KCS-055 Machine Learning Techniques

Theory Exam, Odd Semester 2022-23 Solution

Paper id: 232119

1.
(a) Discuss model representation of artificial neuron.
Answer: The artificial neuron is an elementary unit in an artificial neural network. It is
a mathematical function modeled on the working of biological neurons. It can be
represented as, set of inputs, multiplied by some weights, then sum up(summation)
this product of weights and input and passing it to the activation function which
mapped the output between 0 and 1. The figure below shows the architecture of an
artificial neuron.

Figure 1: Artificial Neuron

(b) Explain general to specific ordering hypothesis in concept learning.

Answer: General-to-specific ordering hypothesis in machine learning is a process of
refining the categories more accurately to distinguish between different objects,
situations or events. General-to-specific ordering hypothesis can be illustrated as:
Suppose we have two hypotheses, h1 and h1 for weather to play a game, such as:
h1 = < Sunny,?,? Strong,?,?>
h2 = < Sunny,?,? ?,?,?>

Any instance classified positive by h1 will be classified positive by h2, therefore h2 is

more general than h1. And it can be defined mathematically as: Let hj and hk be
Boolean-valued functions defined over X. Then hj is more-general-than-or-equal-to hk
if and only if (∀x∈X)[( hk(x)=1) -> (hj(x)=1)]

(c) Discuss support vectors in SVM.

Answer: Support vectors are the data points that lie closest to the decision boundary,
also known as the hyperplane, between the two classes in the dataset. These data
points are the most informative in determining the location and orientation of the
hyperplane, and as such, they are crucial in the construction of the SVM model.
 Figure: Support Vectors

(d) Compare Artificial Intelligence and Machine Learning.

Answer: Artificial Intelligence (AI) is a broad field that encompasses the development
of intelligent systems that can perform tasks that typically require human-like
intelligence, such as natural language processing, decision making, and perception.
Machine Learning (ML) is a subfield of AI that involves the development of algorithms
and statistical models that enable machines to automatically learn and improve their
performance on a specific task, without being explicitly programmed. The key
differences between AI and ML:

 AI is a broader field that includes many different techniques and approaches,

while ML is a specific subset of AI that uses statistical techniques to learn from
data.
 AI can involve both rule-based and learning-based systems, while ML is
exclusively based on learning from data.
 AI is a more theoretical and philosophical concept, while ML is more practical
and technical, focused on developing algorithms and models for specific tasks.

(e) Discuss reinforcement learning.

Answer: Reinforcement Learning(RL) is a type of machine learning technique that
enables an agent to learn in an interactive environment by trial and error using
feedback from its own actions and experiences. The goal of reinforcement learning is
to enable machines to learn how to make optimal decisions or take actions based on
their environment and the rewards or penalties associated with those actions.
The main components of a reinforcement learning system are: Agent, Environment,
State, Action, and Reward.
(f) Illustrate the advantages of instance-based learning techniques over other machine
learning techniques.
Answer: Instance-based learning techniques are particularly well-suited for problems
where the underlying relationships between the features and the target variable are
complex and not well understood, and where the data is noisy or has many outliers.
Instance-based learning, also known as lazy learning, is a machine learning technique
that relies on storing and classifying new data based on the similarity of the new data
to previously seen examples. Unlike other machine learning techniques such as
decision trees, support vector machines, and neural networks, instance-based
learning has several advantages that make it a popular choice for certain types of
problems.

(g) Differentiate between Gradient Descent and Stochastic Gradient Descent.

Answer: In both gradient descent (GD) and stochastic gradient descent (SGD), we
update a set of parameters in an iterative manner to minimize an error function.

Gradient descent (GD) calculates the gradients of the cost function with respect to
each parameter and updates the parameters by taking small steps in the opposite
direction of the gradient. The algorithm repeats this process until it reaches the
minimum of the cost function.
Stochastic gradient descent (SGD) computes the gradient and updates the model
parameters using a random subset of the training data at each iteration, rather than
the entire dataset. In SGD, the model parameters are updated incrementally after
each subset of training data, making the algorithm more computationally efficient
than gradient descent.

(h) Compare ANN and Bayesian network.

Answer: Key differences between ANNs and Bayesian networks are:
 Bayesian networks are well-suited for dealing with uncertainty, as they
explicitly represent the probabilistic relationships between variables. ANNs, on
the other hand, do not explicitly represent uncertainty.
 Bayesian networks are generally more interpretable than ANNs, as they
provide a clear graphical representation of the relationships between
variables. In ANN it can be difficult to understand how the model arrived at a
particular prediction.
 ANNs are trained using an iterative process called backpropagation, to adjust
weights and minimize error . Bayesian networks, are trained using a process
called maximum likelihood estimation, which estimates the parameters of the
network by maximizing the likelihood of the observed data.
 ANNs generally require a large amount of training data to perform well,
whereas Bayesian networks can often be trained on smaller datasets.

(i) Illustrate Markov decision model.

Answer: Markov decision models are a class of mathematical models that are used to
represent decision-making problems that involve sequential actions and uncertain
 outcomes. Markov decision models are based on the principles of Markov processes,
which are mathematical models of random processes that satisfy the Markov
property. Markov Decision Process(MDP) the policy is the mechanism to take
decisions. An MDP consists of a set of states, a set of actions, transition probabilities,
rewards, and a discount factor.

