AI Quiz

1. 1What is the analogy in Artificial Neural Networks to synaptic efficiencies in biological neurons?

   1. AThe strength of the weights between neurons
   2. BThey can be updated depending on learning performance
   3. CThe hidden layers with neurons
   4. DThey can classify different objects
2. 2Which of the following non-derivative functions can only produce outputs in the range 0 to 1? [may be more than one correct answer]

   1. ALinear activation
   2. BThe ReLU activation function
   3. CThe sigmoid function
   4. DThe softmax function
3. 3How many weights (excluding threshold/bias) does a feedforward, fully-connected, neural network with a single hidden layer and with 5 inputs nodes, 3 hidden nodes and 1 output node and no biases have? [only one correct answer]

   1. A13
   2. B18
   3. C9
   4. D20
4. 4What is supervised learning?

   1. AEntails the ANN learning to predict specific outputs (targets) labelled by human
   2. BEntails the ANN learning without specific human-labelled targets.
5. 5What is unsupervised learning?

   1. AEntails the ANN learning to predict specific outputs (targets) labelled by human
   2. BEntails the ANN learning without specific human-labelled targets.
6. 6What is the vanishing gradient problem?

   1. AThe deeper is the artificial neural network, the harder it is to train
   2. BDetermining how the success of a system’s overall performance is due to the variance of parameters
   3. CLack of computational power and availability of data sets
   4. DInefficient learning algorithms
7. 7How can you deal with the problem of the ‘vanishing gradient’ when training MLPs using backpropagation with gradient descent.

   1. AUsing the ReLU function instead of sigmoid
   2. BUsing the sigmoid function instead of ReLU
   3. CPre-training layers of the deeper network
   4. DThe vanishing gradient can't be solved
8. 8Which of the following, given appropriate connection weights, can a Multi-Layer Perceptron with one hidden layer find linearly-separable classifications for its inputs? [may be more than one correct answer]

   1. AAND logic functions
   2. BOR logic functions
   3. CXOR logic functions
   4. DNot-AND logic functions
9. 9What is the composition of a ‘shallow’ autoencoder?

   1. AA single hidden layer compresses information into a lower dimensional space
   2. BFeedforward connected (every neuron in one layer is connected to every neuron)
   3. CThe number of input nodes are the same as the output nodes
   4. DThe number of output nodes are twice as many as the input nodes
10. 10How can autoencoders help us understand representations in the brain (e.g. visual cortex)?

    1. AThe output layer of the autoencoder re-presents the inputs.
    2. BAutoencoders when trained in an unsupervised way appear to learn a number of
       feature detectors – individual hidden nodes connected to input nodes.
    3. CThe first layer weights represent complex features
    4. DThe input of an Autoencoder is converted to 2D

11. 11The below image depicts the input-hidden node weight vectors for a single hidden layer (non-classifier) autoencoder visualized at different stages of learning (left early stage; right: late stage).
    
    Input vectors are taken from large input images (for example 512 by 512 pixels) using small patches (samples) of the images. The input vectors are fully (feedforward) connected to the hidden layer, i.e. every input node connects to every hidden node. Look carefully at the figures:
    
    How many hidden nodes are there?

    

    1. A25
    2. B18
    3. C100
    4. D512
12. 12The below image depicts the input-hidden node weight vectors for a single hidden layer (non-classifier) autoencoder visualized at different stages of learning (left early stage; right: late stage).
    
    Input vectors are taken from large input images (for example 512 by 512 pixels) using small patches (samples) of the images. The input vectors are fully (feedforward) connected to the hidden layer, i.e. every input node connects to every hidden node. Look carefully at the figures:
    
    What is the receptive field of each hidden node?

    

    1. A8x8
    2. B5x5
    3. C10x10
    4. D2x2
13. 13The below image depicts the input-hidden node weight vectors for a single hidden layer (non-classifier) autoencoder visualized at different stages of learning (left early stage; right: late stage).
    
