Font size
S
M
L
XL
WorksheetsWeb Mining 2.2
Total questions: 88
Worksheet time: 44mins
Name
Class
Date
1.
In the slides, what is an Artificial Neural Network (ANN) described as simulating?
a)
A biological neural system
b)
A relational database schema
c)
A compression algorithm for text
d)
A deterministic finite automaton
e)
A blockchain consensus protocol
2.
In a perceptron, what does the bias term represent?
a)
A constant input added to shift activation
b)
The learning rate used in training
c)
The number of hidden layers
d)
A method for reducing overfitting
e)
A voting rule across models
3.
Which range of values is stated for the sigmoid activation function output?
a)
Between 0 and 1
b)
Between -1 and 1
c)
Any real number
d)
Only integers
e)
Between 0 and infinity
4.
According to the slides, which connection pattern defines a feedforward neural network?
a)
Connections go only from a previous layer to a following layer
b)
Connections exist only within the same layer
c)
Connections go only from later layers back to earlier layers
d)
Connections are random at every iteration
e)
Connections exist only between output neurons
5.
In the FNN example, what determines how many output signals the network returns?
a)
The number of neurons in the output layer
b)
The number of neurons in the input layer
c)
The size of the training set
d)
The learning rate value
e)
The number of filters in a convolution layer
6.
What does the slide define as the loss function for the whole dataset D?
a)
The average loss over all examples in D
b)
The maximum loss among examples in D
c)
The minimum loss among examples in D
d)
The loss of the first example only
e)
The loss of the last example only
7.
In gradient descent, what does the negative gradient direction indicate?
a)
The direction of fastest decrease of the loss
b)
The direction of fastest increase of the loss
c)
A direction unrelated to the loss
d)
A direction that keeps weights unchanged
e)
A direction that increases model size
8.
What key step is included in backpropagation as stated in the slides?
a)
Backpropagating an error signal from the output layer toward the input layer
b)
Deleting misclassified training examples
c)
Randomly reinitializing weights each epoch
d)
Clustering inputs before training
e)
Removing hidden layers during training
9.
In the incremental gradient descent algorithm shown, what is the first step done to weights w?
a)
Initialize weights to small random values
b)
Set all weights to zero
c)
Set weights to their optimal values
d)
Copy weights from the test set
e)
Quantize weights to integers
10.
Why do the slides warn against using large initial weights with sigmoid units?
a)
Large weights can cause sigmoid saturation
b)
Large weights always improve generalization
c)
Large weights remove the need for a loss function
d)
Large weights guarantee a global optimum
e)
Large weights eliminate noise in data
11.
What is a stated risk of using a very large learning rate?
a)
It can overlook the global optimum
b)
It always makes training too slow
c)
It forces the model to be linear
d)
It prevents any weight updates
e)
It removes the need for hidden layers
12.
What approach is recommended in the slides for choosing the number of neurons in hidden layers?
a)
Choose it based on experiments, starting small
b)
Always match the number of input neurons
c)
Always use exactly one hidden neuron
d)
Always use a fixed number for all problems
e)
Choose it based only on test set performance
13.
According to the slides, what can a 2-hidden-layer ANN perform?
a)
Continuous functions
b)
Only linearly separable functions
c)
Only binary functions
d)
Only strictly discrete functions
e)
Only clustering tasks
14.
Which item is listed as a disadvantage of ANN in the slides?
a)
It is a black-box method that is hard to explain
b)
It cannot learn from data
c)
It requires no hyperparameters
d)
It cannot run in parallel
e)
It always fails with noisy data
15.
Which condition is listed as an application scenario suitable for ANN?
a)
High-dimensional input data
b)
Only low-dimensional input data
c)
Outputs must be strictly symbolic
d)
Training time must be extremely short
e)
Results must always be fully interpretable
16.
Why do the slides say a feedforward neural network (FNN) is limited for image tasks like digit recognition?
a)
It cannot utilize spatial relations among pixels
b)
It cannot produce any numeric output
c)
It cannot be trained with gradient descent
d)
It cannot use hidden layers
e)
It cannot accept an input image
17.
In a CNN, what is the input pattern for a neuron in a convolution hidden layer as described in the slides?
a)
A local area (receptive field) of the input
b)
The entire input image at once
c)
Only the output labels
d)
Only the model bias terms
e)
A random subset of training examples
18.
