WorksheetsADM0307_ML_Unit 2
Total questions: 100
Worksheet time: 50mins
Name
Class
Date
1.
A Perceptron has weights $w_1 = 0.5, w_2 = -0.2$ and bias $b = -0.4$. If the input is $(1, 1)$ and the activation is a step function (1 if $x \geq 0$, else 0), what is the output?
a)
1
b)
0
c)
0.3
d)
-0.1
2.
In a backpropagation network, if the error at the output layer is 0.02 and the derivative of the activation function is 0.25, what is the weight update value given a learning rate of 0.5 and input of 1?
a)
0.0025
b)
0.01
c)
0.005
d)
0.025
3.
Why is a single-layer perceptron unable to solve the XOR problem?
a)
Linear inseparability
b)
High bias
c)
Vanishing gradient
d)
Overfitting
4.
Given a Genetic Algorithm with a 4-bit chromosome, if the initial population is {1011, 0001} and a single-point crossover occurs at the 2nd position, what are the offspring?
a)
{1001, 0011}
b)
{1111, 0000}
c)
{1011, 0001}
d)
{1100, 0011}
5.
Analyzing the "Vanishing Gradient" problem in Deep Networks, which activation function is most likely to mitigate this?
a)
Sigmoid
b)
Tanh
c)
ReLU
d)
Step
6.
A neural network for a 3-class classification problem has a Softmax output layer. If the raw scores are [2.0, 1.0, 0.1], which class is the most probable?
a)
Class 1
b)
Class 2
c)
Class 3
d)
All equal
7.
In Genetic Programming, what differentiates it primarily from standard Genetic Algorithms?
a)
Fitness function
b)
Chromosome structure (Trees vs Strings)
c)
Mutation rate
d)
Population size
8.
If a Backpropagation algorithm is stuck in a local minimum, which modification to the learning process would best analyze and fix the issue?
a)
Decrease learning rate
b)
Add Momentum
c)
Increase bias
d)
Reduce layers
9.
For a perceptron to behave as an AND gate with inputs {0,1}, what is a valid set of weights $(w_1, w_2)$ and bias $b$?
a)
(1, 1, -1.5)
b)
(1, 1, -0.5)
c)
(0.5, 0.5, 0)
d)
(1, 0, -1)
10.
In GA, if the fitness of individuals A, B, and C are 10, 20, and 70 respectively, what is the probability of selecting C using Roulette Wheel selection?
a)
0.7
b)
0.2
c)
0.1
d)
1
11.
A perceptron uses the weight vector $W = [0.2, -0.5, 0.4]$ for inputs $[x_1, x_2, 1]$. If the learning rate $\eta = 0.1$ and the target output for input $[1, 0, 1]$ is 1, but the current output is 0, what is the updated weight vector?
a)
$[0.3, -0.5, 0.5]$
b)
$[0.1, -0.5, 0.3]$
c)
$[0.2, -0.4, 0.4]$
d)
$[0.3, -0.4, 0.5]$
12.
In a Multilayer Perceptron, if the hidden layer has 4 units and the output layer has 2 units, how many total weight parameters (excluding biases) exist between these two layers?
a)
6
b)
8
c)
16
d)
4
13.
Analyze the effect of a learning rate $\eta = 1.0$ in a Backpropagation algorithm. What is the most likely consequence for a complex error surface?
a)
Faster convergence
b)
Oscillation or divergence
c)
Getting stuck in local minima
d)
Improved generalization
14.
Given a chromosome represented by the bit string $10101101$, what is the result of a bit-flip mutation at the $3^{rd}$ and $5^{th}$ positions (from left)?
a)
$10000101$
b)
$10111101$
c)
$10001101$
d)
$10100101$
15.
In a fitness-proportionate selection, if the fitness values are $f_1=5, f_2=15, f_3=30$, what is the probability of selecting the second individual ($f_2$)?
a)
0.15
b)
0.25
c)
0.3
d)
0.5
16.
A neuron has inputs $[2, 3]$ with weights $[0.5, -1]$ and a bias of $1.5$. Using the ReLU activation function, what is the final output?
a)
0
b)
0.5
c)
-0.5
d)
1.5
17.
