MODULE 1 – (Basic notes)

Basics of Machine Learning (ML)

Machine Learning (ML) stands as a pivotal subfield within Artificial Intelligence

(AI), fundamentally concerned with equipping computer systems with the ability
to "learn" from data. Unlike traditional programming, where every step and rule
is explicitly coded, ML empowers machines to automatically identify intricate
patterns, make informed predictions, and adapt their behavior progressively,
without being given explicit instructions for every possible scenario.

The core essence of ML lies in its iterative process: an ML algorithm receives a

set of "training data," which is a collection of examples relevant to the problem
at hand. From this data, the algorithm constructs a "model." This model is
essentially a mathematical representation or set of rules derived from the
training data, designed to generalize beyond the observed examples. Once
trained, this model can then be applied to new, unseen data to generate
predictions, classify information, or make decisions. The remarkable aspect of
ML is its capacity for self-improvement; as more data becomes available or as
the model encounters new situations, its performance on a specific task can
continually refine and enhance over time, leading to increasingly accurate and
insightful outcomes.

Types of Machine Learning Systems

Machine learning methodologies are broadly categorized into distinct paradigms,

primarily differentiated by the nature of the training data and the type of
feedback the learning system receives during the learning process.

1. Supervised Learning:
o Concept: Supervised learning is the most prevalent and arguably
the most intuitive form of machine learning. It operates on the
principle of learning from "labeled data." A labeled dataset is a
collection of input-output pairs, where each input (e.g., features of
an object) is explicitly associated with its corresponding correct
output (e.g., the object's category or value). The algorithm's
primary objective is to learn a mapping function or a set of rules
that accurately transforms given inputs into their correct outputs.
This learned mapping can then be used to predict the outputs for
entirely new, unobserved inputs.
o Process: During the training phase, the supervised learning model
processes the labeled data. It makes a prediction, compares that
 prediction against the known "ground truth" label, and calculates an
error. This error signal is then used to iteratively adjust the
model's internal parameters (weights and biases) through
optimization algorithms (like gradient descent) to minimize the
discrepancy between its predictions and the actual labels. This
iterative refinement allows the model to generalize well to future
data.
o Common Tasks: Supervised learning encompasses two primary
types of tasks:
 Classification: Predicting a categorical label (e.g., identifying
whether an email is "spam" or "not spam").
 Regression: Predicting a continuous numerical value (e.g.,
forecasting the price of a house).
o Examples of Algorithms: Linear Regression, Logistic Regression,
Support Vector Machines (SVMs), Decision Trees, Random Forests,
K-Nearest Neighbors (KNN), and Neural Networks (for both
classification and regression).
2. Unsupervised Learning:
o Concept: Unsupervised learning distinguishes itself by working with
"unlabeled data." In this paradigm, the training data consists solely
of input features, with no corresponding output labels provided.
The algorithm's fundamental goal is to autonomously discover
inherent patterns, underlying structures, hidden relationships, or
natural groupings within the data. It seeks to make sense of data
where explicit guidance is absent.
o Process: Unlike supervised methods, there's no error signal based
on correct answers. Instead, unsupervised algorithms employ
statistical and mathematical techniques to identify similarities,
differences, and latent variables. They aim to simplify complex
data, making it more digestible and revealing insights that might
not be immediately apparent.
o Common Tasks:
 Clustering: Grouping similar data points together based on
their intrinsic characteristics (e.g., customer segmentation).
 Dimensionality Reduction: Reducing the number of features
or variables while retaining most of the important
information (e.g., for visualization or noise reduction).
 Association Rule Mining: Discovering relationships between
variables in large datasets (e.g., "customers who buy bread
also tend to buy milk").
 o Examples of Algorithms: K-Means, Hierarchical Clustering,
Principal Component Analysis (PCA), Independent Component
Analysis (ICA), Autoencoders.
3. Reinforcement Learning (RL):
o Concept: Reinforcement Learning is a behavioral learning paradigm
inspired by psychology. It concerns how an intelligent "agent" (e.g.,
a software program, a robot) should make a sequence of decisions
or take actions in a dynamic "environment" to maximize a
cumulative "reward" over time. The agent learns through direct
interaction with its environment.
o Process: The agent performs an action, and the environment
responds, providing a new state and a scalar reward (or penalty).
The agent's goal is to learn a "policy" – a strategy that maps states
to actions – which maximizes its total expected reward over the
long run. There's no labeled dataset; instead, learning occurs
through trial-and-error, exploring different actions and observing
their consequences.
o Key Components:
 Agent: The learner or decision-maker.
 Environment: Everything outside the agent that it interacts
with.
 State: The current situation of the environment.
 Action: What the agent can do in a given state.
 Reward: A numerical signal that indicates how good or bad an
action was.
 Policy: The agent's strategy for choosing actions based on
states.
o Common Applications: Robotics (learning to walk), game playing (e.g.,
AlphaGo beating Go champions, DeepMind's agents mastering Atari
games), autonomous driving, resource management, personalized
recommendations.