When a filter is applied across an input with a stride, what does it produce conceptually?
a)
A set of responses at different positions, forming an output map
b)
A new training dataset with fewer examples
c)
A decision tree of rules
d)
A list of class labels without scores
e)
A compressed database index
19.
What term do the slides use for the output of a convolution hidden layer for a filter?
a)
Feature map
b)
Confusion matrix
c)
Gradient vector
d)
Learning curve
e)
Decision boundary
20.
If a convolution layer uses multiple different filters, what does it produce?
a)
Multiple feature maps, one per filter
b)
A single feature map shared by all filters
c)
Only a single scalar output
d)
A set of one-hot encoded vectors
e)
A set of training labels
21.
According to the slides, what does a CNN filter learn?
a)
Patterns in the input image
b)
Only the final class label
c)
Only the learning rate schedule
d)
Only the dataset size
e)
Only the bias initialization
22.
Why do CNNs typically have fewer parameters than fully connected FNNs for images, as stated in the slides?
a)
They use small filters and parameter sharing
b)
They use no weights at all
c)
They require no hidden layers
d)
They always use larger fully connected layers
e)
They avoid any activation functions
23.
What is the main role of a pooling layer according to the slides?
a)
Aggregate important features from convolution outputs
b)
Increase the number of parameters
c)
Replace the need for training data
d)
Convert labels into one-hot vectors
e)
Guarantee perfect accuracy
24.
What effect does pooling have on the representation, as implied by the slides' pooling example?
a)
It reduces the spatial size while keeping key patterns
b)
It increases spatial size while keeping key patterns
c)
It removes the need for convolution filters
d)
It turns images into text
e)
It makes outputs strictly binary
25.
In the full CNN example, which layer is used to output class probabilities for multiple classes?
a)
A softmax output layer
b)
A pooling layer
c)
A convolution filter bank
d)
An input layer
e)
A bias-only layer
26.
What issue does the softmax activation function slide say it addresses?
a)
Inefficiency of mean square error for classification
b)
The need for any loss function
c)
The need for training data
d)
The inability to do pooling
e)
The impossibility of parallel computing
27.
Which pair of activation and loss functions is explicitly contrasted in the slides for CNN training?
a)
Sigmoid with cross entropy versus softmax with log likelihood
b)
ReLU with hinge loss versus tanh with MSE
c)
Linear with MSE versus sigmoid with MSE
d)
Softmax with MSE versus sigmoid with MSE
e)
Step function with L2 loss versus tanh with L1 loss
28.
Which type of problem is listed as a typical application for Recurrent Neural Networks (RNNs)?
a)
Time-series prediction
b)
Sorting numbers without any order
c)
Static image compression only
d)
Database normalization only
e)
Hardware circuit design only
29.
What is a key capability of an RNN according to the slides?
a)
It can store information from recent history
b)
It always removes noise without training
c)
It has no parameters to learn
d)
It uses different parameters at each time step
e)
It cannot process sequences
30.
In the unfolded RNN example, what is stated about the model parameters over time?
a)
They are shared across time steps
b)
They are reinitialized at each time step
c)
They are learned only at the last time step
d)
They are fixed to zero
e)
They depend only on the output labels
31.
In the RNN parameter count slide, which component corresponds to recurrent connections within the hidden layer?
a)
H times H
b)
I times H
c)
H times K
d)
I times K
e)
K times K
32.
In the RNN forward phase definitions, what does a recurrent weight connect?
a)
A previous hidden neuron output to a current hidden neuron input
b)
An input neuron to an output neuron directly
c)
An output neuron to an input neuron directly
d)
A bias term to a training example index
e)
A dictionary word to a sentiment label
33.
In the RNN definitions, what does the activation function (theta) produce for a hidden neuron at time t?
a)
The hidden neuron output at time t
b)
The training label at time t
c)
A new learning rate at time t
d)
A new filter kernel at time t
e)
The dataset size at time t
34.
In the RNN definitions, what does the weight whk connect?
a)
A hidden neuron to an output neuron
b)
An input neuron to an input neuron
c)
An output neuron to a hidden neuron
d)
A bias term to an input neuron
e)
A training example to a classifier
35.