    Input vectors are taken from large input images (for example 512 by 512 pixels) using small patches (samples) of the images. The input vectors are fully (feedforward) connected to the hidden layer, i.e. every input node connects to every hidden node. Look carefully at the figures:
    
    This network has 2 output nodes. How many weights (parameters) exist between the hidden layer and the output layer?

    

    1. A25 x 2 = 50
    2. B8 x 8 = 64
    3. C5 x 5 = 25
    4. D2 x 2 = 4
14. 14In what ways can Convolutional Neural Networks be seen as similar to how the Visual Cortex processes stimuli?

    1. ACells in V1 are receptive to simple features
    2. BCells in Visual cortex also appear to encode for pooling.
    3. CCells in higher levels of visual cortex and Inferotemporal cortex are responsive to objects and relatively invariant to the spatial and orientation presentation of the object.
    4. DCells in V1 are receptive to complex features
    5. ECells in Visual cortex also appear to decode for pooling.
15. 15How many pooling maps does a standard ‘shallow’ Convolutional Neural Network (1 layer of feature maps + 1 layer of pooling maps) have that uses 5 filters of dimensionality 4x4 over the input image?

    1. A5
    2. B10
    3. C16
    4. D32
16. 16How does Convolutional Neural Networks encode for spatial invariance when being trained on multiple images of the same class?

    1. ACNNs use pooling maps that encode a lower dimensional ‘summary’ of the space of the features maps to which they respond by encoding
    2. BDeeper CNNs have increasingly lower dimensionality (fewer neurons) in deeper layers
    3. CThe pooling neuron’s receptive field covers the entire image
    4. DThe receptive field becomes smaller, and neurons become increasingly variant to shifts of the object in the image
17. 17In what way does Deep Convolutional Neural Networks allow for a small number of weights/parameters with respect to MLPs like stacked autoencoders.

    1. AThey use low-dimensional filters over the images they are processing and these filters are shared for the feature map that is encoding the same feature in different positions in space
    2. BA large number of weights are found between the final pooling layer and the fully connected layers.
    3. CThe CNN is fully connected all the way through
    4. DA large number of weights are found between the first pooling layer and the fully connected layers.
18. 18Why might deeper Convolutional Neural Networks be considered superior to more shallow Convolutional Neural Networks?

    1. ATraining efficiency, spatial invariance and feature combination
    2. BBottleneck layer, training efficiency and filtering out noise
    3. CSpatial invariance, larger datasets and more weight connections
    4. DHierarchical decomposition, feature combination and pre training
19. 19What non-linear logical function/functions can a Single Layer Perceptron not do?

    1. AXOR
    2. BAND
    3. COR
    4. DNOR
20. 20How can we train a simple ANN?

    1. AManually select weight values or use a learning rule for weight values
    2. BManually select weight values or use backpropagation for choosing the weight values
    3. CUse backpropagation or use another AI for choosing the weight values
    4. DBy not using any weights or let the algorithm train itself
21. 21What are the steps of training the backpropagation algorithm?
    
    Forward pass
    [blank_start]1.[blank_end]
    [blank_start]2.[blank_end]
    [blank_start]3.[blank_end]
    [blank_start]4.[blank_end]
    
    Backward pass
    [blank_start]5.[blank_end]
    [blank_start]6.[blank_end]
    [blank_start]7.[blank_end]
    [blank_start]8.[blank_end]

    1. APresent data vector at the input layer
    2. BPass weighted input layer activations
    3. CPass weighted input layer activations
    4. DCalculate the δ (delta) output error
    5. EPass weighted hidden layer activation
    6. FApply the target value to the output
    7. GApply the target value to the output
    8. HCalculate δ h error value for each node
    9. ICalculate δ h error value for each node
    10. JPass weighted hidden layer activation
    11. KUpdate hidden-output layer weights
    12. LCalculate δ h error value for each node
    13. MCalculate δ h error value for each node
    14. NUpdate input-hidden layer weights
    15. OUpdate input-hidden layer weights
    16. PPass weighted hidden layer activation
22. 22What does the input vector of a neural network encode?