Markov decision models are widely used in many fields, including finance, operations
research, and artificial intelligence. They are particularly useful for modeling problems
where the outcome of an action is uncertain and where decisions need to be made
over time.

(j) Differentiate between Q learning and Deep Learning.

Answer: Q-learning and deep learning are two different machine learning techniques
that are used in different contexts.

Q-Learning Deep Learning

 Q-learning is a reinforcement  Deep learning, on the other hand, is a
learning algorithm that is used to type of machine learning that is used
learn an optimal policy for a Markov for pattern recognition and data
decision process (MDP). analysis.

 The goal of Q-learning is to find the  Deep learning involves training

optimal action to take in a given artificial neural networks with
state, by iteratively updating a table multiple layers of processing units,
of Q-values that represent the which are capable of learning
expected reward of each action in complex patterns in data.
each state.

 Q-learning is a model-free learning  Deep Learning used for its ability to

algorithm, which means it does not learn hierarchical representations of
require a model of the environment data, which can be used to extract
to learn an optimal policy. meaningful features from complex
datasets.

2. Attempt any three of the following:

(a) Explain supervised and unsupervised learning techniques.

Answer: Supervised learning and unsupervised learning are two popular techniques
used in machine learning to train and develop models that can learn and make
predictions or decisions based on data inputs.

Supervised Learning: Supervised learning is a machine learning technique in which a

model is trained on labeled data, which means that the data used to train the model
has already been labeled with the correct output. The goal of supervised learning is to
learn a mapping function that can predict the output for new inputs. The model is
 trained on a dataset that includes input variables (features) and the corresponding
output variables (labels), which are already known.
For example: Linear regression is a supervised learning technique. In linear
regression, the goal is to create a model that can predict a continuous output variable
(also known as the response or dependent variable) based on one or more input
variables (also known as the predictor or independent variables). The model is trained
on labeled data, which means that both the input variables and the corresponding
output variable are known during training, so linear regression is a supervised learning
technique. The goal is to learn a relationship between the input variables and the
output variable, so that the model can make accurate predictions for new input data.

Unsupervised learning: Unsupervised learning is a machine learning technique in

which a model is trained on unlabeled data, which means that the data used to train
the model is not labeled with the correct output. The goal of unsupervised learning is
to identify patterns or relationships in the data without any prior knowledge of the
correct output.
For example, Clustering is an unsupervised learning technique. In clustering, the goal
is to group similar data points together based on their characteristics or features,
without the use of pre-existing labels or categories. The algorithm learns the inherent
structure of the data and assigns data points to different clusters based on their
similarity to other data points. Unsupervised learning is often used for data
exploration and identifying patterns in the data, as well as for various applications
such as customer segmentation, anomaly detection, and image recognition.

(b) Discuss linear regression and logistic regression in detail.

Answer: Linear regression and logistic regression are both popular and widely used
techniques in supervised learning for regression and classification problems,
respectively.

Linear regression: Linear regression is a supervised learning technique used for

regression problems, where the goal is to create a model that can predict a continuous
output variable based on one or more input variables. The model assumes that there
is a linear relationship between the input variables and the output variable, and tries
to learn the best-fit line or hyperplane that can predict the output variable for new
input data. The simplest form of linear regression is simple linear regression, where
there is only one input variable, and the model tries to learn the best-fit line that can
predict the output variable. The equation for simple linear regression can be
represented as:

y = b0 + b1*x

where y is the output variable, x is the input variable, b 0 is the y-intercept, and b1 is
the slope of the line. The goal is to learn the values of b0 and b1 that can best fit the
data and make accurate predictions for new input data.
 Linear regression is a widely used technique in various fields, such as finance,
economics, engineering, and social sciences, and is used for various applications such
as predicting stock prices, housing prices, and sales revenue.

Linear Regression Logistic Regression

Logistic regression: Logistic regression is a supervised learning technique used for

classification problems, where the goal is to create a model that can predict a binary
output variable (0 or 1) based on one or more input variables. The model assumes that
there is a linear relationship between the input variables and the log-odds of the
output variable, and tries to learn the best-fit line or hyperplane that can predict the
log-odds for new input data. The log-odds (also known as the logit) is defined as the
natural logarithm of the odds, which is the ratio of the probability of the event
occurring to the probability of the event not occurring. The equation for logistic
regression can be represented as:

logit(p) = b0 + b1x1 + b2x2 + ... + bn*xn

where p is the probability of the output variable being 1, x1, x2, ..., xn are the input
variables, and b0, b1, b2, ..., bn are the coefficients that need to be learned.

The logistic regression model uses a sigmoid function to convert the log-odds to a
probability value between 0 and 1. The sigmoid function has an S-shaped curve and is
defined as:

p = 1 / (1 + e(-logit(p)))

where e is the base of the natural logarithm.

Logistic regression is a widely used technique in various fields, such as healthcare,

marketing, and social sciences, and is used for various applications such as predicting
the likelihood of a patient developing a disease, the likelihood of a customer buying a
product, and the likelihood of a student passing an exam.