What technique do the slides mention to deal with the vanishing gradient problem in RNNs?
a)
Using LSTM cells
b)
Using k-means clustering
c)
Removing all hidden layers
d)
Using only mean square error
e)
Using a larger dataset label space
36.
In RNNs, vanishing gradients are described as happening along which dimension?
a)
The time dimension
b)
Only the input dimension
c)
Only the output dimension
d)
Only the batch dimension
e)
Only the dictionary dimension
37.
Which gates are listed as components of an LSTM unit in the slides?
a)
Forget, input, and output gates
b)
Start, stop, and reset gates
c)
Convolution, pooling, and softmax gates
d)
Batch, layer, and group gates
e)
Encode, decode, and merge gates
38.
In the LSTM section, what is the purpose of using gates in an LSTM cell?
a)
To control what information is kept, added, and exposed over time
b)
To randomize weights at each time step
c)
To remove the need for hidden states
d)
To convert images into vectors
e)
To eliminate the loss function
39.
In an RNN-style unrolled view over a time axis, what repeats at each time step?
a)
The same computation with shared parameters
b)
A different network architecture each step
c)
A different set of output classes each step
d)
A different dictionary size each step
e)
A different training set each step
40.
Why does the LSTM parameter formula in the slides include a factor of 4?
a)
Because LSTM uses four sets of parameters for its internal transformations
b)
Because LSTM has four input layers
c)
Because LSTM always uses four output classes
d)
Because LSTM uses four different datasets
e)
Because LSTM removes four types of noise
41.
What is the defining idea of a bidirectional RNN as shown in the slides?
a)
It processes the sequence in both forward and backward directions
b)
It processes only the first word of a document
c)
It replaces recurrent connections with convolution filters
d)
It shares no parameters across time
e)
It requires no training data
42.
In the sentiment analysis example, what is used as the representation of the whole document for classification?
a)
The output corresponding to the last word
b)
The first word only
c)
A random word in the middle
d)
Only the dictionary size V
e)
Only the number of output neurons
43.
In one-hot encoding as described, how is a word represented at the input layer?
a)
One neuron is 1 and all others are 0
b)
All neurons are 1
c)
All neurons are 0
d)
Neuron values are random real numbers
e)
Neuron values are the word length
44.
According to the sentiment analysis setup, how many sentiment outputs are produced by the output layer?
a)
Three sentiments
b)
Two sentiments
c)
Four sentiments
d)
Five sentiments
e)
Ten sentiments
45.
What does the term bagging stand for in the slides?
a)
Bootstrap aggregating
b)
Binary aggregation
c)
Batch gradient
d)
Bayesian grouping
e)
Boosted averaging
46.
According to the slides, bagging tends to increase performance for which type of learning algorithm?
a)
Unstable algorithms such as decision trees
b)
Stable algorithms such as Naive Bayes
c)
Algorithms with no parameters
d)
Algorithms that cannot be trained
e)
Algorithms that require no data
47.
In boosting, what happens to misclassified examples during training according to the slides?
a)
Their weights are adjusted so later classifiers focus on them
b)
They are removed from the dataset permanently
c)
They are always assigned zero weight
d)
They become the only test examples
e)
They are converted into new output classes
48.
In the AdaBoost algorithm slide, what condition causes the algorithm to stop early for a given round?
a)
When a classifier error is greater than 1/2
b)
When the learning rate is too small
c)
When the dataset has no noise
d)
When the model has more than one output
e)
When all weights are initialized randomly
49.
If an ANN performs learning by updating connection weights, what is the primary quantity adjusted during training?
a)
Weights between neurons
b)
The number of input features
c)
The number of classes in the dataset
d)
The size of the test set
e)
The dictionary size for one-hot encoding
50.
The slides describe an ANN as a distributed and parallel computing architecture. Which implication best matches this description?
a)
Computation can be carried out by many neurons at the same time
b)
Only one neuron can be active at a time
c)
The model must be trained without data
d)
The network cannot generalize beyond training examples
e)
The output layer must be a decision tree
51.
Why is a continuous activation function with a continuous derivative important for gradient-based training in the slides?
a)
It allows gradients to be computed for weight updates
b)
It guarantees zero training error
c)
It removes the need to choose a learning rate
d)
It makes all functions linearly separable
e)
It prevents any saturation effects
52.
