Convolutional Neural Networks Flashcards (82 cards)

All cards

1Question

What does machine learning algorithms improve through experience?

Answer

Their performance on a task.

2Question

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

Answer

Labeled data with a numeric target for each observation.

3Question

What kind of variables does classification predict?

Answer

Categorical variables, such as classes or class probabilities.

4Question

What does clustering find in an unsupervised learning task?

Answer

Natural groupings based on the dataset’s inherent structure.

5Question

How does deep learning learn useful representations directly from data?

Answer

By composing nonlinear transformations of the data.

6Question

How do deep learning feature extractors differ from traditional machine-learning feature extractors?

Answer

Deep learning feature extractors are trainable and learned from data.

7Question

What does an end-to-end recognition system map raw pixels to?

Answer

Labels.

8Question

How does an end-to-end recognition system map raw pixels to labels?

Answer

By composing simple nonlinear functions with learnable parameters.

9Question

What is a neural network?

Answer

A set of connected neurons with randomly initialized weights and nonlinear activation functions, optimized on training data to minimize prediction error.

10Question

What formula gives a single-layer neuron's pre-activation value zz?

Answer

z=∑iWixi+bz = \sum_i W_i x_i + b

11Question

What formula gives a single-layer neuron's output aa?

Answer

a=f(z)a = f(z)

12Question

What sequence does neural-network training follow for each training example?

Answer

Initialize weights, compute outputs left to right, measure loss, propagate error backward, and adjust weights with gradient descent.

13Question

What does backpropagation calculate in a multilayer feedforward neural network?

Answer

Each weight’s contribution to prediction error.

14Question

What is the gradient descent update formula for weight wiw_i?

Answer

winew=wiold−η∂E∂wiw_i^{\mathrm{new}} = w_i^{\mathrm{old}} - \eta \frac{\partial E}{\partial w_i}

15Question

What does η\eta represent in the gradient descent update formula?

Answer

The learning rate.

16Question

What does a convolutional neural network extract from images?

Answer

Topological properties.

17Question

How does image filtering apply a convolution kernel?

Answer

It slides a window across the image and approximates the result from neighboring pixels.

18Question

Which Sobel kernels detect orientations in an image?

Answer

Horizontal and vertical gradient kernels.

19Question

Which reference is associated with the HOG method?

Answer

Dalal, 2005.

20Question

What two quantities are calculated in a HOG descriptor?

Answer

Gradient norm and gradient direction.

21Question

How does HOG detect points of interest?

Answer

It applies a threshold to the gradient norm.

22Question

Why do CNNs reuse the same kernel across image locations?

Answer

Features such as edges can occur anywhere in the image.

23Question

What can a CNN learn from raw pixels in an early image-classification layer?

Answer

Edges.

24Question

What two properties let CNNs detect the same feature at different image positions?

Answer

Local connections and shared weights.

25Question

What does pooling do to each convolutional feature map’s spatial resolution?

Answer

It reduces the spatial resolution.

26Question

Which invariances can pooling provide?

Answer

Some shift and distortion invariance.

27Question

What is overfitting?

Answer

Fitting training data, including noise or outliers, while failing to generalize to new test examples.

28Question

How does dropout encourage robust features during training?

Answer

It randomly removes units and their connections, reducing complex co-adaptations and overfitting, especially in dense layers.

29Question

What is a pre-trained network?

Answer

A saved network previously trained on a large dataset.

30Question

How does feature extraction differ from fine-tuning?

Answer

Feature extraction trains a new classifier from scratch; fine-tuning jointly trains selected unfrozen top layers with the new model part.

31Question

What features do earlier convolutional layers extract?

Answer

Local, generic features such as edges, colors, and textures.

32Question

Along how many dimensions does a Conv2D kernel slide?

Answer

Two dimensions.

33Question

What kind of data is Conv2D generally used on?

Answer

Image data.

34Question

Along how many dimensions do Conv3D kernels slide?

Answer

Three dimensions.

35Question

What data types are Conv3D mainly used with?

Answer

3D images and video sequences.

36Question

How many classes and labeled photos does ImageNet contain?

Answer

1,000 classes and 1.2 million labeled photos.

37Question

What is the approximate human top-5 error on ImageNet?

Answer

Approximately 5%.

38Question

Which three outcomes are examples of regression?

Answer

Real-estate prices, stock price movements, and student test scores.

39Question

Which three outcomes are examples of classification?

Answer

Email spam, financial fraud, and student letter grades.

40Question

Which three applications are examples of clustering?

Answer

Customer segmentation, grouping similar e-commerce items, and social network analysis.

41Question

What can traditional feature extraction require?

Answer

Expert knowledge, time-consuming hand-tuning, and problem-specific choices.

42Question

What two learning settings can deep learning methods use?

Answer

Supervised or unsupervised learning.

43Question

What is a multilayer perceptron?