(c) Describe the following concepts in decision tree in detail:

(i) Avoiding overfitting in decision tree.
(ii) Incorporating continuous valued attributes.
Answer:
Avoiding overfitting in decision tree.
Overfitting is a common problem in decision trees, where the model is too complex
and is tailored too closely to the training data, leading to poor performance on new,
unseen data. Overfitting occurs when the decision tree is too deep and has too many
branches, resulting in a model that is too specific to the training data and does not
generalize well to new data. To avoid overfitting in decision trees, several techniques
can be used:
 Pruning: One of the most effective techniques to avoid overfitting in decision trees
is pruning, which involves removing branches from the tree that do not improve its
accuracy. Pruning can be done in two ways: pre-pruning and post-pruning. Pre-
pruning involves setting a limit on the depth of the tree or the minimum number of
instances required to split a node, while post-pruning involves growing the tree to
its maximum depth and then removing the branches that do not improve its
accuracy.
 Regularization: Regularization is a technique used to reduce the complexity of the
decision tree by adding a penalty term to the cost function. The penalty term
discourages the model from having too many branches and encourages it to have
simpler decision rules. Two common regularization techniques used in decision
trees are L1 regularization and L2 regularization.
 Ensemble Methods: Ensemble methods like Random Forest and Gradient Boosting
can help avoid overfitting in decision trees. In these methods, multiple decision
trees are created and the final prediction is made by combining the results of all
the trees. This helps to reduce the variance of the model and improve its accuracy.
 Cross-validation: Cross-validation is a technique used to evaluate the performance
of the decision tree on new, unseen data. It involves dividing the data into k-folds,
training the model on k-1 folds, and evaluating its performance on the remaining
fold. This process is repeated k times, with each fold serving as the validation set.
Cross-validation helps to estimate the performance of the model on new data and
avoid overfitting.

By using these techniques, it is possible to build decision trees that generalize well to
new, unseen data and avoid overfitting.

Incorporating continuous valued attributes in decision trees

Incorporating continuous valued attributes in decision trees can improve the accuracy
and interpretability of the model, especially when dealing with datasets that contain
a mixture of categorical and continuous features. However, it is important to carefully
choose the appropriate method based on the nature of the data and the specific
problem being solved. In order to incorporate continuous valued attributes in decision
trees, there are several methods that can be used:
 Binning: One method is to bin the continuous values into discrete intervals or bins.
This means that the continuous attribute is divided into a finite number of intervals,
and each interval is treated as a separate categorical feature. This way, the decision
tree algorithm can make binary splits on the categorical feature to create the tree.
 Threshold-based splitting: Another method is to use threshold-based splitting,
where the algorithm selects a threshold value to divide the continuous attribute
 into two or more binary values. The threshold can be chosen based on various
criteria such as maximizing the information gain or minimizing the impurity of the
resulting subsets.
 Regression Trees: For regression tasks, another approach is to use regression trees.
Instead of making binary splits on categorical features, regression trees make splits
on continuous features based on the value of the feature. The algorithm tries to
find the best split that minimizes the sum of squared errors for the resulting
subsets.
 Hybrid Approach: A hybrid approach can also be used, where both threshold-based
splitting and regression trees are combined. In this method, the algorithm starts by
selecting the best threshold to split the continuous attribute into two or more
binary values. Then, for each resulting subset, a separate decision tree is
constructed using regression trees.

(d) Explain various types of activation functions with examples.

Answer: Activation functions are mathematical functions that are applied to the
output of a neuron in a neural network. These functions introduce non-linearity to the
model and allow it to learn complex patterns and relationships in the data. There are
several types of activation functions used in deep learning, some of the most
commonly used ones are:

 Linear function: The linear activation function, also known as the identity
function, is a simple activation function that is commonly used in neural
networks for regression problems. It is a linear transformation that scales the
input by a fixed factor and adds a constant offset, without changing the shape
of the function. The equation for the linear activation function is:
f(x) = x
where x is the input to the neuron, and f(x) is the output of the neuron. An
example of a linear function is shown below:

 Sigmoid function: The sigmoid function maps any input to a value between 0
and 1. It is often used in binary classification problems to predict the
probability of a certain class. The equation for the sigmoid function is:
f(x) = 1 / (1 + e(-x))
An example of a sigmoid function is shown below:

 ReLU (Rectified Linear Unit) function: The ReLU function sets any negative
input to zero and leaves positive inputs unchanged. It is one of the most
commonly used activation functions in deep learning models due to its
simplicity and effectiveness in reducing vanishing gradients. The equation for
the ReLU function is:
f(x) = max(0, x)
An example of a ReLU function is shown below:

 Tanh (hyperbolic tangent) function: The tanh function maps any input to a
value between -1 and 1. It is similar to the sigmoid function but has a range
 that is centered around zero, which makes it useful for normalization of data.
The equation for the tanh function is:
f(x) = (ex - e-x) / (ex + e-x)
An example of a tanh function is shown below:

 Softmax function: The softmax function is often used as the activation

function in the output layer of a neural network for multiclass classification
problems. It maps the output of each neuron to a probability distribution over
the classes. The equation for the softmax function is:
f(xi) = e(xi) / ∑(e(xj)) for j=1 to n, where n is the number of classes
An example of a softmax function is shown below:

(e) Illustrate the process of Q-learning and discuss the following terms:
(i) Q-values or action value (ii) Rewards and Episode (iii) Temporal difference or TD
update.