The dataset loss ED(w) is defined as an average over examples. What is a direct consequence of averaging rather than summing without normalization?
a)
The scale of the loss is less dependent on dataset size
b)
The loss becomes independent of model outputs
c)
The loss cannot be minimized by gradient descent
d)
The loss becomes a classification accuracy measure
e)
The loss no longer depends on weights
53.
If the gradient points in the direction of fastest increase of the loss, which update rule best aligns with decreasing the loss as described?
a)
Move weights opposite the gradient
b)
Move weights along the gradient
c)
Do not change weights at all
d)
Randomly permute weights each step
e)
Increase the number of layers instead of updating weights
54.
The slides state that a perceptron can only perform linearly separable functions. What change enables modeling non-linearly separable functions?
a)
Using a multi-layer neural network
b)
Removing the activation function
c)
Setting the learning rate to zero
d)
Replacing weights with fixed constants
e)
Using only an input layer
55.
The slides caution that large initial weights can saturate sigmoid units. What training issue is most consistent with saturation?
a)
Very small gradients that slow learning
b)
Exploding dataset size
c)
Perfect interpretability of the model
d)
Automatic selection of the best learning rate
e)
Removal of the need for backpropagation
56.
A large learning rate can speed up learning but overlook the global optimum. Which adjustment mentioned in the slides can mitigate this trade-off?
a)
Changing the learning rate during training
b)
Removing hidden layers
c)
Using only one training example
d)
Switching to one-hot encoding
e)
Replacing CNN filters with fully connected layers
57.
The slides recommend starting with a small number of hidden neurons and increasing it if the model cannot converge. What does this imply about model capacity?
a)
Increasing hidden neurons can increase representational capacity
b)
Increasing hidden neurons always decreases capacity
c)
Hidden neurons have no effect on capacity
d)
Capacity depends only on the output layer size
e)
Capacity depends only on the dataset size
58.
One listed disadvantage of ANN is being hard to explain. In what scenario from the slides would this be less of a concern?
a)
When results explanation is not mandatory
b)
When the input has very low dimension
c)
When training time must be extremely short
d)
When outputs must be symbolic only
e)
When the model must be fully interpretable
59.
The slides list that ANN is suitable when long training time is acceptable and prompt predictions are required. What does this combination suggest?
a)
High upfront training cost is tolerated to enable fast inference
b)
Training is skipped and only inference is done
c)
Inference must be slower than training
d)
The model cannot generalize
e)
The model must use only kNN
60.
For handwritten digit recognition, why would a CNN be preferred over an FNN according to the slides?
a)
CNN can exploit spatial relations in images
b)
CNN cannot be trained with backpropagation
c)
FNN has no learnable weights
d)
CNN requires no training data
e)
FNN always has fewer parameters than CNN
61.
A CNN filter connects each hidden neuron to a local area of the input. What advantage does this local connectivity provide for image data?
a)
It focuses on local patterns that can appear anywhere in the image
b)
It forces the model to ignore all patterns
c)
It guarantees exact translation of the image
d)
It eliminates the need for pooling
e)
It prevents overfitting without any other technique
62.
Parameter sharing in CNN is motivated by translational invariance. What does translational invariance mean in this context?
a)
A feature can be useful regardless of where it appears in the input
b)
The input must be translated into another language
c)
The model output is invariant to the number of classes
d)
Weights must be different for each position
e)
The dataset must be duplicated many times
63.
Pooling is described as removing positional information while keeping useful patterns. What kind of robustness does this mainly support?
a)
Robustness to small shifts in feature location
b)
Robustness to changes in dataset size
c)
Robustness to the learning rate value
d)
Robustness to the number of output classes
e)
Robustness to missing labels
64.
Why can a CNN have fewer parameters than a fully connected FNN for the same image input, based on the slides?
a)
Because filters reuse the same weights across many locations
b)
Because CNN has no biases
c)
Because CNN has no hidden layers
d)
Because CNN uses only one training example
e)
Because FNN cannot use any activation function
65.