Challenges in ML

While immensely powerful, the implementation and deployment of ML systems

are fraught with significant challenges:

 Data Quality and Quantity:

o Insufficient Data: Many ML algorithms, especially deep learning
models, require vast amounts of data to learn effectively and
generalize well. Scarcity of data can lead to poor model
performance.
 o Noisy Data: Irrelevant or incorrect data can mislead the learning
algorithm, resulting in erroneous models. Data cleaning and
preprocessing are critical.
o Biased Data: If the training data is biased (e.g., unrepresentative
of the real-world population or containing societal prejudices), the
model will learn and perpetuate these biases, leading to unfair or
discriminatory outcomes.
o Irrelevant Features: Including too many irrelevant features can
obscure the true signal in the data, increase computational cost,
and sometimes lead to overfitting.
 Overfitting vs. Underfitting: This is a fundamental trade-off in model
complexity.
o Overfitting: Occurs when a model learns the training data too well,
memorizing noise and specific examples rather than capturing the
underlying general patterns. Such a model performs exceptionally
on the training set but poorly on new, unseen data. It's like a
student who memorizes answers but doesn't understand the
concepts. Techniques to mitigate include regularization, cross-
validation, more training data, and simplifying the model.
o Underfitting: Occurs when a model is too simplistic to capture the
essential patterns within the training data. It performs poorly on
both the training set and new data. This often happens if the model
lacks sufficient complexity or is trained for too short a time. It's
like a student who hasn't grasped the basic concepts. Solutions
involve using a more complex model, adding more features, or
reducing regularization.
 Feature Engineering:
o Definition: This is the process of transforming raw data into
features that better represent the underlying problem to the
predictive models, thereby improving model accuracy on unseen
data. It often involves creating new features, selecting important
existing features, or transforming them (e.g., normalization, scaling,
one-hot encoding).
o Complexity: It frequently requires deep domain expertise and can
be the most time-consuming and labor-intensive part of the ML
workflow. The quality of features often has a greater impact on
model performance than the choice of algorithm itself.
 Model Interpretability/Explainability (XAI):
o The "Black Box" Problem: Many powerful ML models, especially
deep neural networks, operate as "black boxes"—it's difficult to
understand why they make a particular prediction or decision. This
 lack of transparency can be problematic in critical domains like
healthcare (diagnosis), finance (loan approvals), or legal systems,
where accountability and trustworthiness are paramount.
o Need for XAI: The emerging field of Explainable AI (XAI) aims to
develop techniques that make ML models more transparent and
interpretable, allowing humans to understand, trust, and
effectively manage AI systems.
 Computational Resources:
o Training Demands: Training large-scale ML models, particularly
deep learning architectures with millions or billions of parameters,
demands substantial computational power. This often necessitates
specialized hardware like Graphics Processing Units (GPUs) or
Tensor Processing Units (TPUs), large amounts of RAM, and
distributed computing setups.
o Cost Implications: Acquiring and maintaining these resources can
be very expensive, limiting accessibility for some researchers and
organizations.
 Ethical Concerns:
o Bias and Fairness: If training data reflects existing societal
biases (e.g., racial, gender, socioeconomic), the ML model will learn
and amplify these biases, leading to unfair or discriminatory
outcomes when applied in real-world scenarios (e.g., biased hiring
algorithms, facial recognition systems).
o Privacy: ML models often require access to sensitive personal data.
Ensuring data privacy and compliance with regulations (like GDPR) is
a significant challenge, especially when sharing or using data across
different entities.
o Security: ML models can be vulnerable to adversarial attacks,
where subtly perturbed inputs can cause the model to make
incorrect predictions.
o Accountability: Determining who is responsible when an autonomous
ML system makes a harmful error.