Which of the following best describes the Hypothesis Space Search in Genetic Algorithms compared to Gradient Descent?
a)
Local search vs Global search
b)
Continuous vs Discrete
c)
Stochastic global search vs Gradient-based local search
d)
Linear vs Non-linear
18.
In Genetic Programming, if we represent the expression $(a + b) * c$ as a tree, which node serves as the root?
a)
$a$
b)
$+$
c)
$*$
d)
$c$
19.
Evaluate the derivative of the Sigmoid function $\sigma(x)$ at $x=0$. Given $\sigma(0) = 0.5$.
a)
0
b)
0.25
c)
0.5
d)
1
20.
In a deep neural network, if the input dimension is 100 and the first hidden layer has 50 neurons, what is the shape of the weight matrix $W_1$?
a)
$50 \times 100$
b)
$100 \times 50$
c)
$100 \times 100$
d)
$50 \times 50$
21.
Using the Perceptron Learning Rule, if the input is $x = [1, 2]$, weight $w = [0, 0]$, bias $b = 0$, target $d = 1$, and $\eta = 0.5$, calculate the new bias after one update.
a)
0
b)
0.5
c)
1
d)
-0.5
22.
Analyze the impact of "Dropout" in a Deep Neural Network during training. What is its primary numerical effect on the weights?
a)
Increases weight values
b)
Prevents co-adaptation of features
c)
Zeroes out the bias
d)
Doubles the learning rate
23.
Two 6-bit strings $110110$ and $011011$ undergo two-point crossover between points 2 and 4. What is one of the possible offspring?
a)
$111010$
b)
$111011$
c)
$010111$
d)
$110111$
24.
In a Backpropagation network, the error $E$ is $0.5(d-y)^2$. If $d=1, y=0.8$, and the activation is linear ($f(net)=net$), find the gradient $\frac{\partial E}{\partial w}$ for an input $x=10$.
a)
-2
b)
2
c)
-0.2
d)
0.2
25.
How does the Baldwin Effect differ from Lamarckian Evolution in the context of learning and evolution models?
a)
Inherits acquired traits directly
b)
Learning accelerates evolution without direct inheritance
c)
Ignores fitness
d)
Focuses only on mutation
26.
Consider a 3-layer MLP (Input, 1 Hidden, Output). If the input is $[1, 1]$, weights to hidden neuron are $[0.5, 0.5]$ with bias 0, and weight from hidden to output is 1 with bias 0. What is the output if all neurons use Step activation?
a)
0
b)
1
c)
0.5
d)
2
27.
If a Genetic Algorithm population has converged to a single point in the hypothesis space, which operator is most effective to escape this local optimum?
a)
Crossover
b)
Mutation
c)
Tournament Selection
d)
Elitism
28.
Calculate the output of a Tanh neuron for a net input of 0.
a)
0
b)
1
c)
-1
d)
0.5
29.
In Genetic Programming, what is the result of performing a "Subtree Crossover" on two parent trees?
a)
Swapping leaf nodes only
b)
Replacing a branch in one tree with a branch from another
c)
Averaging the node values
d)
Deleting a random node
30.
A neural network is underfitting the training data. Analyzing the architecture, which change is most likely to resolve this?
a)
Increase number of layers
b)
Increase Dropout rate
c)
Decrease training epochs
d)
Increase L2 regularization
31.
In a binary GA, how many possible individuals exist in the hypothesis space if the chromosome length is 10 bits?
a)
100
b)
512
c)
1024
d)
2048
32.
Analyze why the "Mean Squared Error" is often preferred over "Mean Absolute Error" for Backpropagation.
a)
MSE is easier to calculate
b)
MSE is more robust to outliers
c)
MSE has a smoother gradient near zero
d)
MSE prevents overfitting
33.
Given inputs $x_1=1, x_2=0$ and weights $w_1=0.6, w_2=0.3$ and bias $b=-0.5$, calculate the input to the activation function ($net$).
a)
0.1
b)
0.4
c)
0.6
d)
0.9
34.
In the context of Deep Learning, what does "Stochastic Gradient Descent" (SGD) numerically imply?
a)
Using the entire dataset for one update
b)
Using a single random sample for each update
c)
Using a fixed learning rate
d)
Using only positive weights
35.
Compare "Tournament Selection" with "Roulette Wheel Selection". Which one allows for easier control of selection pressure?
a)
Roulette Wheel
b)
Tournament
c)
Both are same
d)
Neither
36.