The slides mention softmax with log likelihood and sigmoid with cross entropy. For a single-label multi-class problem, which pair is most appropriate?
a)
Softmax with log likelihood
b)
Sigmoid with mean square error
c)
Linear activation with L2 pooling
d)
Step activation with voting
e)
No activation with no loss
66.
The slides state CNN learns high-level abstraction using convolution and pooling. What is the most direct role of stacking these operations?
a)
Building increasingly abstract features from simpler ones
b)
Guaranteeing no overfitting in any case
c)
Eliminating the need for labeled data
d)
Making gradients unnecessary
e)
Forcing the network to be linear
67.
RNNs are said to be useful when output depends on current input and previous outputs. What property of RNNs supports this behavior?
a)
Using the previous hidden state as part of the current input
b)
Using independent weights at every time step
c)
Removing recurrence to avoid cycles
d)
Replacing sequence data with images
e)
Using only pooling layers
68.
Because RNN parameters are shared over time, what practical benefit follows for processing sequences?
a)
The same model can be applied to different sequence lengths
b)
The model requires a different architecture for each time step
c)
The model cannot generalize to new sequences
d)
The model output must be a single number only
e)
The model cannot be trained with backpropagation
69.
The slides connect vanishing gradients in RNNs to the time dimension. What learning difficulty does this most directly cause?
a)
Difficulty learning long-range dependencies
b)
Difficulty reading the current input value
c)
Difficulty producing any output at all
d)
Difficulty defining a loss function
e)
Difficulty performing parallel computing
70.
Given that LSTM is introduced to address vanishing gradients, which LSTM mechanism is most directly intended to manage information flow over time?
a)
Gates that control remembering and forgetting
b)
Max pooling over time steps
c)
Using fully connected filters
d)
Randomly reinitializing hidden states
e)
Removing the hidden layer entirely
71.
Bagging trains k classifiers on bootstrap samples and combines them by voting. Which goal best matches this strategy?
a)
Reducing variance of an unstable learner
b)
Increasing bias of a stable learner
c)
Eliminating the need for training labels
d)
Guaranteeing a global optimum
e)
Converting regression into clustering
72.
Boosting creates a sequence of classifiers and increases weight on misclassified examples. What is the main intent of this reweighting?
a)
Force later classifiers to focus on harder cases
b)
Remove hard cases from the dataset
c)
Make all examples have identical predictions
d)
Prevent any classifier from being created
e)
Turn training into unsupervised learning
73.
A network uses sigmoid activations and is initialized with very large weights. Based on the slides, what is the most likely training failure mode?
a)
Training slows because activations saturate and gradients become very small
b)
Training becomes impossible because the loss is undefined
c)
Training instantly reaches the global optimum
d)
Training becomes faster because gradients are larger everywhere
e)
Training removes the need for backpropagation
74.
You observe oscillating or divergent loss during gradient descent. Using only slide guidance, which change is the most defensible first intervention?
a)
Reduce the learning rate or use a schedule that changes it during training
b)
Increase the learning rate further to escape local minima
c)
Remove the activation function to make the model linear
d)
Freeze all weights and train only biases
e)
Replace gradient descent with voting across random models
75.
The slides suggest increasing hidden neurons if a model cannot converge. Which interpretation best reconciles this advice with neural network behavior?
a)
The current model may lack capacity to represent the target mapping, so more neurons can help optimization find a better fit
b)
More neurons always guarantee faster convergence regardless of data
c)
More neurons always reduce the number of parameters
d)
Convergence depends only on the output layer size, not hidden layers
e)
If a model cannot converge, it must be because the dataset is too small
76.
A task requires exact location of a detected feature to remain available to later layers. Based on the pooling description in the slides, what is a key risk of aggressive pooling?
a)
It can discard positional information that the task needs
b)
It guarantees overfitting by increasing parameters
c)
It prevents any feature extraction in convolution layers
d)
It makes the model unable to accept image inputs
e)
It forces the output to be only binary
77.
A classification problem has exactly one correct class per example. Which activation and loss pairing is most aligned with the slide comparison, and why?
a)
Softmax with log likelihood, because it models a normalized distribution across classes
b)
Sigmoid with cross entropy, because it forces probabilities to sum to 1 across classes
c)
Mean square error with softmax, because it is stated as most efficient for classification
d)
No activation with no loss, because classification needs only voting
e)
Sigmoid with mean square error, because the slides recommend it for multi-class
78.
