wayground logo

Free Printable Worksheets

Font size

S
M
L
XL
Worksheets

Linear Algebra and Machine Learning Quiz

Total questions: 111

Worksheet time: 56mins

Name
Class
Date
1.

What does a scalar represent in linear algebra?

a)

A 1D array

b)

A single numerical value

c)

A matrix with one row

d)

A function

2.

Which of the following best describes a matrix?

a)

A list of vectors

b)

A 2D array of numbers

c)

A collection of scalars only

d)

A line in space

3.

If a matrix has dimensions 3×4, it has:

a)

3 columns and 4 rows

b)

4 rows and 3 columns

c)

3 rows and 4 columns

d)

12 columns

4.

In deep learning, vectors and matrices are often used to represent:

a)

Only images

b)

Mathematical constants

c)

Data and model parameters

d)

Programming functions

5.

Which operation is not valid between two matrices?

a)

Addition of same-shaped matrices

b)

Multiplication with proper dimensions

c)

Transposing a matrix

d)

Dividing one matrix by another

6.

The dot product of two vectors results in a:

a)

Matrix

b)

Vector

c)

Scalar

d)

Tensor

7.

What is the transpose of a matrix?

a)

Flipping matrix values

b)

Swapping rows with columns

c)

Taking the inverse

d)

Reducing dimensions

8.

What is required for two matrices A and B to be multiplied (i.e., AB)?

a)

Number of columns in A = number of columns in B

b)

Number of rows in A = number of rows in B

c)

Number of columns in A = number of rows in B

d)

Both must be square matrices

9.

In neural networks, the weight matrices are often:

a)

Used for storing losses

b)

Randomly initialized

c)

Always identity matrices

d)

Deterministically fixed

10.

What is the identity matrix?

a)

A matrix full of zeros

b)

A matrix with all ones

c)

A square matrix with 1s on the diagonal and 0s elsewhere

d)

A matrix with mirrored values

11.

What is classification in machine learning?

a)

Finding clusters in unlabeled data

b)

Mapping input to a discrete output

c)

Finding correlations between variables

d)

Generating continuous outputs

12.

In supervised learning, what is provided with the input data?

a)

Data clusters

b)

Noise

c)

Labels

d)

Hidden layers

13.

Which component defines the possible outputs a learning algorithm can predict?

a)

Hypothesis space

b)

Output space

c)

Feature vector

d)

Cost function

14.

The hypothesis space is:

a)

All possible datasets

b)

All outputs a model can predict

c)

All candidate functions a model can learn

d)

All possible errors a model can make

15.

Which component of learning evaluates prediction error?

a)

Optimizer

b)

Loss function

c)

Gradient

d)

Feature vector

16.

What is the role of the learning algorithm?

a)

To define the loss function

b)

To find the best hypothesis in the hypothesis space

c)

To generate the training data

d)

To normalize inputs

17.

What does the input space represent?

a)

All possible labels

b)

Set of possible inputs to the model

c)

Set of outputs the model can produce

d)

Distribution of the loss function

18.

The feature vector refers to:

a)

The label assigned to a sample

b)

The space of learned weights

c)

A numerical representation of input data

d)

A type of classifier

19.

What is Empirical Risk Minimization (ERM)?

a)

Estimating risk using test data

b)

Calculating the maximum loss

c)

Minimizing loss on training data

d)

Ignoring loss on training data

20.

True risk is also known as:

a)

Empirical error

b)

Test loss

c)

Expected risk

d)

Hypothesis error

21.

What do we aim to minimize in machine learning?

a)

True labels

b)

Training speed

c)

Expected risk

d)

Model size

22.

Generalization is the model's ability to:

a)

Memorize training data

b)

Perform well on training data only

c)

Perform well on unseen data

d)

Store all input-output pairs

23.

Overfitting happens when:

a)

The model generalizes well

b)

The training error is high

c)

The test error is lower than training error

d)

The model memorizes training data

24.

Underfitting occurs when:

a)

The model is too complex

b)

The model fails to learn from training data

c)

The model performs well on test data

d)

The hypothesis space is large

25.

What does the bias in bias-variance tradeoff refer to?

a)

Sensitivity to noise

b)

Error due to incorrect assumptions in the model

c)

Variation due to data sampling

d)

Errors in the test set

26.

Variance in a model refers to:

a)

Noise in the data

b)

Sensitivity to small changes in training data

c)

Assumptions made by the model

d)

Error on the training set

27.

