Convolutional Neural Networks Practice Quiz (24 questions)

All questions and answers

1. Which statements correctly describe machine learning and regression?

Machine learning performance improves through manually written rules, not experience with a task.
Regression models discover natural groupings without using target values in the dataset.
A regression dataset assigns each observation a categorical target rather than a numeric value.
Machine learning algorithms learn to solve problems from data and improve through experience.
Regression is a supervised task that predicts categorical labels from labeled observations.

Machine learning algorithms learn to solve problems from data and improve through experience.

Explanation

Machine learning algorithms learn from data and can improve their task performance through experience. Regression uses labeled data with numeric targets to predict continuous values; categorical targets describe classification, while finding natural groupings describes clustering.

2. A project predicts whether each email is spam and estimates house prices; which statements classify these tasks correctly?

Predicting spam status is classification because the target is a categorical class.
Estimating house prices is regression because the target is a continuous numeric value.
Predicting student test scores is a regression example involving numeric targets.
A model predicting spam probabilities is performing classification rather than regression.
Predicting letter grades is a regression example because grades are numeric targets.

Predicting spam status is classification because the target is a categorical class. · Estimating house prices is regression because the target is a continuous numeric value. · Predicting student test scores is a regression example involving numeric targets. · A model predicting spam probabilities is performing classification rather than regression.

Explanation

Spam status and spam-class probabilities are categorical prediction tasks, so they are classification. House prices and student test scores are regression examples because their targets are numeric; letter grades are categorical classes, not numeric regression targets.

3. A retailer groups similar products, segments customers, and studies social networks; which statements characterize these clustering applications?

Customer segmentation is an application of clustering in a dataset.
Social network analysis is listed as a clustering application.
Predicting email spam is a clustering example because spam forms a group.
Grouping similar e-commerce products is an example of clustering observations.
Predicting house prices is a clustering example because prices are continuous.

Customer segmentation is an application of clustering in a dataset. · Social network analysis is listed as a clustering application. · Grouping similar e-commerce products is an example of clustering observations.

Explanation

Grouping similar e-commerce items, customer segmentation, and social network analysis are all clustering examples. House-price prediction is regression, while predicting spam status is classification.

4. A model learns image representations directly from training data; which statements fit deep learning and its contrast with traditional feature extraction?

Deep learning relies on fixed feature extractors selected before training begins.
Deep learning composes nonlinear transformations to learn useful representations from data.
Traditional feature extractors are learned directly from data as part of deep learning.
Traditional feature extraction can require expert knowledge and automated hand-tuning.
Traditional machine learning uses trainable feature extractors learned from the data.

Deep learning composes nonlinear transformations to learn useful representations from data.

Explanation

Deep learning uses compositions of nonlinear transformations to learn useful features directly from data. Traditional approaches use fixed feature extractors, which can require expert knowledge and time-consuming hand-tuning, whereas deep learning uses trainable extractors learned from data.

5. Which statements describe feature extraction and the range of deep learning methods?

Traditional feature choices can be specific to the problem being addressed.
Deep learning methods may use either supervised or unsupervised learning.
Hand-tuning traditional feature extractors can be time-consuming.
Deep learning methods are restricted to supervised learning with labeled data.
Traditional feature extraction can require expert knowledge when choosing useful features.

Traditional feature choices can be specific to the problem being addressed. · Deep learning methods may use either supervised or unsupervised learning. · Hand-tuning traditional feature extractors can be time-consuming. · Traditional feature extraction can require expert knowledge when choosing useful features.

Explanation

Traditional feature extraction may require expertise, time-consuming hand-tuning, and problem-specific choices. Deep learning has supervised and unsupervised varieties, so it is not restricted to labeled-data training.

6. A recognition system receives raw pixels and produces category labels; which descriptions fit an end-to-end deep-learning approach?

It maps raw pixels directly to labels by composing simple nonlinear functions.
It replaces nonlinear components with a single fixed linear transformation of the pixels.
It first relies on a fixed feature extractor before a separate classifier assigns labels.
Its component functions can have learnable parameters adjusted during training.
It produces labels without composing intermediate functions into a larger mapping.

It maps raw pixels directly to labels by composing simple nonlinear functions. · Its component functions can have learnable parameters adjusted during training.

Explanation

An end-to-end recognition system maps raw pixels to labels through a composition of simple nonlinear functions, and those functions have learnable parameters. A fixed feature extractor followed by a separate classifier is not the described jointly learned approach; neither a single fixed linear map nor a system without composition matches the unit.