CNN parameter sharing is motivated by translational invariance. In which scenario would this assumption be most questionable?
a)
When the meaning of a pattern depends strongly on its absolute position in the input
b)
When the input contains local patterns repeated across the image
c)
When the dataset has noise in input values
d)
When training time is acceptable but prediction must be fast
e)
When outputs are numeric vectors
79.
The slides say FNNs cannot utilize spatial relations of pixels, while CNNs can. Which design choice is the most direct reason CNNs can exploit spatial structure?
a)
Local receptive fields combined with shared filters
b)
Using a larger learning rate
c)
Using only fully connected layers
d)
Removing the loss function
e)
Using one-hot encoding on pixels
80.
An RNN must produce outputs in real time as each new input arrives. Based on the bidirectional RNN concept, what is the main limitation of a bidirectional RNN for this setting?
a)
It needs future context, so it cannot fully operate causally in strict real-time
b)
It has no learnable parameters
c)
It cannot share parameters over time
d)
It cannot store any history
e)
It cannot be trained by backpropagation
81.
The slides suggest using the last output in an RNN as a document representation for sentiment analysis. What is a plausible failure case of this approach for long documents?
a)
Early important information may be poorly preserved due to vanishing gradients or limited memory
b)
The output layer cannot represent three sentiment classes
c)
One-hot encoding makes the input dense and information-rich
d)
Using the last output guarantees perfect interpretability
e)
Pooling layers prevent any sequence modeling
82.
Vanishing gradients occur from later to earlier layers in deep networks and also across time in RNNs. What common consequence follows in both cases?
a)
Early layers or early time steps learn much more slowly than later ones
b)
Later layers always have zero gradients while early layers are large
c)
The model cannot compute any forward pass outputs
d)
The model becomes linear regardless of activations
e)
The dataset loss no longer depends on weights
83.
LSTM is introduced to address vanishing gradients, but it has more parameters than a basic RNN. What trade-off is most consistent with this change?
a)
Improved ability to keep information over time at the cost of higher model complexity
b)
Lower complexity with identical ability to model long-term dependencies
c)
Fewer parameters and shorter training time in all cases
d)
Guaranteed interpretability due to gates
e)
Elimination of the need to choose hyperparameters
84.
Bagging may decrease performance of stable algorithms like Naive Bayes or kNN, according to the slides. Which explanation best fits this claim?
a)
If a learner is already stable, resampling adds little benefit and may inject extra variability or bias
b)
Stable learners cannot be trained on bootstrap samples
c)
Bagging forces all base learners to be identical and perfectly correlated
d)
Voting always decreases accuracy regardless of base learner
e)
Bagging converts classification into regression automatically
85.
Boosting reweights misclassified examples so later classifiers focus on them. What is a common risk implied by this strategy when data contains noise?
a)
The ensemble may over-focus on noisy or mislabeled examples
b)
The ensemble cannot use weighted voting at test time
c)
The ensemble loses the ability to represent any non-linear function
d)
The ensemble becomes a stable learner by definition
e)
The ensemble cannot be trained sequentially
86.
In AdaBoost, a round can be skipped if the classifier error exceeds 0.5. What is the most principled reason for this threshold?
a)
A weak classifier must perform better than random guessing for boosting to be meaningful
b)
Any error rate above 0.5 guarantees overfitting
c)
0.5 is the maximum possible error for any classifier
d)
The threshold ensures the model uses exactly two classes
e)
The threshold makes training independent of the dataset
87.
The slides state that ANN structure and hyperparameters are chosen by experiment. Which combined set of choices best matches this experimental tuning as presented?
a)
Learning rate and number of hidden neurons
b)
Dataset label names and slide background color
c)
Pixel brightness thresholds and font sizes
d)
Database index type and file format
e)
Choice of university motto and logo
88.
A CNN uses multiple filters, pooling, and a fully connected layer before softmax. If overfitting is a concern, which part of this pipeline is most likely to add many parameters and thus increase overfitting risk?
a)
The fully connected layer
b)
Parameter sharing in convolution filters
c)
Pooling layers
d)
Using local receptive fields
e)
Using translational invariance
Reset