    1. AData values
    2. BNumerical values (e.g. binary)
    3. CString ("word") values
    4. DWeight values
23. 23What does the hard limiter (threshold function) of a single neuron do?

    1. ATransforms the activation value (input) into a binary value (output)
    2. BChecks the activation value against the threshold value.
    3. CChanges the activation value into the value of the threshold
    4. DTransforms the activation value into a zero only if the activation is lower than the threshold
24. 24What is the dot product of x and w where:
    
    x = [0.4 1.2] and w = [0.2 0.3]

    1. A0.44
    2. B[0.08 0.36]
    3. C0.94
    4. D[0.48 0.06]
25. 25If you have completed 3 epochs of training for a dataset with 4 data samples (e.g. with inputs 0 0; 0 1; 1 0; 1 1) where weights are updated for each data sample, how many iterations of training have been carried out?

    1. A12
    2. B3
    3. C8
    4. D4
26. 26From the below which options best describes a decision boundary for an ANN classifier?

    1. AA line or lines that is a geometric expression of the weights
    2. BA line or lines that approximately separates two or more data types?
    3. CA straight line or lines that is a geometric expression of the weights?
    4. DA line or lines that approximately bisects each data point?
27. 27What is the minimum number of nodes/neurons for a single layer perceptron classifier (non augmented vector version) required to encode a 2 input AND logic function?

    1. A3
    2. B2
    3. C4
    4. D7
28. 28What does credit assignment refer to when training ANNs?

    1. AHow much each weight is responsible for the error of the network?
    2. BHow well the network predicts the target (class label) for the particular data sample?
    3. CHow much each node is responsible for the error of the network?
    4. DWhether the network has finished training (no error)?
29. 29How many tunable weights in total, does a Convolutional Neural Network with one convolutional layer have consisting of the below-listed aspects? Each filtered input to each feature map has the same receptive field. Consider feature and pooling map units as neurons.
    
    Input image size 8x8
    4 input convolution filters of 2x2 dimensions
    Feature map stride 2
    Pooling ‘filter’ dimension 2x2
    Pooling map stride 2
    Classes = 4
    Hints:
    
    Consider that you have weights in the feature extractor part of the network and weights from the extractor part to the classification part of the network
    Account for a single fully connected layer (one neuron per class) in your calculation of weights, but not the softmax layer!

    1. A80
    2. B32
    3. C336
    4. D16
30. 30How many neurons in total, does a Convolutional Neural Network with one convolutional layer have consisting of the below-listed aspects? Each filtered input to each feature map has the same receptive field. Consider feature and pooling map units as neurons.
    
    Input image size 12x12
    4 input convolution filters of 2x2 dimensions
    Feature map stride 2
    Pooling ‘filter’ dimension 2x2
    Pooling map stride 2
    Classes = 4
    Hints:
    
    Do not include the input pixels as neurons
    Include a single fully connected layer (one neuron per class) but not the softmax layer!

    1. A68
    2. B36
    3. C184
    4. D144
31. 31Which of the following is true concerning convolutional filters and pooling "filters"?

    1. APooling filters do not have trainable weight values
    2. BBoth convolutional filters and pooling filters scan over values that represent their "inputs" (e.g. image or feature map)
    3. COnly pooling filters allow for "downsampling"
    4. DConvolutional filters extract the maximum value from the scanned part of the image
32. 32Which of the following are typically used as stopping criteria when training ANNs?

    1. AWhen the loss for the training dataset is below a certain threshold
    2. BWhen the classification error for the training dataset is below a certain threshold
    3. CWhen validation error does not reduce for successive checks
    4. DWhen all data samples have been presented to the network
33. 33If your network has trained on all the data samples (total = 12) but updates weights in "minibatches", i.e. only after a minibatch of 3 (when 3 data samples are presented to the network, the activations are stored and then the weights are updated), how many iterations of training will there be per epoch?