7. A single neuron receives inputs xix_i with learned weights WiW_i and bias bb; which descriptions match its computation and neural-network context?

It forms z=∑iWixi+bz = \sum_i W_i x_i + b before applying an activation function.
Its weights can begin with random values and be optimized on training data to reduce prediction error.
Neural networks can be used in supervised, semi-supervised, and unsupervised learning.
It applies the nonlinear activation first, then computes z=∑iWixi+bz = \sum_i W_i x_i + b.
A multilayer perceptron is a non-convolutional architecture formed by composing layers of simple models.

It forms $$z = \sum_i W_i x_i + b$$ before applying an activation function. · Its weights can begin with random values and be optimized on training data to reduce prediction error. · Neural networks can be used in supervised, semi-supervised, and unsupervised learning. · A multilayer perceptron is a non-convolutional architecture formed by composing layers of simple models.

Explanation

The neuron first calculates the weighted sum plus bias and then outputs a=f(z)a=f(z); network weights may be randomly initialized and optimized to reduce prediction error. An MLP is non-convolutional and built from composed layers, and neural networks span all three listed learning paradigms. Reversing the weighted-sum and activation order distorts the neuron’s computation.

8. A network’s sigmoid units operate near an extreme output, while some ReLU units receive negative inputs; which conclusions apply?

Leaky ReLU addresses the dead-neuron problem when its input is positive.
Leaky ReLU can help address dead ReLU units when their inputs are negative.
The sigmoid gradients can become close to zero as their outputs approach extreme values.
A numeric-valued network output corresponds to a regression task.
Tanh gradients become largest as outputs approach their extreme values.

Leaky ReLU can help address dead ReLU units when their inputs are negative. · The sigmoid gradients can become close to zero as their outputs approach extreme values. · A numeric-valued network output corresponds to a regression task.

Explanation

Sigmoid gradients saturate near extreme outputs, while a numeric-valued output indicates regression. Leaky ReLU helps with dead ReLU units for negative inputs. Tanh gradients also become close to zero near extreme outputs, rather than becoming largest, and the stated Leaky ReLU benefit concerns negative rather than positive inputs.

9. During training of a multilayer feedforward network, which sequence and role correctly describe output calculation, error handling, and backpropagation?

Backpropagation carries prediction error backward to calculate each weight’s contribution to that error.
The forward pass propagates error from the final layer toward the input to calculate weight contributions.
For each training example, the forward pass computes node outputs from left to right.
Backpropagation computes the network’s outputs from left to right before comparing them with targets.
The loss function measures the difference between the network’s final output and the training target.

Backpropagation carries prediction error backward to calculate each weight’s contribution to that error. · For each training example, the forward pass computes node outputs from left to right. · The loss function measures the difference between the network’s final output and the training target.

Explanation

Training computes outputs left to right, compares the final output with the target using a loss, and propagates error backward. Backpropagation determines each weight’s contribution to error and supports weight adjustment; the two false statements reverse the forward and backward pass roles.

10. A training update uses learning rate η=0.1\eta=0.1, current weight wi=2w_i=2, and error derivative ∂E∂wi=3\frac{\partial E}{\partial w_i}=3; which statements are correct?

The update rule gives 2+0.1×3=2.32+0.1\times3=2.3 for these values.
Training can repeat adjustments on randomly selected instances thousands or millions of times.
A positive error derivative makes this gradient-descent update decrease the weight.
Gradient descent gives the updated weight as 2−0.1×3=1.72-0.1\times3=1.7.
Training repeats adjustments using randomly selected test instances thousands of times.

Training can repeat adjustments on randomly selected instances thousands or millions of times. · A positive error derivative makes this gradient-descent update decrease the weight. · Gradient descent gives the updated weight as $$2-0.1\times3=1.7$$.

Explanation

The update subtracts the learning-rate-scaled derivative, so winew=2−0.1×3=1.7w_i^{\mathrm{new}}=2-0.1\times3=1.7; a positive derivative therefore decreases the weight. Training may repeat adjustments on randomly selected training instances thousands or millions of times, not test instances; adding the derivative uses the wrong update direction.

11. Which capabilities are associated with backpropagation as a training procedure for multilayer feedforward neural networks?

It can learn solutions to problems that are linearly inseparable.
It enables training of multilayer feedforward neural networks.
It can theoretically perform any input-output mapping.
It adjusts weights using gradient descent before propagating error backward.
It is restricted to learning mappings whose classes are linearly separable.