Which of the following can reduce variance?

a)

Increasing model complexity

b)

Decreasing training size

c)

Using simpler models

d)

Using fewer epochs

28.

Which technique can help reduce overfitting?

a)

Increasing training error

b)

Using a large hypothesis space

c)

Regularization

d)

Increasing model depth

29.

High bias and low variance usually lead to:

a)

Overfitting

b)

Underfitting

c)

Perfect generalization

d)

No learning

30.

A model with low bias and high variance tends to:

a)

Underfit

b)

Ignore the data

c)

Overfit

d)

Generalize well

31.

Which of these is a common loss function for classification tasks?

a)

Mean squared error

b)

Cross-entropy

c)

Hinge loss

d)

Both B and C

32.

Cross-validation helps in:

a)

Reducing test error

b)

Estimating model performance on unseen data

c)

Increasing training accuracy

d)

Normalizing input data

33.

What happens if a model is too complex for the data?

a)

It generalizes better

b)

It underfits

c)

It overfits

d)

It becomes regularized

34.

Which component selects the best model based on performance?

a)

Feature extractor

b)

Output function

c)

Learning algorithm

d)

Cost function

35.

The tradeoff between bias and variance is crucial for:

a)

Training speed

b)

Memory efficiency

c)

Model generalization

d)

Choosing optimizers

36.

What was the main topic discussed in the previous week's lecture before introducing deep learning?

a)

Decision Trees

b)

Logistic Regression for Classification

c)

Support Vector Machines

d)

Neural Networks

37.

What framework was used to implement Logistic Regression in the previous week's lecture?

a)

Empirical Risk Minimization (ERM)

b)

Gradient Boosting

c)

K-Means Clustering

d)

Principal Component Analysis (PCA)

38.

What application was built using Softmax Regression in the previous week's lecture?

a)

Image Segmentation

b)

MNIST Handwritten Digit Classification

c)

Speech Recognition

d)

Sentiment Analysis

39.

What limitation of Logistic Regression was observed when dealing with non-linearly separable data?

a)

It cannot handle high-dimensional data

b)

It is only effective for linearly separable data

c)

It requires manual feature selection

d)

It cannot perform multiclass classification

40.

Why is feature extraction a challenge in traditional machine learning for unstructured datasets like images?

a)

It reduces the dataset

b)

It results in high-dimensional data, e.g., 784 columns for 28x28 images

c)

It is computationally inexpensive

d)

It automatically adapts to image variations

41.

What is a key difference between machine learning and deep learning according to the lecture?

a)

Machine learning requires no human intervention

b)

Deep learning can learn features directly from data

c)

Machine learning is only used for image data

d)

Deep learning requires manual feature engineering

42.

Why is manual feature extraction considered a limitation in traditional machine learning?

a)

It is always computationally inexpensive

b)

It requires expertise and may not generalize well

c)

It automatically adapts to different datasets

d)

It eliminates the need for dimensionality reduction

43.

What is a major breakthrough of deep learning mentioned in the lecture?

a)

It has replaced all traditional machine learning algorithms

b)

It has established state-of-the-art results in various pattern recognition tasks

c)

It requires no data preprocessing

d)

It only works with structured datasets

44.

What is a common element in modern deep learning breakthroughs?

a)

Use of decision trees

b)

Neural networks

c)

Manual feature extraction

d)

Linear regression model

45.

What is a common element in modern deep learning breakthroughs?

a)

Use of decision trees

b)

Neural networks

c)

Manual feature extraction

d)

Linear regression models

46.

What is an example application of neural networks mentioned in the lecture?

a)

Auto Caption Generator

b)

K-Means Clustering

c)

Principal Component Analysis

d)

Logistic Regression

47.

What inspired the development of neural networks according to the lecture?

a)

Human ability of thinking

b)

Quantum computing

c)

Statistical modeling

d)

Graph theory

48.

What is the theory of Associationism, as described in the lecture?

a)

Learning is a process of forming associations between temporally related phenomena

b)

Learning is purely based on genetic factors

c)

Learning requires no external stimuli

d)

Learning is a random process

49.

Who proposed the idea that the brain is a network of interconnected neurons in 1873?

a)

Alan Turing

b)

Alexander Bain

c)

Warren McCulloch

d)

Frank Rosenblatt

50.