Supervised Learning Model Example: Regression Models

 Purpose and Definition: Regression models are a class of supervised

learning algorithms specifically designed to predict a continuous numerical
output variable based on one or more input features. Unlike classification,
which predicts discrete categories, regression aims to estimate a real-
valued output. The objective is to find a function that best describes the
relationship between the independent variables (features) and the
dependent variable (the target value).
  Mathematical Concept: At its simplest, linear regression tries to fit a
straight line (or hyperplane in higher dimensions) to the data that
minimizes the sum of squared errors between the predicted values and
the actual values. More complex regression models can fit non-linear
relationships.
 Common Applications:
o Financial Forecasting: Predicting stock prices, currency exchange
rates, or housing market trends.
o Healthcare: Predicting blood pressure based on age and diet, or
predicting hospital readmission rates.
o Environmental Science: Predicting temperature, rainfall, or
pollution levels.
o Business: Predicting sales figures, customer lifetime value, or call
center wait times.
 Example: Predicting the selling price of a house.
o Input Features: Square footage, number of bedrooms, number of
bathrooms, age of the house, location (e.g., zip code), presence of a
garden, proximity to schools, etc.
o Output Variable: The continuous numerical value of the house price.
o A regression model would learn from historical sales data how
these features influence the price. When given details of a new
house, it would output a specific dollar value as its predicted price.

Classification Model Example: Logistic Regression

 Purpose and Definition: Classification models are supervised learning

algorithms used when the output variable is categorical (or discrete),
meaning it belongs to a finite set of classes or categories. The goal is to
predict which category an input instance belongs to.
 Logistic Regression: Despite having "regression" in its name, Logistic
Regression is a fundamental and widely used algorithm for binary
classification tasks (predicting one of two classes, often represented as
0 or 1). It's called "regression" because it calculates a "score" which is
then transformed into a probability.
 How it works:
1. It first computes a linear combination of the input features and
their corresponding weights (similar to linear regression).
2. This linear score is then passed through a sigmoid (or logistic)
activation function. The sigmoid function maps any real-valued
number into a value between 0 and 1, which can be interpreted as
the probability of the input belonging to the "positive" class (e.g.,
Class 1).
 3. A threshold (commonly 0.5) is then applied to this probability: if
the probability is above the threshold, the instance is classified
into one category; otherwise, it's classified into the other.
 Common Applications:

o Spam Detection: Classifying an email as "spam" or "not spam."

o Medical Diagnosis: Predicting if a patient has a "disease" or "no
disease" based on symptoms and test results.
o Customer Churn Prediction: Determining if a customer will "churn"
(cancel service) or "not churn."
o Fraud Detection: Identifying if a transaction is "fraudulent" or
"legitimate."
 Example: Classifying an email as spam or not spam.
o Input Features: Presence of certain keywords ("buy now," "free"),
sender's email address, number of exclamation marks, email length,
etc.
o Output Variable: A binary category ("spam" or "not spam").
o A logistic regression model would learn from a dataset of labeled
emails. For a new email, it would calculate the probability of it
being spam. If the probability is, say, 0.95, it would classify it as
spam.