It can learn solutions to problems that are linearly inseparable. · It enables training of multilayer feedforward neural networks. · It can theoretically perform any input-output mapping.

Explanation

Backpropagation is a training procedure for multilayer feedforward networks, and it can theoretically perform any input-output mapping. It can also learn linearly inseparable problems. During training, error is propagated backward before weights are adjusted using gradient descent; restricting learning to linearly separable tasks contradicts its stated capabilities.

12. A technician compares three image filters: a left-shifting kernel and the two Sobel orientations. Which statements correctly identify their effects?

The kernel with its 1 in the middle-left position shifts the image left by one pixel.
The Sobel kernel with vertical sign changes detects horizontal edges by absolute response.
The vertical-edge Sobel kernel has the horizontal-edge kernel's orientation.
The left-shifting kernel moves the image one pixel to the right.
The Sobel kernel with horizontal sign changes detects vertical edges by absolute response.

The kernel with its 1 in the middle-left position shifts the image left by one pixel. · The Sobel kernel with vertical sign changes detects horizontal edges by absolute response. · The Sobel kernel with horizontal sign changes detects vertical edges by absolute response.

Explanation

The kernel with its 1 in the middle-left position shifts the image left by one pixel. The Sobel kernel with horizontal sign changes detects vertical edges, while the vertically oriented Sobel kernel detects horizontal edges; the shift direction and the Sobel orientations in the false statements are reversed.

13. A HOG procedure calculates gradient measurements and then selects points of interest. Which statements describe the procedure?

The procedure calculates gradient norm but omits gradient direction.
The HOG approach described for this procedure is attributed to Dalal in 2005.
The procedure thresholds gradient direction to detect points of interest.
The procedure calculates gradient norm and gradient direction before identifying points of interest.
A threshold applied to gradient norm is used to detect points of interest.

The HOG approach described for this procedure is attributed to Dalal in 2005. · The procedure calculates gradient norm and gradient direction before identifying points of interest. · A threshold applied to gradient norm is used to detect points of interest.

Explanation

HOG calculates both gradient norm and gradient direction, and a threshold on the norm detects points of interest. The approach is attributed to Dalal (2005); the threshold is on norm, not direction, and direction is also calculated.

14. A CNN is classifying a face from raw pixels; which descriptions of its feature learning and architecture are accurate?

An early layer may detect edges, which a later layer combines into simple shapes.
A typical CNN architecture omits pooling or subsampling layers in favor of fully connected layers.
Shared weights allow the same feature to be detected at different positions while reducing free parameters.
Higher-level facial shapes may be detected before simple shapes are formed from edges.
For an RGB input, a separate k×kk\times k filter is applied to each of its three channels.

An early layer may detect edges, which a later layer combines into simple shapes. · Shared weights allow the same feature to be detected at different positions while reducing free parameters. · For an RGB input, a separate $$k\times k$$ filter is applied to each of its three channels.

Explanation

A CNN may progress from edges to simple shapes and then to higher-level features such as facial shapes. Local connections and shared weights let features be recognized at different positions and reduce free parameters. RGB inputs use separate filters for each channel; the architecture can include pooling or subsampling, and the described hierarchy places simple shapes before higher-level facial features.

15. When a convolution filter reaches an image boundary, which statements describe the operation correctly?

Padding leaves the original image unchanged, and filter size alone determines the feature-map size.
Increasing stride beyond 1 reduces the spatial size of the resulting feature map.
Padding can enlarge the image, for example by adding zeros, so border locations can be convolved.
The learned filters correspond to convolution-layer weights, whose activations form a two-dimensional feature map.
The activations are the learned filters, while the layer weights are the two-dimensional response map.

Increasing stride beyond 1 reduces the spatial size of the resulting feature map. · Padding can enlarge the image, for example by adding zeros, so border locations can be convolved. · The learned filters correspond to convolution-layer weights, whose activations form a two-dimensional feature map.

Explanation

Padding extends the input, often with zeros, to address border locations; feature-map size depends on padding and filter size. Learned filters are the convolution-layer weights, and their responses form a two-dimensional feature map. A stride greater than 1 reduces feature-map size; the false alternatives reverse the roles of weights and activations or misstate padding's effect.

16. A CNN applies 2 × 2 pooling with stride 2 to convolutional feature maps; which outcomes are expected?

Average pooling is another commonly used pooling type.
The illustrated pooling setup uses 2 × 2 filters with stride 2.
Max pooling is one of the common pooling types.
Pooling reduces the spatial resolution of each convolutional feature map.
Pooling increases feature-map spatial resolution to improve shift invariance.