    1. A4
    2. B3
    3. C12
    4. D1
34. 34Which of the following isn't a hyper-parameter?

    1. AThe number of connection weights
    2. BNumber of hidden nodes
    3. CNumber of epochs of training
    4. DThe learning rate
35. 35Which of the following activation functions can produce an output of greater than 1 for a single neuron?

    1. AReLU function
    2. BThreshold function
    3. CSoftmax function
    4. DSigmoid function
36. 36How many neurons does an autoencoder have that has 4 inputs nodes and 3 nodes at the bottleneck (is only trained on its reconstruction error not for classification)?

    1. A11
    2. B24
    3. C14
    4. D7
37. 37Which of the following is true when referring to the notion of transfer learning?

    1. ATransfer learning can be seen as a form of pre-training using the 1st layer of weights
    2. BTransfer learning is typically applied to the weights at the output layer
    3. CTransfer learning is best applied to ANNs to be trained on a very different problem to that which the original weights were trained on
    4. DTransfer learning is most useful for ANNs with multiple hidden layers
38. 38Which of the following loss functions is most typically used for binary class problems or regression for MLPs?

    1. AHalf mean squared error (or Half mean squared classification error)
    2. BCross entropy
    3. CSoftmax
    4. DBinary classification error (t - y)
39. 39Which of the following best describes what the sigmoid function does?

    1. AIt is a squashing function
    2. BIt is a probabilistic version of a binary threshold function
    3. CIt is a continuous and derivable function
    4. DIt allows for outputs that are negative
40. 40When referring to an ANN's classification capabilities for a specific task, what is meant by an "input-output mapping that is not linearly separable"?

    1. AIt cannot find a straight line decision boundary between the classes of data in a geometric space
    2. BAll classified data falls on one side of a decision boundary
    3. CIt is not possible to classify the data using any kind of ANN
    4. DThe classification problem is not solvable by a SLP
41. 41What are deep learning algoritms?

    1. ASuper powerful neural networks
    2. BAn AI learning from deep dream images
    3. CLess powerful neural networks
    4. DA bunch of machine learning algoritms training each other
42. 42What does deep learning enable?

    1. AQuantitative (big data)
    2. BQualitative gains (better technology)
    3. CBackpropagation
    4. DCalculus
43. 43What was the problem at the first AI Winter (1973-1980)?

    1. ACredit assignment
    2. BVanishing gradient problem
    3. CComputational Power
    4. DInefficient learning algorithms
44. 44What was the problem at the second AI Winter?

    1. ACredit assignment
    2. BVanishing gradient problem
    3. CComputational power
    4. DInefficient learning algorithms
45. 45What problems did we have in the 90-00s?

    1. AComputational Power, Availability of Data sets, Inefficient learning algorithms
    2. BVanishing gradient problem, Availability of Data sets, Backpropagation
    3. CComputational Power, Credit Assignment, Inefficient learning algorithms
    4. DVanishing gradient problem, Computational Power, Backpropagation
46. 46How do we find the delta δ (error) values for each layer?

    1. AUse gradient descent principle
    2. BUse the chain rule (from calculus)
    3. CUse a linear activation function
    4. DUse ReLU
47. 47What is backpropagation?

    1. ADemands backchaining error derivatives (delta values) to update the weights
    2. BDemands backchaining error derivatives (delta values) to update the neurons
    3. CDemands backchaining error derivatives (delta values) to update the input
    4. DDemands backchaining error derivatives (delta values) to update the output
48. 48What is the problem of Credit Assignment?

    1. AProblem of knowing which weights in which layers should be updated and by what amount
    2. BProblem of knowing how many nodes the network should have
    3. CProblem of knowing which errors the network will get
    4. DProblem of knowing which layers that should connect to each other
49. 49How is the Credit Assignment problem solved?

