11

Hyper-parameter Tunning

1. What are hyperparameters in machine learning?


a. Model parameters learned during training b. User-defined settings that control the learning process c.
Features extracted from the data d. Output labels for the training data
Answer: b. User-defined settings that control the learning process
2. What is the primary goal of hyperparameter tuning?
a. Maximizing the number of training samples b. Achieving the highest training accuracy c. Finding the
optimal set of hyperparameters for improved model performance d. Minimizing the number of features in
a model
Answer: c. Finding the optimal set of hyperparameters for improved model performance
3. Which technique involves systematically searching for the best hyperparameters using various
combinations and strategies?
a. Regularization b. Data augmentation c. Grid search d. Feature extraction
Answer: c. Grid search
4. What is the risk of overfitting during hyperparameter tuning?
a. Improved generalization on unseen data b. Increased training accuracy c. Decreased model complexity
d. Potentially worse performance on unseen data
Answer: d. Potentially worse performance on unseen data
5. What is cross-validation used for in hyperparameter tuning?
a. To evaluate the model's performance on the training data b. To create synthetic data for training c. To
assess the model's performance on unseen data d. To minimize the number of hyperparameters
Answer: c. To assess the model's performance on unseen data.
6. Which hyperparameter is related to the trade-off between bias and variance in a machine learning model?
a. Learning rate b. Number of epochs c. Batch size d. Regularization strength
Answer: d. Regularization strength
7. When using random search for hyperparameter tuning, how are hyperparameter values typically selected?
a. Randomly from a predefined range b. Sequentially from a grid of values c. Based on the performance of
the model d. According to alphabetical order
Answer: a. Randomly from a predefined range
8. Which technique involves gradually reducing the learning rate during training to reach the optimal
hyperparameters?
a. Early stopping b. Feature selection c. Learning rate annealing d. Principal Component Analysis (PCA)
Answer: c. Learning rate annealing
9. Which hyperparameter determines the number of times the entire training dataset is passed through a
model?
a. Batch size b. Learning rate c. Epochs d. Regularization strength
Answer: c. Epochs
10. What is the benefit of using a validation dataset during hyperparameter tuning?
a. It reduces the need for cross-validation. b. It speeds up the training process. c. It allows for direct model
evaluation without any hyperparameter changes. d. It helps in selecting the best hyperparameters while
avoiding overfitting.
Answer: d. It helps in selecting the best hyperparameters while avoiding overfitting.

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida
Deep Learning (ACSML0702) UNIT-II Convolution Neural Network 12

Emerging Neural Network Architecture

1. What is the purpose of the GPT (Generative Pre-trained Transformer) series of models in deep learning?
a. Object detection b. Image classification c. Natural language understanding and generation d. Speech
recognition
Answer: c. Natural language understanding and generation
2. What is the primary focus of Siamese Networks in deep learning?
a. Reinforcement learning b. Semi-supervised learning c. Face recognition and similarity learning d. Object
detection
Answer: c. Face recognition and similarity learning
3. In computer vision, what is a key advantage of the YOLO (You Only Look Once) architecture?
a. High interpretability of model predictions b. Real-time object detection with a single forward pass c.
Robustness to image noise d. Pixel-level image segmentation
Answer: b. Real-time object detection with a single forward pass
4. What is the primary application of CycleGAN (Cycle-Consistent Adversarial Networks) in deep learning?
a. Machine translation b. Style transfer and image-to-image translation c. Natural language understanding
d. Object detection
Answer: b. Style transfer and image-to-image translation
5. What does the term "capsule" refer to in Capsule Networks (CapsNets)?
a. A model architecture for speech recognition b. A building block for recurrent neural networks (RNNs) c.
A fundamental unit for handling hierarchical relationships in data d. A technique for image compression
Answer: c. A fundamental unit for handling hierarchical relationships in data
6. What is the primary advantage of Self-Attention mechanisms in neural networks?
a. Enhanced image resolution b. Improved language generation c. Handling long-range dependencies in
data d. Reduced computational complexity
Answer: c. Handling long-range dependencies in data
7. What is the key application area of GANs (Generative Adversarial Networks)?
a. Image classification b. Image generation and data synthesis c. Natural language processing d.
Reinforcement learning
Answer: b. Image generation and data synthesis
8. What is the primary focus of Federated Learning in deep learning?
a. Aggregating data from various sources into a single dataset b. Collaboratively training models across
decentralized devices while keeping data localized c. Feature extraction from large-scale datasets d.
Hyperparameter tuning techniques
Answer: b. Collaboratively training models across decentralized devices while keeping data
localized
9. In neural network architectures, what does the term "attention" refer to in models like BERT (Bidirectional
Encoder Representations from Transformers)?
a. Concentration of computing power b. Mechanism for focusing on relevant parts of input sequences c. A
type of recurrent layer d. The learning rate of the model
Answer: b. Mechanism for focusing on relevant parts of input sequences
10. What is the primary application of Spiking Neural Networks (SNNs) in emerging neural network
architectures?
a. Image classification b. Emulating human brain functions c. Audio processing d. Cryptography
Answer: b. Emulating human brain functions

Prepared by: Mr. Raj u, Assistant Professor, Department of Data Science, NIET, Greater Noida

You might also like