Answer:
(i) Q-values or action value
The Q-value, also known as the action-value, is the expected total reward that an
agent can obtain by taking a specific action a in a specific state s, and then following
the optimal policy thereafter. The Q-value function Q(s,a) represents the expected
long-term discounted reward that an agent will receive if it takes action a in state s,
and then acts optimally from that point onwards. Mathematically, the Q-value
function can be defined as:
∞

𝑄(𝑠, 𝑎) = 𝐸[∑ γ𝑡 𝑟𝑡+1 | 𝑠, 𝑎]

𝑡=0
Q(s, a) = E[∑t=0∞ γ^t r_t+1 | s, a]

where rt+1 is the reward obtained at time t+1, γ is the discount factor, and E denotes
the expected value.
The Q-value function is learned through an iterative process, where the agent explores
the environment by taking actions, and updates the Q-values based on the observed
rewards. The Q-value update rule is given by the Bellman equation:

Q(s, a) ← Q(s, a) + α [r + γ max_a' Q(s', a') - Q(s, a)]

where α is the learning rate, s' is the next state, a' is the next action, and max_a' Q(s',
a') is the maximum Q-value over all possible actions in the next state s'. The Bellman
equation represents the optimal Q-value of a state-action pair as the immediate
reward plus the maximum expected future reward.

Over time, the Q-values converge to their optimal values, which correspond to the
maximum expected cumulative reward that the agent can obtain by following the
optimal policy. Once the Q-value function is learned, the optimal policy can be
obtained by selecting the action with the highest Q-value in each state.

(ii) Rewards and Episode:

Q-learning is a type of reinforcement learning algorithm that learns by trying to find
the optimal action to take in a given state in order to maximize a reward. The two key
concepts in Q-learning are rewards and episodes.

Rewards: In Q-learning, the agent receives a reward for each action it takes in a given
state. The rewards can be positive, negative or zero, and are defined by the
environment. The goal of the agent is to maximize the total reward it receives over
the course of an episode or a sequence of actions.

Episodes: An episode is a sequence of actions taken by the agent in the environment,

starting from an initial state and ending when the agent reaches a terminal state. A
terminal state is a state in the environment where no further actions can be taken,
and the episode ends. At the end of each episode, the agent receives a final reward
that reflects the success or failure of its actions.

The Q-learning algorithm uses the rewards and episodes to update the Q-values,
which are estimates of the expected rewards for each action in a given state. The Q-
values are updated based on the rewards received and the predicted rewards of the
next state, using the Bellman equation. By learning the optimal Q-values, the agent
can choose the best action to take in each state, based on the maximum expected
reward.

(iii) Temporal difference or TD update:

In Q-learning, the temporal difference (TD) update is a method for iteratively updating
the Q-values based on the rewards received and the predicted rewards of the next
state. TD update is a key aspect of Q-learning algorithm that enables it to learn and
update Q-values online, as the agent interacts with the environment.

The TD update rule for Q-learning is given by:

 Q(s, a) <- Q(s, a) + α(r + γ * maxa' Q(s', a') - Q(s, a))

where Q(s, a) is the current Q-value for the state-action pair (s, a), α is the learning
rate, r is the immediate reward for taking action a in state s, γ is the discount factor,
s' is the next state reached after taking action a, and a' is the optimal action to take in
state s'.

The TD update rule calculates the difference between the Q-value of the current state-
action pair and the estimated optimal Q-value of the next state-action pair. This
difference is multiplied by the learning rate α and added to the current Q-value for
the state-action pair. The learning rate determines the extent to which the new
information is incorporated into the Q-value.

By repeatedly updating the Q-values based on the rewards and the predicted rewards
of the next state, the Q-learning algorithm learns the optimal Q-values for each state-
action pair, and the agent can use these Q-values to choose the best action to take in
each state.

3. Attempt any one part of the following

(a) Illustrate various areas where you can apply machine leaning.
Answer: Machine learning is a powerful tool that can be applied to a wide variety
of fields and industries. Here are some areas where machine learning is being used
today:

 Finance: Machine learning is used in the finance industry for fraud detection,
credit scoring, and algorithmic trading.
 Healthcare: Machine learning is used in healthcare for medical diagnosis,
patient monitoring, and drug discovery.
 Retail: Machine learning is used in retail for recommendation systems,
inventory management, and supply chain optimization.
 Manufacturing: Machine learning is used in manufacturing for predictive
maintenance, quality control, and supply chain optimization.
 Transportation: Machine learning is used in transportation for traffic
prediction, route optimization, and autonomous vehicles.
 Marketing: Machine learning is used in marketing for customer segmentation,
personalized advertising, and churn prediction.
 Natural Language Processing (NLP): Machine learning is used in NLP for speech
recognition, language translation, and sentiment analysis.
 Image and Video Analysis: Machine learning is used in image and video analysis
for object recognition, facial recognition, and video surveillance.
 Agriculture: Machine learning is used in agriculture for crop yield prediction,
disease detection, and pest management.
 Energy: Machine learning is used in energy for demand forecasting, energy
optimization, and predictive maintenance.
 These are just a few examples of the many areas where machine learning is being
applied today. As machine learning continues to develop, it is likely that we will
see it being used in even more areas, improving efficiency, accuracy, and decision-
making across a range of industries.