    1. ABy using dedrivable cost function and derivable activation functions
    2. BBackpropagation
    3. CBy using linear cost function and linear activation functions
    4. DSoftmax
50. 50What is the Vanishing Gradient Problem?

    1. ADeeper layers are harder to train with backpropagation because that the weights continuous update gets small after many iterations
    2. BDeeper layers are easier to train with backpropagation because that the weights continuous update gets small after many iterations
51. 51How can we reduce the Vanishing Gradient Problem?

    1. APre-training
    2. BReLU
    3. CBackpropagation
    4. DSigmoid
52. 52What does the chain rule do?

    1. AFind the value of a derivative through multiplying the values of a chain of derivatives
    2. BFind the value of the weights through multiplying the values of a chain of weights
    3. CFind the value of the layers through multiplying the values of a chain of derivatives
53. 53What two part does an Autoencoder contains of?

    1. AEncoder
    2. BDecoder
    3. CPre-training
    4. DY and Y^
54. 54What is training parts of the autoencoder (a load of shallow encoders) first?
    

    1. APre-training
    2. BPre-learning
    3. CFine-tuning
    4. DFine-training
55. 55What is when you train the whole stacked autoencoder in the end with backpropagation?

    1. AFine-tuning
    2. BFine-training
    3. CPre-training
    4. DPre-learning
56. 56What are the benefits of sparsity?

    1. AForces training to find only the most important features in the data
    2. BWe can have more hidden nodes/units so can potentially find more complex solutions more quickly
    3. CWe can train network on earlier knowledge from another network
    4. DForces training to indicate if it is overfitting to the training data
57. 57What do we have if we might get stuck in a local minimum before we "navigate" to the global minimum.

    1. AToo low learning rate
    2. BToo high learning rate
58. 58What do we have if we might fly out of the global minimum and oscillate?
    

    1. AToo high learning rate
    2. BToo low learning rate
59. 59How can we evaluate performance?

    1. ATrain, Validate, Test
    2. BConfusion Matrices
    3. CRunning multiple times
    4. DUse pre-settings
60. 60Use ___________ to check whether our ANN is becoming increasingly poor at predicting data that it is not trained on. This indicates it is overfitting to the training data .

    1. AValidation data
    2. BTest data
    3. CInput data
    4. DOutput data
61. 61When do we use cross entropy loss?

    1. AFor multiclass problems
    2. BFor SLPs
    3. CFor bi-class problems
    4. DFor MLPs
62. 62What is cross entropy loss?

    1. AThe performance of a classification model whose output is a probability value between 0 and 1
    2. BThe performance of a classification model whose output is 0 or 1
63. 63What is true about Convolutional Neural Networks?

    1. ASpecialize in feature extraction –scan images using small "filters" to look for simple features in the image
    2. BCNNs are not densely connected but use ”shared” weights
    3. CPre-training approach
    4. DWeights learn correlations between inputs and
       output/hidden nodes
64. 64What does invariance refer to in CNNs?

    1. AThe CNN may recognize a data class even if dimensions of the
       data vary slightly
    2. BPartitioning of data into train/validate/test datasets is random
    3. CIt indicates it is overfitting to the training data
    4. DCNNs systematically scan the image
65. 65What is useful for: Solving linearly separable problems? (e.g. OR and AND Boolean logic functions)

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
66. 66What neural network has this algorithmic approach?
    
    • Activation: feedforward dot product (between inputs and weights) and step function,
    • Learning: perceptron learning rule.

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
67. 67Which neural network has this cognitive/biological relevance?
    
    • Resemble firing (on/off) or a single neuron
    • Implement logic computations (like the ALU of a computer)
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
68. 68Which neural network has this model structure?
    