What is a key difference between connectionist machines and Von Neumann machines?

a)

Connectionist machines use a single processing unit

b)

Connectionist machines store programs in connections between units

c)

Von Neumann machines are non-linear

d)

Connectionist machines are only used for numerical calculations

51.

What are the major elements of a biological neuron mentioned in the lecture?

a)

Dendrite, Synapse, Soma, Axon

b)

Input, Output, Weight, Bias

c)

Node, Edge, Vertex, Graph

d)

Processor, Memory, Bus, Cache

52.

Who proposed the first computational model of a neuron in 1943?

a)

Frank Rosenblatt

b)

Alan Turing

c)

Warren McCulloch and Walter Pitts

d)

Alexander Bain

53.

What type of logic does the McCulloch-Pitts (MCP) neuron model?

a)

Fuzzy logic

b)

Boolean logic

c)

Probabilistic logic

d)

Quantum logic

54.

What happens when an inhibitory synapse is active in an MCP neuron?

a)

It amplifies the excitatory inputs

b)

It prevents the neuron from firing regardless of other inputs

c)

It increases the threshold value

d)

It has no effect on the neuron

55.

What is the output of an MCP neuron when the sum of excitatory inputs is below the threshold?

a)

1

b)

0

c)

-1

d)

Undefined

56.

Which Boolean function cannot be implemented by a single MCP neuron due to its non-linear separability?

a)

AND

b)

OR

c)

XOR

d)

NOT

57.

Who introduced the Perceptron model in 1958?

a)

Warren McCulloch

b)

Frank Rosenblatt

c)

Alan Turing

d)

Marvin Minsky

58.

What is a key feature of the Perceptron compared to the MCP neuron?

a)

It uses fixed weights

b)

It incorporates learnable weights and an adaptive learning process

c)

It only processes binary inputs

d)

It does not use a threshold function

59.

What is the role of the bias term in the Perceptron model?

a)

It scales the input values

b)

It adjusts the decision boundary

c)

It reduces the dimensionality of inputs

d)

It replaces the activation function

60.

Why is the step function unsuitable for real-valued inputs in a Perceptron?

a)

It is non-differentiable and produces binary outputs

b)

It is computationally expensive

c)

It cannot handle binary classification

d)

It requires manual tuning

61.

What is the purpose of the sigmoid activation function in sigmoid neurons?

a)

To introduce non-linearity and map inputs to a range between 0 and 1

b)

To perform linear transformations

c)

To reduce the dimensionality of inputs

d)

To eliminate the need for weights

62.

What is a key limitation of the perceptron discussed in the lecture?

a)

It can handle non-linearly separable data effectively

b)

It only works well for linearly separable data

c)

It uses a sigmoid activation function

d)

It requires no training data

63.

Why is the unit step function not ideal for modern neural network problems?

a)

It is computationally expensive

b)

It is non-differentiable and produces binary outputs

c)

It allows for complex decision boundaries

d)

It supports gradient-based optimization

64.

What problem was highlighted in the tutorial regarding the perceptron's inability to solve certain Boolean functions?

a)

AND function

b)

OR function

c)

XOR function

d)

NOT function

65.

What is a major drawback of the perceptron learning algorithm when applied to multi-layer networks?

a)

It supports backpropagation

b)

It cannot propagate errors to hidden layers

c)

It is computationally efficient for deep networks

d)

It handles non-linear data effectively

66.

What does arranging perceptrons in layers allow, according to the lecture?

a)

Learning linearly separable functions only

b)

Learning non-linearly separable functions like XOR

c)

Reducing the number of neurons required

d)

Eliminating the need for activation functions

67.

What is a key property of a good activation function?

a)

Linear output to simplify computations

b)

Non-differentiability to avoid optimization

c)

Non-linearity to learn complex patterns

d)

Binary output for all inputs

68.

What is the output range of the sigmoid activation function?

a)

-1 to 1

b)

0 to 1

c)

0 to infinity

d)

-infinity to infinity

69.

Why is the sigmoid activation function useful for binary classification?

a)

It produces discrete outputs

b)

It maps inputs to probabilities between 0 and 1

c)

It is non-differentiable

d)

It avoids non-linear decision boundaries

70.

What activation function is typically used in the output layer for multi-class classification?

a)

Sigmoid

b)

ReLU

c)

Softmax

d)

Tanh

71.

What issue can arise from activation functions in deeper networks, as mentioned in the lecture?

a)

Vanishing or exploding gradients

b)

Linear decision boundaries

c)

Overly simple computations

d)

Inability to handle binary inputs

72.