(b) Compare regression, classification and clustering in machine learning along with
suitable real-life applications.
Answer: Regression, classification, and clustering are three fundamental
techniques in machine learning, each with its own unique characteristics and
applications. Here's a brief comparison of these techniques:

Regression: Regression is a type of supervised learning that is used to predict

continuous numerical values. The goal of regression is to build a model that can
learn the relationship between the input features and the output variable, and
then use that model to make predictions on new data. Regression algorithms
include linear regression, polynomial regression, and decision trees. Regression
algorithms can be used for a variety of applications, such as:
 Predicting stock prices based on historical trends.
 Estimating the price of a house based on its features.
 Predicting the amount of rainfall in a particular region based on past data.
 Forecasting the sales of a product based on marketing spend and other
factors.

Classification: Classification is also a type of supervised learning that is used to

predict discrete categorical values. The goal of classification is to build a model
that can learn the relationship between the input features and the target class
label, and then use that model to classify new data. Classification algorithms
include logistic regression, decision trees, random forests, and support vector
machines. Classification algorithms can be used for a variety of applications, such
as:
 Identifying spam emails in your inbox.
 Detecting fraudulent credit card transactions.
 Diagnosing a medical condition based on patient symptoms.
 Classifying images based on their content.

Clustering: Clustering is a type of unsupervised learning that is used to group

similar data points together. The goal of clustering is to identify groups of data
points that share similar features, without any prior knowledge of the group
labels. Clustering algorithms include k-means clustering, hierarchical clustering,
and DBSCAN. Clustering algorithms can be used for a variety of applications, such
as:
 Customer segmentation for marketing purposes.
 Identifying groups of genes that are expressed similarly in a particular
disease.
 Grouping similar documents together for text analysis.
 Identifying groups of stars in the night sky based on their positions and
brightness
 Here are some key differences between these three techniques:

 Supervision: Regression and classification are types of supervised learning,

which means that they require labeled data to train the model. Clustering
is a type of unsupervised learning, which means that it doesn't require
labeled data.
 Output: Regression predicts a continuous numerical value, whereas
classification predicts a discrete categorical value. Clustering groups similar
data points together without any predetermined output.
 Objective: Regression and classification aim to minimize prediction error
or maximize accuracy, respectively. Clustering aims to identify similarities
and differences in the data.
 Training: Regression and classification require a training set of labeled data
to build a model. Clustering does not require labeled data and is performed
on an unlabeled dataset.
 Evaluation: Regression and classification are evaluated based on their
prediction accuracy, while clustering is evaluated based on its ability to
group similar data points together.
Overall, regression, classification, and clustering are important
techniques in machine learning, each with its own unique set of applications
and characteristics. Choosing the right technique depends on the problem at
hand, the data available, and the desired output.

4. Attempt any one part of the following:

(a) Discuss the role of Bayes theorem in machine learning. How naive Bayes algorithm
is different from Bayes theorem?

Answer: Bayes' theorem is a fundamental concept in probability theory that plays a

significant role in machine learning. It is used to compute the probability of an event
based on prior knowledge of conditions that might be related to the event. In machine
learning, Bayes' theorem is used to build probabilistic models that can be used for a
variety of applications, including classification, regression, and clustering.

The Naive Bayes algorithm is a popular machine learning algorithm that is based on
Bayes' theorem. It is called "naive" because it assumes that the features are
independent of each other, which is often not the case in real-world applications.
Despite this simplifying assumption, Naive Bayes has been found to be surprisingly
effective in many practical applications.

The Bayes theorem formula is as follows:

P(A|B) = P(B|A) * P(A) / P(B)
where P(A) is the prior probability of A, P(B|A) is the conditional probability of B given
A, and P(B) is the prior probability of B. P(A|B) is the posterior probability of A given
B, which is the quantity we want to compute.
The Naive Bayes algorithm is used for classification problems, where we want to
predict the class of a given input based on its features. For example, we might want
to predict whether an email is spam or not based on its content.

To use Naive Bayes for classification, we first need to calculate the prior probabilities
and the conditional probabilities for each feature given the class. We then use Bayes'
theorem to calculate the posterior probability of each class given the features. The
class with the highest probability is the predicted class for the input.

The Naive Bayes algorithm differs from Bayes' theorem in that it makes the
assumption that the features are independent of each other. This assumption greatly
simplifies the calculations and makes the algorithm computationally efficient.
However, in practice, features are often correlated, and this assumption may not hold.
Despite this limitation, Naive Bayes has been found to be surprisingly effective in many
practical applications, such as text classification and spam filtering.