Average pooling is another commonly used pooling type. · The illustrated pooling setup uses 2 × 2 filters with stride 2. · Max pooling is one of the common pooling types. · Pooling reduces the spatial resolution of each convolutional feature map.

Explanation

Pooling downsamples convolutional outputs by reducing each feature map's spatial resolution, which yields some shift and distortion invariance. Max and average pooling are common types, and the illustrated setup uses 2 × 2 filters with stride 2. The final statement reverses the direction of the spatial-resolution change.

17. A model fits noisy training examples closely but performs poorly on new test examples; which statements apply?

During training, dropout randomly removes units together with their connections.
Overfitting can arise when a learned hypothesis fits noise or outliers in its training data.
Dropout addresses overfitting chiefly in convolutional layers rather than dense layers.
Dropout removes units during testing to prevent complex co-adaptations.
This gap between close training fit and poor test generalization describes overfitting.

During training, dropout randomly removes units together with their connections. · Overfitting can arise when a learned hypothesis fits noise or outliers in its training data. · This gap between close training fit and poor test generalization describes overfitting.

Explanation

Overfitting is a strong fit to training data, including noise or outliers, coupled with poor generalization to new examples. Dropout randomly removes units and connections during training to promote robust features and reduce complex co-adaptations; it is especially helpful in dense layers. It is not applied at test time.

18. A CNN's final dense layer receives extracted features from earlier layers. Which statements describe its structure and role?

Deeper replication of convolution and pooling layers was reported to reduce performance.
The dense layer has a two-dimensional structure, with nodes arranged as image pixels.
The dense layer uses high-level features from earlier layers to classify the input.
Each dense-layer node connects to every input node and computes a weighted sum.
The dense layer classifies inputs using low-level pixels instead of extracted features.

The dense layer uses high-level features from earlier layers to classify the input. · Each dense-layer node connects to every input node and computes a weighted sum.

Explanation

A fully connected layer is one-dimensional, and each node connects to all input nodes to compute a weighted sum. It uses high-level features from preceding layers for classification. Experiments reported better results with deeper replication of convolution and pooling layers, while replacing extracted features with low-level pixels misstates the dense layer's role.

19. A team reuses a saved network trained on a large image dataset for a new vision task. Which statements correctly distinguish transfer-learning choices?

A saved network's learned spatial hierarchy can serve as a generic model across vision problems.
Higher convolutional layers tend to encode generic edges, while earlier layers encode object concepts.
Feature extraction feeds reused representations into a new classifier trained from scratch.
Earlier convolutional layers tend to capture edges, colors, and textures more generically.
Fine-tuning freezes selected top layers while training only the newly added classifier.

A saved network's learned spatial hierarchy can serve as a generic model across vision problems. · Feature extraction feeds reused representations into a new classifier trained from scratch. · Earlier convolutional layers tend to capture edges, colors, and textures more generically.

Explanation

A pre-trained network is a saved network whose learned hierarchy can be useful across computer-vision problems. Feature extraction uses prior representations with a newly trained classifier, whereas fine-tuning unfreezes selected top layers and trains them jointly with the new model part. Earlier layers capture generic local features; higher layers capture more abstract concepts, not the reverse.

20. A new image dataset differs greatly from the dataset used to pre-train an ImageNet CNN. Which strategy is supported?

A model pre-trained on ImageNet cannot be further trained on new data.
Using the first few layers for feature extraction may be preferable to using the entire convolutional base.
Using the entire convolutional base is preferred when the new dataset differs greatly.
Fine-tuning means using the pre-trained CNN as fixed features without updating weights.
ImageNet pre-training provides no option to reuse the network as fixed features.

Using the first few layers for feature extraction may be preferable to using the entire convolutional base.

Explanation

When the new dataset differs greatly, using only the first few layers for feature extraction may be preferable because early representations are more generic. The entire base is not the preferred strategy in that case. ImageNet-pre-trained networks can be used as fixed features or initializations, and continued training on new data can fine-tune their weights.

21. A developer processes a color image with input_shape (128, 128, 3) and kernel_size (3, 3). Which statements apply to this Conv2D setup?

Conv2D is mainly used for video sequences because its kernel spans time.
The input dimensions represent width, height, and kernel depth, in that order.
The input dimensions represent image height, width, and depth, in that order.
The kernel dimensions specify height and width, while kernel depth matches image depth.
The kernel slides across three spatial dimensions when processing the image.

The input dimensions represent image height, width, and depth, in that order. · The kernel dimensions specify height and width, while kernel depth matches image depth.