What does the Universal Approximation Theorem state, as per the lecture?

a)

Neural networks cannot represent non-linear functions

b)

Neural networks with non-linear activation functions can represent any non-linear function

c)

Single-layer perceptrons are sufficient for all tasks

d)

Deep networks are less expressive than shallow ones

73.

How many neurons are theoretically required in a single hidden layer to represent any Boolean function with n n n inputs?

a)

n n n

b)

2n 2n 2n

c)

2n 2^n 2n

d)

n2 n^2 n2

74.

Why is a single hidden layer with many neurons impractical for real-world tasks like image recognition?

a)

It is computationally inefficient due to the large number of neurons required

b)

It cannot learn non-linear patterns

c)

It eliminates the need for multiple layers

d)

It is only suitable for binary classification

75.

What defines a shallow neural network, as discussed in the lecture?

a)

A network with no hidden layers

b)

A network with one or a few hidden layers

c)

A network with more than two hidden layers

d)

A network with a single neuron per layer

76.

What is a deep neural network, according to the lecture?

a)

A network with exactly one hidden layer

b)

A network with more than two hidden layers

c)

A network with no activation functions

d)

A network with fewer than 10 neurons

77.

How is the number of neurons in the input layer determined for the MNIST dataset?

a)

It equals the number of classes

b)

It equals the size of the input image (e.g., 784 for 28x28 images)

c)

It is a hyperparameter chosen by the engineer

d)

It depends on the number of hidden layers

78.

What is a hyperparameter in the context of hidden layers?

a)

The number of input neurons

b)

The number of hidden layers and neurons per layer

c)

The activation function used in the output layer

d)

The size of the output layer

79.

What happens if the number of neurons in hidden layers is too high?

a)

The network underfits the data

b)

The network may overfit the data

c)

The network becomes linearly separable

d)

The network requires fewer layers

80.

How many neurons are typically used in the output layer for multi-class classification?

a)

One neuron

b)

Two neurons

c)

Equal to the number of classes

d)

Equal to the number of inputs

81.

What is a fully connected neural network?

a)

A network where only some neurons are connected

b)

A network where all neurons in one layer are connected to all neurons in the next layer

c)

A network with no hidden layers

d)

A netw

82.

What is a requirement for backpropagation, as stated in the lecture?

a)

A non-differentiable activation function

b)

A dataset with input-output pairs

c)

A single-layer network

d)

A fixed learning rate

83.

Which gradient descent variant computes the gradient for the entire training set?

a)

Stochastic Gradient Descent

b)

Mini-Batch Gradient Descent

c)

Batch Gradient Descent

d)

Online Gradient Descent

84.

What is a benefit of Mini-Batch Gradient Descent compared to Batch Gradient Descent?

a)

It uses the entire dataset for each update

b)

It reduces variance in parameter updates for more stable convergence

c)

It is slower but more accurate

d)

It eliminates the need for a learning rate

85.

What is a challenge of gradient descent mentioned in the lecture?

a)

It always converges to the global minimum

b)

It can get trapped in suboptimal local minima or saddle points

c)

It does not require a learning rate

d)

It is only applicable to single-layer networks

86.

What is a key challenge in image classification mentioned in the lecture?

a)

Lack of computational resources

b)

Non-linear decision boundaries for non-linearly separable data

c)

Inability to use activation functions

d)

Absence of labeled datasets

87.

What issue arises when extracting features from images by flattening them into vectors?

a)

Increased computational efficiency

b)

Loss of spatial information between pixels

c)

Improved feature detection

d)

Reduced dataset size

88.

Which of the following is NOT listed as a challenge in processing image data for feature extraction?

a)

Viewpoint variation

b)

Illumination changes

c)

Lack of activation functions

d)

Background clutter

89.

What happens when an image dataset like MNIST is represented as a vector with 784 columns?

a)

The dataset becomes low-dimensional

b)

The dataset becomes high-dimensional, increasing computational complexity

c)

Spatial relationships are preserved

d)

The dataset is automatically classified

90.

What is a consequence of intra-class variation in image classification?

a)

It simplifies feature extraction

b)

It makes it harder to generalize across similar objects

c)

It reduces the need for deep networks

d)

It eliminates the need for pooling layers

91.