Calculate the Manhattan distance between two chromosomes $1011$ and $1101$ if treated as vectors.
a)
1
b)
2
c)
3
d)
4
37.
In an MLP, if the activation function is $f(x) = 2x$, and the input is 5, what is the output?
a)
5
b)
10
c)
25
d)
2.5
38.
Analyze the hypothesis space of a single perceptron. It is limited to:
a)
All possible logic gates
b)
Linearly separable hyperplanes
c)
Non-linear manifolds
d)
Polynomial functions
39.
In GA, what is the "Schema Theorem" primarily used to analyze?
a)
Convergence speed
b)
The propagation of small, high-fitness building blocks
c)
Mutation probability
d)
Population size requirements
40.
If the learning rate $\eta$ is reduced to near zero, what happens to the weight update in Backpropagation?
a)
Becomes infinite
b)
Weights stay nearly constant
c)
Weights change randomly
d)
Accuracy improves instantly
41.
Calculate the output of a Softmax layer for inputs $[1, 1]$.
a)
$[0.5, 0.5]$
b)
$[1, 0]$
c)
$[0.7, 0.3]$
d)
$[0, 1]$
42.
What is the primary analytical reason for using "Mini-batch" gradient descent?
a)
To increase memory usage
b)
To balance between speed of SGD and stability of Batch GD
c)
To eliminate the need for bias
d)
To make the network shallower
43.
In a 4-bit chromosome GA, if the fitness function is $f(x) = x^2$ (where $x$ is the decimal value), find the fitness of $1010$.
a)
10
b)
20
c)
100
d)
64
44.
A deep network has 10 layers. If the gradient at layer 10 is 0.1 and each weight is 0.1, what happens to the gradient by the time it reaches layer 1?
a)
It explodes
b)
It vanishes
c)
It remains 0.1
d)
It becomes 1.0
45.
Determine the number of crossover points required to perform a "Uniform Crossover" conceptually.
a)
1
b)
2
c)
$L-1$ (where $L$ is length)
46.
In an MLP, if the weights are initialized to all zeros, what is the analytical problem?
a)
Exploding gradient
b)
Symmetry breaking problem
c)
Overfitting
d)
Vanishing gradient
47.
Given a target $T=0$ and actual output $Y=0.5$, calculate the error using Absolute Error.
a)
0.5
b)
0.25
c)
-0.5
48.
What happens to the diversity of a population in GA if the selection pressure is extremely high?
a)
Increases
b)
Decreases rapidly
c)
Stays same
d)
Becomes random
49.
In a Deep Neural Network, if the input is $32 \times 32 \times 3$ (image), how many neurons are in the input layer?
a)
32
b)
96
c)
1024
d)
3072
50.
Analyze the role of the "Momentum" term in Backpropagation. It numerically helps to:
a)
Slow down training
b)
Overcome local minima and accelerate in flat directions
c)
Reduce the number of layers
d)
Eliminate the bias term
51.
Assertion: Multilayer Perceptrons (MLP) can approximate any continuous function. Reason: Non-linear activation functions allow MLPs to create complex decision boundaries.
a)
Both A and R are true; R is correct explanation
b)
Both A and R true; R is NOT correct explanation
c)
A is true; R is false
d)
A is false; R is true
52.
Assertion: Genetic Algorithms are guaranteed to find the global optimum. Reason: The mutation operator ensures the entire search space can be explored.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R is true
d)
Both false
53.
Assertion: Deep Neural Networks require more data than shallow networks. Reason: They have more parameters, increasing the risk of overfitting.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is true; R is false
d)
A is false; R is true
54.
Assertion: A single-layer perceptron cannot solve the XOR problem. Reason: The XOR function is not linearly separable.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
55.
Assertion: Genetic Algorithms are considered "Stochastic" search methods. Reason: They use random operators like mutation and crossover to explore the hypothesis space.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
56.
Assertion: In Backpropagation, weights are updated in the direction of the gradient. Reason: We aim to maximize the error function during training.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is true; R is false
d)
Both A and R are false
57.
Assertion: Deep Neural Networks are prone to the vanishing gradient problem. Reason: Repeated multiplication of small gradients through many layers causes the update signal to disappear.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
58.