Explanation

For Conv2D, the input shape gives height, width, and depth; the two kernel-size values give kernel height and width, and kernel depth matches image depth. Conv2D kernels slide along two dimensions, while three-dimensional sliding and video applications are associated with Conv3D. The stated input ordering is not width, height, and kernel depth.

22. A researcher needs to analyze MRI volumes and sequences of video frames with a 3D CNN. Which statements are correct?

A kernel size of (3, 3, 3) specifies height, width, and color-channel count.
The shape (128, 128, 128, 3) has four dimensions for a Conv3D input.
The fourth input dimension in this example represents kernel depth, not color channels.
Conv3D activation maps discard temporal and volumetric context during analysis.
Conv3D activation maps retain temporal or volumetric context for such analyses.

The shape (128, 128, 128, 3) has four dimensions for a Conv3D input. · Conv3D activation maps retain temporal or volumetric context for such analyses.

Explanation

Conv3D activation maps preserve temporal or volumetric context, supporting video-frame analysis and medical imaging such as MRI and CT. The given input shape has four dimensions; the three kernel-size values specify height, width, and depth, while the fourth dimension matches the color channel. Thus the kernel-size tuple does not specify color-channel count, and the fourth input dimension is not kernel depth.

23. A benchmark summary reports ImageNet's scale and human top-5 error. Which statements accurately describe the dataset and the listed CNN architectures?

AlexNet was trained for six days on two Nvidia Geforce GTX 780 GPUs.
VGG-16 was presented in 2014 by Karen Simonyan and Andrew Zisserman.
Human top-5 error on ImageNet is approximately 1.2%.
LeNet5 reached 99.2% accuracy on MNIST handwritten-character recognition.
ImageNet contains 1,000 classes and about 1.2 million labeled photographs.

VGG-16 was presented in 2014 by Karen Simonyan and Andrew Zisserman. · LeNet5 reached 99.2% accuracy on MNIST handwritten-character recognition. · ImageNet contains 1,000 classes and about 1.2 million labeled photographs.

Explanation

ImageNet is described as having 1,000 classes and 1.2 million labeled photos, while human top-5 error is about 5%, not 1.2%. LeNet5 achieved 99.2% accuracy on MNIST, AlexNet used two GTX 580 GPUs, and VGG-16 was presented by Simonyan and Zisserman in 2014.

24. Which propositions correctly match these CNN architectures with their cited papers and publication years?

Inception-v3 is attributed to Christian Szegedy and colleagues in a 2015 paper on rethinking Inception.
ResNet-50 is attributed to Kaiming and colleagues in a 2015 paper on deep residual learning.
Inception-v1, also called GoogLeNet, is attributed to Christian Szegedy and colleagues in 2014.
Inception-v3 is attributed to Christian Szegedy and colleagues in the 2014 paper “Going Deeper with Convolutions.”
VGG-16 is attributed to Christian Szegedy and colleagues in the 2014 paper “Very Deep Convolutional Networks for Large-Scale Image Recognition.”

Inception-v3 is attributed to Christian Szegedy and colleagues in a 2015 paper on rethinking Inception. · ResNet-50 is attributed to Kaiming and colleagues in a 2015 paper on deep residual learning. · Inception-v1, also called GoogLeNet, is attributed to Christian Szegedy and colleagues in 2014.

Explanation

Inception-v1 (GoogLeNet) is attributed to Christian Szegedy and colleagues’ 2014 paper “Going Deeper with Convolutions,” while Inception-v3 is attributed to their 2015 paper on rethinking the Inception architecture. ResNet-50 is attributed to Kaiming and colleagues’ 2015 paper on deep residual learning. The 2014 “Going Deeper with Convolutions” paper is not the cited paper for Inception-v3, and the 2014 “Very Deep Convolutional Networks for Large-Scale Image Recognition” paper is associated with VGG-16, not Christian Szegedy and colleagues.

Study with flashcards

Lock in the answers with 82 flashcards on Convolutional Neural Networks.

What does machine learning algorithms improve through experience?

Their performance on a task.

What kind of data does regression use to predict continuous numeric variables?

Labeled data with a numeric target for each observation.

What kind of variables does classification predict?

Categorical variables, such as classes or class probabilities.

See the flashcards →

Read the study guide

Go over the full study guide for Convolutional Neural Networks.

Read the study guide →

Related classes

Make your own practice quizzes

Upload your class notes and Revizly writes a practice quiz with explained answers in about a minute.

Quiz maker