What is a strength of FCNs highlighted in the lecture?

a)

They preserve spatial information in images

b)

They use hierarchical layers to learn complex patterns

c)

They require fewer parameters than CNNs

d)

They are designed specifically for image data

92.

Why are FCNs not ideal for image recognition tasks?

a)

They cannot handle non-linear data

b)

They treat pixels as independent features, losing spatial relationships

c)

They are computationally inexpensive

d)

They automatically extract features

93.

What is a major drawback of having too many parameters in an FCN?

a)

Reduced risk of overfitting

b)

Increased computational cost and risk of overfitting

c)

Improved generalization to unseen data

d)

Simplified training process

94.

How many total trainable parameters were calculated for the FCN example with three hidden layers and an output layer?

a)

50,240

b)

94,154

c)

33,024

d)

8,320

95.

What is a proposed solution to reduce the number of parameters in FCNs for image processing?

a)

Increase the number of hidden layers

b)

Extract meaningful features to reduce input dimensionality

c)

Use linear activation functions

d)

Flatten images into larger vectors

96.

What is a key property of a good activation function?

a)

Non-differentiability to simplify optimization

b)

Non-linearity to learn complex patterns

c)

Binary output for all inputs

d)

High computational complexity

97.

Why is the step function not recommended for image classification tasks?

a)

It is computationally expensive

b)

It is non-differentiable, hindering backpropagation

c)

It supports complex decision boundaries

d)

It avoids vanishing gradients

98.

What is a disadvantage of the sigmoid activation function in deep networks?

a)

It is not differentiable

b)

It causes vanishing gradients for large inputs

c)

It is zero-centered

d)

It is computationally inexpensive

99.

What is the primary advantage of the ReLU activation function?

a)

It avoids the vanishing gradient problem for positive inputs

b)

It is zero-centered

c)

It produces binary outputs

d)

It is computationally expensive

100.

How does Leaky ReLU address the "dying ReLU" problem?

a)

It sets all outputs to zero

b)

It allows a small gradient for negative inputs

c)

It eliminates non-linearity

d)

It increases computational complexity

101.

What distinguishes CNNs from FCNs?

a)

CNNs assume inputs are in image format and preserve spatial information

b)

CNNs use only fully connected layers

c)

CNNs are less efficient in parameter usage

d)

CNNs cannot use backpropagation

102.

What is the typical structure of a CNN architecture?

a)

Input → Fully Connected → Output

b)

Input → Convolution → Activation → Pooling → Fully Connected

c)

Input → Pooling → Convolution → Output

d)

Input → Activation → Fully Connected

103.

What is the role of the convolution operation in CNNs?

a)

To flatten the image into a vector

b)

To detect local patterns like edges and textures

c)

To reduce the depth of the feature map

d)

To eliminate non-linearity

104.

What is the purpose of the pooling layer in CNNs?

a)

To increase spatial dimensions

b)

To reduce spatial dimensions while retaining essential information

c)

To introduce learnable parameters

d)

To perform classification

105.

Which pooling operation is commonly used in CNNs?

a)

Sum pooling

b)

Max pooling

c)

Median pooling

d)

Random pooling

106.

How are CNNs trained, according to the lecture?

a)

Using forward pass and backpropagation with gradient descent

b)

Using only forward computation without weight updates

c)

Using manual feature extraction

d)

Without activation functions

107.

What is computed during the backward pass of a convolutional layer?

a)

The input image

b)

Gradients with respect to the filter and input

c)

The final class probabilities

d)

The spatial dimensions of the output

108.

How is the gradient with respect to the filter computed in the backward pass?

a)

As a convolution between the input and the error gradient

b)

As a sum of all pixel values

c)

As a flattening operation

d)

As a pooling operation

109.

What is the role of the fully connected layer in a CNN for image classification?

a)

To extract low-level features

b)

To classify the input based on high-level features

c)

To reduce spatial dimensions

d)

To introduce non-linearity

110.

Why does a deeper CNN with convolutional and pooling layers achieve higher accuracy (91% vs. 86%) compared to a shallow FCN, as mentioned in the lecture?

a)

It uses fewer parameters

b)

It leverages convolution and pooling to extract and preserve spatial features

c)

It avoids backpropagation

d)

It eliminates the need for activation functions

111.

What is a primary challenge in training CNNs mentioned in the lecture?

a)

Inability to use activation functions

b)

Balancing overfitting and underfitting for generalization

c)

Lack of computational resources

d)

Absence of labeled datasets