    Model Structure (1 input layer, 1 output layer, 1
    layer of weights: fully connected input->hidden)

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
69. 69Which neural network is useful for:
    
    Solving non-linearly separable problems (e.g. XOR, Not-XOR Boolean logic functions)

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
70. 70Which neural network has this algoritmic approach:
    
    • Activation: feedforward dot product (between inputs and weights) at each layer, and sigmoid (or tanh) function, softmax at output layer (for classification of >2 classes),
    • Learning: Backpropagation/gradient descent (following forward ‘pass’).

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
71. 71Which neural network has this cognitive/biological relevance:
    
    • Resemble networks of neurons
    • Backpropagation algorithm biological plausibility controversial
    • Compute non-linearly separable logic functions
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
72. 72Which neural network has this model structure: (input layer, 1+ hidden layer, output
    layer, fully connected weights between all layers)
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
73. 73Which neural network is useful for:
    
    • Denoising (re-presenting inputs at output layer)
    • Learning basic features (useful for training on other problems)
    • Good at image classification

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
74. 74Which neural network has this algorithmic approach:
    
    • Activation: feedforward dot product (between inputs and weights) at each layer, and sigmoid (or tanh) function, softmax at output layer (for classification of >2 classes),
    • Pre-training approach
    • Learning: Backpropagation/gradient descent.
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
75. 75Which neural network has this cognitive/biological relevance:
    
    • Visual cortex has many bidirectional synapses, can denoise inputs (at output layer)
    • Reconstruct their inputs according to past learning
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
76. 76Which neural network has this model structure: (input layer, 1+ hidden layer, output layer same number of nodes as input layer, fully connected weights* between all layers; encoder part and decoder part – discard decoder after training for classification, keep decoder for denoising)

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
77. 77In what ANN does weights learn decision boundaries of nested SLPs
    

    1. AMLPs
    2. BSLPs
    3. CAutoencoders
    4. DCNNs
78. 78In what ANN does weights learn decision boundaries?
    

    1. ASLPs
    2. BMLPs
    3. CAutoencoders
    4. DCNNs
79. 79In what ANN does weights learn correlations between inputs and
    output/hidden nodes

    1. AMLPs
    2. BSLPs
    3. CAutoencoders
    4. DCNNs
80. 80Which of the following is an example of a true negative?

    1. AA patient is diagnosed as free from COVID-19 and in fact does not has the disease.
    2. BA patient is diagnosed with COVID-19 but in fact does not have the disease.
    3. CA patient is diagnosed as free from COVID-19 but in fact has the disease.
    4. DA patient is diagnosed with COVID-19 and in fact has the disease.
81. 81Which of the following are examples of stopping criteria that are always desirable for training ANNs?

    1. ANone of the alternatives
    2. BWhen training has gone past a certain (large) number of epochs.
    3. CWhen training error reaches zero.
    4. DWhen validation error has increased following a single test of the model on the validation data set
    5. EWhen training error is no longer reducing (as seen over successive epochs).
82. 82How many weight connections are there in a “dense” (or fully connected) MLP with 3 input nodes, 3 hidden nodes, and 1 output nodes (don’t include bias nodes or connections):

    1. A12
    2. B7
    3. C10
    4. D27
83. 83When activating (feedforward) a MLP with one hidden layer, if we have a dense MLP with 3 inputs nodes, 3 hidden nodes, and 1 output node, how many times do we apply the dot product computation between vectors of weights and vectors of neurons from which the weights are projecting?

    1. A4
    2. B2
    3. C7
    4. D12
84. 84When training MLPs over epochs our weight values will, hopefully allow our error/loss to reduce over time. What are some of the problems we can face during this training?

    1. AOur training gets stuck in a local minimum
    2. BOur learning rate is low
    3. COur learning rate is high
    4. DOur training gets stuck in a global minimum

Description

1 AI Quiz on AI Quiz, created by Tove . on 13/03/2022.

Author

Tove .

Language

English

Views

256

Tags

ai1

Rating

No ratings yet

Rate this quiz

Rate this quiz

More on Neural Networks and Deep Learning