In summary, Bayes' theorem is a fundamental concept in probability theory that is

used in machine learning to build probabilistic models for a variety of applications.
The Naive Bayes algorithm is a popular machine learning algorithm that is based on
Bayes' theorem and is used for classification problems. The Naive Bayes algorithm
simplifies the calculations by assuming feature independence, which can be a
limitation in some cases, but still works well in many practical applications.

(b) Explain hyperplane (decision boundary) in SVM. Categorize various popular

kernels associated with SVM.

Answer: In Support Vector Machines (SVM), a hyperplane is a decision boundary that

separates the data points into different classes. SVM finds the optimal hyperplane that
maximally separates the classes. In other words, SVM tries to find the hyperplane that
has the largest margin between the two classes. The margin is the distance between
the hyperplane and the closest data points from each class.

The hyperplane in SVM can be defined by the equation:

w*x+b=0

where w is the normal vector to the hyperplane, x is the input vector, and b is the bias
term. The sign of the output of this equation determines which side of the hyperplane
the input vector is on.

SVM can use different types of kernels to find the hyperplane. A kernel is a function
that takes two input vectors and produces a scalar value that measures the similarity
between them. SVM uses the kernel function to map the input vectors into a higher-
dimensional feature space, where it becomes easier to find a hyperplane that
separates the classes.
 Some popular kernels used in SVM are:

 Linear kernel: The linear kernel is the simplest kernel and works well for
linearly separable data. It maps the input vectors to a higher-dimensional
space without introducing any new features.
 Polynomial kernel: The polynomial kernel maps the input vectors to a higher-
dimensional space using a polynomial function. It works well for data that has
some non-linearity.
 Radial basis function (RBF) kernel: The RBF kernel maps the input vectors to an
infinite-dimensional space using a Gaussian function. It works well for data
that is not linearly separable and has a complex boundary.
 Sigmoid kernel: The sigmoid kernel maps the input vectors to a higher-
dimensional space using a sigmoid function. It works well for data that has non-
linearly separable classes.

In summary, the hyperplane in SVM is a decision boundary that separates the data
points into different classes. SVM uses various types of kernels to map the input
vectors into a higher-dimensional space, where it becomes easier to find a hyperplane
that separates the classes. The linear, polynomial, RBF, and sigmoid kernels are some
of the popular kernels used in SVM.

5. Attempt any one part of the following:

(a) Demonstrate K-Nearest Neighbors algorithm for classification with the help of an
example.
Answer: The K-Nearest Neighbors (KNN) algorithm is a simple yet powerful machine
learning algorithm used for classification and regression tasks. In KNN, the idea is to
classify a new data point based on the classification of its nearest neighbors in the
feature space. The algorithm works as follows:
 Choose the value of K: The K in KNN represents the number of nearest
neighbors considered to make the prediction. It is a hyperparameter that we
have to choose based on the data we have. A common approach is to use odd
numbers for K to avoid ties.
 Compute the distances: For a given new data point, compute the distance from
that point to each data point in the training dataset. Euclidean distance is a
popular choice for the distance metric, but there are other metrics such as
Manhattan distance and cosine similarity that can be used.
 Identify the K nearest neighbors: Select the K data points in the training dataset
that are closest to the new data point based on the distance metric.
 Assign a class to the new data point: In the classification problem, the new data
point is assigned the class that is most frequent among its K nearest neighbors.
 Test and refine: Finally, we test the KNN model on a test dataset, and based
on the results, we can fine-tune the value of K or use different distance metrics
to improve the performance of the algorithm.
 For example, In the figure shown below, suppose a new data point arrive in the
dataset, and we have to classify this data point among the two classes A or B. If
K=3 and 1 of the nearest neighbors belong to class A, and 2 belong to class B, the
new point is classified as class B. In regression, the new data point's predicted
value is the average of the K nearest neighbors' values.

(b) Explain Instance based learning. Compare locally weighted regression and radial
basis function networks.
Answer: Instance-based learning is a type of machine learning where the model is
built on the basis of memorizing the training data instead of learning a set of
parameters from the data. In instance-based learning, the model stores the training
examples and uses them directly to make predictions on new examples. These training
examples are also called instances or examples.

The core idea behind instance-based learning is that the data points are represented
in a feature space and similar data points tend to have similar labels. Given a new
instance, the algorithm computes the similarity between the new instance and each
training instance, and then predicts the label of the new instance based on the most
similar training instances.

One popular instance-based learning algorithm is the K-Nearest Neighbors (KNN)

algorithm. KNN algorithm stores all the training instances in a multidimensional
feature space and uses the Euclidean distance as a measure of similarity between the
instances. For a new instance, KNN algorithm finds the K closest training instances in
the feature space and then makes a prediction based on the majority label of those K
instances.

Instance-based learning is useful when there is no clear pattern in the data, and the
data is difficult to summarize with a set of fixed parameters. It is also useful when the
data is not stationary and is expected to change over time. However, instance-based
learning can be computationally expensive when the training set is large and can also
suffer from overfitting if the data is noisy or if the number of features is high compared
to the number of instances.
 Locally Weighted Regression (LWR) and Radial Basis Function Networks (RBFN) are
two popular machine learning algorithms that are commonly used for non-linear
regression tasks. Although both algorithms share some similarities, they differ in some
key aspects.