Answer

A non-convolutional neural-network architecture built by composing layers of simple models.

44Question

Which machine-learning paradigms can neural networks address?

Answer

Supervised, semi-supervised, and unsupervised learning.

45Question

How does a neural network map input xx to output yy?

Answer

y=fθ(x)y = f_\theta(x)

46Question

What happens to sigmoid and tanh gradients near extreme output values?

Answer

They saturate and become close to zero.

47Question

What problem does Leaky ReLU help address when its input is negative?

Answer

The dead-neuron problem of ReLU.

48Question

Across what kinds of instances does training repeat weight adjustments?

Answer

Randomly selected training instances.

49Question

How many times may training repeat weight adjustments?

Answer

Thousands or possibly millions of times.

50Question

What input-output mappings can backpropagation theoretically perform?

Answer

Any input-output mapping.

51Question

What class of problems can backpropagation learn?

Answer

Linearly inseparable problems.

52Question

How are convolutional neural networks trained?

Answer

With a version of backpropagation.

53Question

How does convolution process a pixel?

Answer

Using information from neighboring pixels.

54Question

What is multiplication equivalent to in the spatial domain?

Answer

Convolution.

55Question

What are typical convolution-kernel dimensions?

Answer

3×3, 5×5, or 7×7.

56Question

What can describe detected orientations using a histogram?

Answer

A histogram of directions.

57Question

What does CNN topology restrict connections to?

Answer

Local image regions.

58Question

What is a feature map?

Answer

The output produced by convolving one feature.

59Question

How are filters applied to the channels of an RGB input?

Answer

A separate k×kk\times k filter is applied to each channel.

60Question

Which layer types does a CNN architecture combine?

Answer

Feature-extraction, convolution, pooling or subsampling, and fully connected layers.

61Question

Which pooling types are common, and what filter size and stride are illustrated?

Answer

Max and average pooling use 2×2 filters with stride 2.

62Question

When is dropout applied?

Answer

During training, not testing.

63Question

How can data augmentation reduce overfitting?

Answer

By applying small image transformations or adding noise.

64Question

Which image changes can data augmentation use to reduce overfitting?

Answer

Small rotations, translations, scaling, flipping, brightness changes, and noise.

65Question

How does each node in a fully connected layer combine its inputs?

Answer

It computes a weighted sum of all input nodes.

66Question

What does a fully connected layer use to classify an input pattern?

Answer

High-level features from previous layers.

67Question

What did course experiments report about deeper CNNs?

Answer

Deeper replication of convolution and pooling layers produced better results.

68Question

When a new dataset differs greatly from the original training data, what may be preferable for feature extraction?

Answer

Using only the first few layers.

69Question

How can an ImageNet-pretrained CNN be adapted to new data?

Answer

Use it as an initialization, then fine-tune its weights on the new data.

70Question

What do the three values in Conv2D input_shape (128, 128, 3) represent?

Answer

Height, width, and depth.

71Question

What does Conv2D kernel_size (3, 3) specify?

Answer

Kernel height and width.

72Question

What context do Conv3D activation maps preserve?

Answer

Temporal or volumetric context.

73Question

In Conv3D, what does kernel_size (3, 3, 3) specify?

Answer

Kernel height, width, and depth.

74Question

What accuracy did LeNet-5 achieve on MNIST handwritten-character recognition?

Answer

99.2%.

75Question

How long was AlexNet trained, and on how many Nvidia GeForce GTX 580 GPUs?

Answer

Six days on two GPUs.

76Question

What approach did VGG-16 use to improve CNN efficiency?

Answer

Stacking more layers.

77Question

Which work attributes Inception-v1 (GoogLeNet) to Christian Szegedy et al.?

Answer

“Going Deeper with Convolutions” (2014).

78Question

Which work attributes Inception-v3 to Christian Szegedy et al.?

Answer

“Rethinking the Inception Architecture for Computer Vision” (2015).

79Question

Which work attributes ResNet-50 to Kaiming et al.?

Answer

“Deep Residual Learning for Image Recognition” (2015).

80Question

Which authors are cited for the 1995 account of convolutional neural networks trained by backpropagation?

Answer

Yann LeCun and Yoshua Bengio.

81Question

What response results from applying kernel [2, 3, 1] to sequence [1, 2, 3, 4, 5, 6, 7]?

Answer

[11, 17, 23, 29, 35].

82Question

How many weight parameters do 100 shared 10×10 filters require for a single-channel image?

Answer

10,000, excluding biases.

Test yourself with the quiz

Answer 24 questions on Convolutional Neural Networks.

1. Which statements correctly describe machine learning and regression?

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

Take the quiz →

Read the study guide

Go over the full study guide for Convolutional Neural Networks.

Read the study guide →

Related classes

Make your own flashcards

Upload your class notes and Revizly writes your flashcards in about a minute.

Flashcard maker