Assertion: Mutation is essential in Genetic Algorithms. Reason: Without mutation, the algorithm may lose diversity and converge prematurely to a local optimum.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
59.
Assertion: The Sigmoid activation function is often replaced by ReLU in deep networks. Reason: Sigmoid functions suffer from saturation, leading to near-zero gradients.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
60.
Assertion: Genetic Programming creates fixed-length binary strings. Reason: It is a variation of GA used to evolve computer programs or mathematical expressions.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R is true
d)
Both false
61.
Assertion: Using a very high learning rate ($\eta$) is always beneficial for convergence. Reason: A high learning rate ensures that the weight updates are large enough to skip local minima.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R is true
d)
Both A and R are false
62.
Assertion: Multilayer Perceptrons require at least one non-linear hidden layer to solve complex problems. Reason: Multiple linear layers can be mathematically simplified into a single linear layer.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
63.
Assertion: Elitism in GAs ensures that the best-performing individual is carried over to the next generation. Reason: This prevents the loss of the best solution due to crossover or mutation.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
64.
Assertion: The "Baldwin Effect" suggests that learning can influence evolution without direct inheritance. Reason: Individuals who can learn have a higher survival rate, eventually allowing the trait to become instinctual through natural selection.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
65.
Assertion: Dropout is a regularization technique used in Deep Learning. Reason: It randomly shuts down neurons during training to prevent the network from becoming overly reliant on specific nodes.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
66.
Assertion: Crossover is the primary operator for "Exploitation" in GAs. Reason: It combines existing high-quality traits from parents to create better offspring.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
67.
Assertion: The Perceptron algorithm will always converge if the data is linearly separable. Reason: The Perceptron Convergence Theorem guarantees a solution in a finite number of steps for separable data.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
68.
Assertion: Tournament selection is more popular than Roulette wheel selection in modern GAs. Reason: Tournament selection allows for easier tuning of selection pressure and is more computationally efficient.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
69.
Assertion: Backpropagation is an unsupervised learning algorithm. Reason: It calculates the error by comparing the predicted output with a target label provided by a teacher.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R is true
d)
Both false
70.
Assertion: Hypothesis space search in GAs is more robust than in Hill Climbing. Reason: GAs maintain a population of points, reducing the risk of getting trapped in a single local optimum.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
71.
Assertion: Fitness scaling is often used in Genetic Algorithms. Reason: It prevents super-fit individuals from dominating the population too early in the search.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
72.
Assertion: The Softmax function is typically used in the output layer of a multi-class classifier. Reason: It transforms raw output scores into a probability distribution that sums to one.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
73.
Assertion: Lamarckian evolution is the standard model used in biological systems. Reason: Acquired characteristics during an organism's life are passed on to its offspring.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R is true
d)
Both false
74.
Assertion: Momentum helps the Backpropagation algorithm navigate "ravines" in the error surface. Reason: It adds a fraction of the previous weight update to the current one, smoothing out oscillations.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
75.
Assertion: Genetic Programming uses tree-based representations. Reason: Trees naturally represent the nested structure of computer programs and mathematical equations.
a)
Both A and R are true; R is the correct explanation
b)
Both A and R are true; R is NOT the correct explanation
c)
A is true; R is false
d)
A is false; R is true
76.
If the drone's vision system misclassifies obstacles in low light, which analysis of the training data is most relevant?
a)
Variance analysis
b)
Gradient descent speed
c)
Feature distribution in low light
d)
Population diversity
77.
The engineer uses a GA to find the shortest path. If the GA converges too quickly to a sub-optimal path, how should the parameters be adjusted?
a)
Increase mutation rate
b)
Increase crossover rate
c)
Decrease population
d)
Increase selection pressure
78.
To improve the drone's ability to learn from its own mistakes in real-time, which "Model of Learning" is most applicable?
a)
Supervised
b)
Unsupervised
c)
Reinforcement Learning
d)
Genetic Programming
79.
During Backpropagation, the drone's weights stop updating. Analysis reveals the gradients are zero. What is this phenomenon?
a)
Exploding Gradient
b)
Vanishing Gradient
c)
Overfitting
d)
Underfitting
80.
The pathfinding GA uses a fitness function $F = 1/Distance$. If the drone must also avoid "No-Fly Zones," how should the fitness function be modified?
a)
Add a penalty term
b)
Multiply by distance
c)
Ignore it
d)
Increase mutation
81.