 Basis functions: RBFN uses a fixed set of basis functions, usually Gaussian or
sigmoid functions, to transform the input features into a higher-dimensional
space. In contrast, LWR uses a weighted sum of the training instances
themselves as basis functions, with the weights depending on the distance of
the instance from the new data point.
 Model structure: RBFN typically has a three-layer structure with an input layer,
a hidden layer of basis functions, and an output layer, while LWR doesn't have
any hidden layers and uses the weighted sum of the training instances directly
as the output.
 Training procedure: RBFN requires a training procedure to learn the
parameters of the basis functions, such as the width of the Gaussian functions
or the weights of the sigmoid functions. In contrast, LWR doesn't have any
parameters to learn, and the model simply selects the training instances based
on their distance to the new data point.
 Computation complexity: RBFN requires more computation to train than LWR,
as it requires optimization of the basis function parameters. However, once
the model is trained, RBFN can make predictions more efficiently than LWR, as
it doesn't require a weighted sum over all the training instances.
 Performance: RBFN can perform well on a wide range of problems, but it can
suffer from overfitting if the number of basis functions is too large. LWR, on
the other hand, is less prone to overfitting but can be affected by noisy or
irrelevant features.

In summary, LWR and RBFN are two different algorithms that have different strengths
and weaknesses. LWR is more suitable when the data is non-stationary and there are
no clear patterns, and RBFN is more suitable when the data has clear non-linear
relationships and the number of basis functions can be optimized.

6. Attempt any one part of the following:

(a) Explain the different layers used in convolutional neural network with suitable
examples.
Answer: Convolutional Neural Networks (CNNs) are a type of deep neural network
that are commonly used for image recognition and classification tasks. CNNs
consist of several layers, each with a specific purpose in processing the input data.
The main layers used in a typical CNN architecture are:
 Convolutional layer: The convolutional layer is the main building block of a
CNN. This layer performs a convolution operation on the input image using a
set of learnable filters or kernels. Each filter is a small matrix that slides over
 the input image, computing the dot product between the filter and the
corresponding patch of the image. The result of this operation is a set of
feature maps that capture different aspects of the input image.

Example: Suppose we have an input image of size 32x32x3 (width, height, and
channels), and we apply 32 filters of size 3x3x3 to the image. The resulting
output feature map will have a size of 30x30x32, which means we have 32
feature maps, each of size 30x30.

 Pooling layer: The pooling layer is used to downsample the output feature
maps from the convolutional layer. It works by dividing the feature map into
non-overlapping regions and then computing a summary statistic, such as the
maximum or average value, for each region. This reduces the dimensionality
of the feature maps while preserving the important information.

Example: Suppose we apply a max-pooling operation of size 2x2 with a stride

of 2 to the output feature map of size 30x30x32. The resulting output feature
map will have a size of 15x15x32.

 Activation layer: The activation layer applies a non-linear function to the

output of the convolutional or pooling layer, introducing non-linearity to the
model. The most common activation function used in CNNs is the Rectified
Linear Unit (ReLU), which sets negative values to zero and leaves positive
values unchanged.

Example: Suppose we apply ReLU activation to the output feature map of size
15x15x32. The resulting output feature map will have the same size but with
all negative values set to zero.

 Fully connected layer: The fully connected layer is a traditional neural

network layer that connects every neuron in the layer to every neuron in the
previous layer. It is used to combine the information from the previous layers
and make a final prediction.

Example: Suppose we flatten the output feature map from the previous layer
into a vector of size 7,200, and then apply a fully connected layer with 512
neurons and a ReLU activation function. The resulting output will be a vector
of size 512, which can be further processed by other layers for making the
final prediction.

In summary, CNNs consist of multiple layers that work together to learn the
important features of an input image and make a final prediction. The
convolutional layer extracts local features, the pooling layer reduces
dimensionality, the activation layer adds non-linearity, and the fully
connected layer combines information for making the final prediction.
(b) Illustrate backpropagation algorithm by assuming the training rules for output unit
weights and Hidden Unit weights.
Answer: Backpropagation is a widely used algorithm for training neural networks.
It works by propagating the error backwards from the output layer to the input
layer, adjusting the weights along the way to minimize the error. Here's an
illustration of the backpropagation algorithm, assuming the following training
rules for the output unit weights and hidden unit weights:

Training rules for output unit weights:

Update the weight between the kth hidden unit and the output unit j as follows:
wjk = wjk - α * δj * ak

Training rules for hidden unit weights:

Update the weight between the i-th input unit and the jth hidden unit as follows:
wji = wji - α * δj * xi

 Forward Pass:
First, we perform a forward pass through the neural network to compute
the output of each unit. Given an input vector x, the output of the jth hidden
unit hj is computed as follows:
hj = f(∑(wji * xi) + bj)

where f is the activation function, wji is the weight between the ith input
unit and the jth hidden unit, xi is the ith input feature, and bj is the bias term
for the jth hidden unit.

Similarly, the output of the k-th output unit yk is computed as follows:

y_k = g(∑(w_kj * h_j) + b_k)

where g is the activation function for the output unit, wkj is the weight
between the jth hidden unit and the kth output unit, and bk is the bias term
for the kth output unit.