If the MRI model has high training accuracy but low test accuracy, what is the most likely diagnosis of the model's state?
a)
Underfitting
b)
Overfitting
c)
Vanishing Gradient
d)
Convergence
82.
To analyze why the model fails on different manufacturers, the team should check for:
a)
Learning rate
b)
Data bias/distribution shift
c)
Mutation rate
d)
Number of epochs
83.
The team decides to use "Data Augmentation" (rotating/flipping images). This is an application of:
a)
Increasing Hypothesis Space
b)
Regularization
c)
Backpropagation
d)
Selection Pressure
84.
If the MRI image is $256 \times 256$ pixels, why is a standard MLP less efficient than a Convolutional Neural Network (CNN)?
a)
Too few weights
b)
Parameter explosion
c)
No bias terms
d)
Step activation
85.
The robot needs to reach a target while consuming minimum energy. How should the fitness function be structured?
a)
$F = Energy \times Distance$
b)
$F = 1 / (Energy + Distance)$
c)
$F = Energy + Distance$
d)
$F = 1$
86.
During the search, the GA finds the same path repeatedly. Which operator should be analyzed to increase exploration?
a)
Crossover
b)
Mutation
c)
Selection
d)
Elitism
87.
If the robotic arm must avoid a new obstacle, how does the "Hypothesis Space Search" change?
a)
Space expands
b)
Space remains same; fitness landscape changes
c)
Space shrinks
d)
GA becomes Perceptron
88.
The engineer uses "Rank-based Selection" instead of "Roulette Wheel." What is the primary analytical advantage?
a)
Faster calculation
b)
Prevents dominance by one very fit individual
c)
Eliminates mutation
d)
Reduces bias
89.
In this GP setup, what would a "Node" in the tree most likely represent?
a)
A weight value
b)
A logical condition (e.g., IF-THEN)
c)
A bit string
d)
A learning rate
90.
If the evolved AI is too complex and slows down the game, what constraint should be added to the fitness function?
a)
Maximize depth
b)
Penalty for tree depth/size
c)
Increase mutation
d)
Increase population
91.
When two "AI brains" perform crossover, what is swapped?
a)
Individual bits
b)
Sub-trees (logic branches)
c)
Weights only
d)
Biases only
92.
Analyze the benefit of using GP over a fixed MLP for this case.
a)
GP uses less memory
b)
GP evolves the logic structure itself
c)
GP is faster to train
d)
GP uses gradients
93.
The model predicts the same value regardless of input. Analysis shows all weights are tiny. What is the numerical cause?
a)
Exploding gradients
b)
Vanishing gradients
c)
High learning rate
d)
Large bias
94.
To improve the model, the team adds a "Momentum" term of 0.9. Calculate the effect on the update if the current gradient is 0.1 and previous update was 0.5.
a)
0.45
b)
0.55
c)
0.05
d)
0.1
95.
If the network uses a Sigmoid activation in a 20-layer architecture, why would the early layers learn very slowly?
a)
High bias
b)
Derivative of sigmoid < 0.25
c)
ReLU is faster
d)
Too many neurons
96.
The firm decides to combine GA with NN by using GA to find the optimal weights. This is an example of:
a)
Supervised Learning
b)
Neuroevolution
c)
Genetic Programming
d)
Perceptron Learning
97.
The car fails to turn on a sharp "U-curve." Analyzing the Perceptron, what is the limitation?
a)
Low learning rate
b)
Linear decision boundary
c)
High bias
d)
Too many inputs
98.
The developer adds a hidden layer. What does the "Backpropagation Algorithm" allow the hidden layer to do?
a)
Select the best parent
b)
Learn intermediate features (e.g., curves)
c)
Increase mutation
d)
Randomize weights
99.
Using "Models of Evolution," the car improves its driving through a simulator before hitting the road. This matches:
a)
Lamarckian Evolution
b)
Baldwin Effect
c)
Direct Inheritance
d)
Random Search
100.
If the car's sensors provide continuous values (0.0 to 1.0) instead of binary (0 or 1), which activation is best for the output?
a)
Step function
b)
Sigmoid/Softmax
c)
Bit-flip
d)
Tournament
100 %