 Backward Pass:
Next, we compute the error at the output layer and propagate it backwards
to the hidden layer. Let tk be the target output for the kth output unit, and
ek be the error at the output layer:
ek = tk - yk

We can use the error to compute the delta value for the kth output unit:
δk = g'(∑(wkj * hj) + bk) * ek
where g' is the derivative of the activation function for the output unit.
 Using the delta value for the output unit, we can update the weights
between the hidden layer and the output layer using the training rules for
the output unit weights.
Next, we compute the delta values for the hidden layer. Let δj be the delta
value for the j-th hidden unit:

δj = f'(∑(wji * xi) + bj) * ∑(wkj * δk)

where f' is the derivative of the activation function for the hidden unit, and
the sum is taken over all the output units that are connected to the jth
hidden unit.

Using the delta value for the hidden unit, we can update the weights
between the input layer and the hidden layer using the training rules for
the hidden unit weights.

 Repeat:
We repeat the forward pass and backward pass for each training example
in the dataset, adjusting the weights after each example, until the error on
the training set is minimized.

This completes the illustration of the backpropagation algorithm. The

training rules for the output unit weights and hidden unit weights
determine how the weights are updated during the backward pass,
allowing the neural network to learn the optimal weights for making
accurate predictions.

7. Attempt any one part of the following:

(a) Explain various types of reinforcement learning techniques with suitable

examples.
Answer: Reinforcement learning (RL) is a type of machine learning where an agent
learns to make decisions by interacting with an environment and receiving feedback
in the form of rewards or punishments. There are several types of RL techniques,
including:

 Q-learning: Q-learning is a value-based RL technique that learns an optimal action-

value function, also known as a Q-function. The Q-function estimates the
expected cumulative reward of taking a particular action in a particular state, and
the optimal policy is derived from the Q-function. Q-learning is widely used in
problems with discrete action spaces, such as game playing or control of discrete-
event systems.
 Deep Q-Networks (DQNs): DQNs are an extension of Q-learning that use deep
neural networks to approximate the Q-function. DQNs can handle high-
dimensional state spaces and have been used successfully in tasks such as Atari
game playing.
 Policy gradients: Policy gradient methods learn a parameterized policy that
directly maps states to actions. The policy is updated using gradient ascent on the
expected cumulative reward, which is estimated by running multiple trajectories
from the current policy. Policy gradient methods are useful in continuous action
spaces, where it is difficult to use discrete actions.
 Actor-critic methods: Actor-critic methods combine value-based and policy-based
methods by learning both a value function and a policy. The value function is used
to estimate the expected future reward, and the policy is updated using the
estimated value function. Actor-critic methods are particularly effective in
problems with high-dimensional state and action spaces.
 Model-based methods: Model-based methods learn a model of the environment
dynamics and use the model to make decisions. Model-based methods can be
more sample-efficient than model-free methods, but they require more
computational resources to learn and maintain the model. Model-based methods
are particularly useful in robotics and control applications.
 Hierarchical RL: Hierarchical RL uses a multi-level decision-making process, where
high-level policies control lower-level policies. This approach allows for more
efficient learning and decision-making in complex tasks, such as robotics or game
playing.

Each of these RL techniques has its own strengths and weaknesses, and the choice of
which technique to use depends on the specific problem domain and the available
data. For example, Q-learning may be appropriate for game playing tasks with discrete
actions, while policy gradients may be more suitable for continuous control tasks.

(b) How to Identify the reproduction cycle of genetic algorithm? Explain with suitable
example.
Answer: In genetic algorithms, the reproduction cycle refers to the process of
generating new individuals (offspring) from the existing population by applying
genetic operators such as crossover and mutation. The reproduction cycle typically
follows a set of steps, as described below:

 Selection: A subset of individuals from the population is selected based on

their fitness, where fitness is a measure of how well an individual solves the
problem at hand. The most common selection methods are tournament
selection, roulette wheel selection, and rank-based selection.
 Crossover: Two or more selected individuals are combined to create new
offspring. Crossover involves swapping parts of the parents' chromosomes to
create new chromosomes for the offspring.
 Mutation: A random mutation is applied to the offspring's chromosomes to
introduce diversity in the population. Mutation involves randomly changing
one or more bits in the chromosome.
 Replacement: The least fit individuals in the population are replaced with the
new offspring.
This process is repeated for a specified number of generations, or until a stopping
criterion is met, such as reaching a maximum fitness or a certain number of
generations.

For example, consider a genetic algorithm used to optimize a function f(x), where x is
a vector of parameters. The reproduction cycle would involve generating new vectors
of parameters by combining and mutating existing ones, and evaluating their fitness
by computing f(x) for each individual. The selection process would select the most fit
individuals based on their f(x) values, and the crossover and mutation steps would
create new vectors of parameters that are similar to the selected individuals, but with
some random variation. The replacement step would replace the least fit individuals
with the new offspring, and the process would repeat for a specified number of
generations, or until a stopping criterion is met. Over time, the population would
evolve to contain individuals that are increasingly fit for the problem at hand.