Unsupervised Model Example: K-Means Clustering

 Purpose and Definition: K-Means is one of the simplest and most popular
unsupervised learning algorithms used for clustering. Its primary goal is
to partition 'n' data points into 'k' distinct, non-overlapping clusters. The
algorithm aims to make the data points within each cluster as similar as
possible (high intra-cluster similarity) while making clusters as distinct as
possible from each other (low inter-cluster similarity).
 How it works (Iterative Process):
1. Initialization: The algorithm begins by randomly selecting 'k' data
points from the dataset to serve as the initial "centroids" for each
cluster. 'k' is a pre-defined number of clusters that the user needs
to specify.
2. Assignment Step (E-step - Expectation): Each data point in the
dataset is assigned to the cluster whose centroid is closest to it.
Closeness is typically measured using Euclidean distance, but other
distance metrics can be used.
3. Update Step (M-step - Maximization): After all data points have
been assigned, the centroids of the clusters are recalculated. The
new centroid for each cluster is the mean (average) of all the data
 points currently assigned to that cluster. This moves the centroids
to the center of their respective assigned data points.
4. Iteration: Steps 2 and 3 are repeated iteratively. The algorithm
continues to assign data points and update centroids until either:
 The cluster assignments no longer change.
 The centroids no longer move significantly.
 A maximum number of iterations is reached.
 Advantages: Relatively simple to understand and implement,
computationally efficient for large datasets.
 Disadvantages: Requires specifying 'k' beforehand, sensitive to initial
centroid placement, struggles with non-globular clusters, sensitive to
outliers.
 Common Applications:

o Customer Segmentation: Grouping customers with similar

purchasing behaviors or demographics for targeted marketing.
o Document Clustering: Organizing large collections of text
documents into topics.
o Image Segmentation: Dividing an image into regions based on color
or texture.
o Anomaly Detection: Identifying unusual data points that don't fit
into any cluster.
 Example: Customer segmentation for an e-commerce platform.
o Input Data: Customer transaction history (e.g., frequency of
purchases, average spending, types of products bought).
o Goal: To group customers into, say, 3 distinct segments (e.g., "high-
value frequent buyers," "occasional discount shoppers,"
"new/inactive users").
o K-Means would identify these groupings based on the patterns in
their behavior without being told what these groups should be in
advance.

Artificial Neural Network (ANN)

 Origin and Inspiration: Artificial Neural Networks (ANNs), often simply

called Neural Networks (NNs), are computational models fundamentally
inspired by the structure and functioning of biological neural networks,
particularly the human brain. They aim to mimic the way biological neurons
communicate and process information.
 Fundamental Unit: The Neuron (or Perceptron): The basic building block
of an ANN is the "artificial neuron." Each neuron receives one or more
 inputs, processes them (typically by summing weighted inputs and passing
them through an activation function), and produces an output.
 Structure: ANNs are typically organized into layers of interconnected
neurons:
o Input Layer: This layer receives the raw input data. Each neuron in
this layer corresponds to an input feature.
o Hidden Layers: These are intermediate layers between the input
and output layers. They perform most of the complex computations
and feature transformations. An ANN can have one or many hidden
layers.
o Output Layer: This layer produces the final output of the network
(e.g., a prediction, a classification).
 Connections and Weights: Neurons in one layer are connected to neurons
in subsequent layers. Each connection has a numerical "weight" associated
with it, representing the strength or importance of that connection.
During the learning process, these weights are adjusted.
 Activation Functions: Each neuron (except input layer neurons) applies an
"activation function" to its weighted sum of inputs. Activation functions
introduce non-linearity into the network, enabling it to learn complex,
non-linear relationships in the data. Common activation functions include
ReLU (Rectified Linear Unit), Sigmoid, and Tanh.
 Learning Process: ANNs learn by iteratively adjusting their weights and
biases (parameters) based on the training data. The process typically
involves:

1. Forward Propagation: Input data passes through the network,

layer by layer, producing an output.
2. Loss Calculation: The network's output is compared to the true
target, and a "loss" (error) is calculated.
3. Back propagation: The calculated loss is propagated backward
through the network. This involves computing the gradient of the
loss with respect to each weight and bias in the network.
4. Parameter Update: An optimization algorithm (e.g., Gradient
Descent, Adam) uses these gradients to adjust the weights and
biases in a direction that reduces the loss. This cycle repeats for
many "epochs" (passes over the entire training dataset).

Perceptron

 Historical Significance: The Perceptron, introduced by Frank Rosenblatt

in 1957, is a foundational algorithm in the history of neural networks and
the simplest form of an ANN. It represents a single artificial neuron.
  Architecture: A Perceptron consists of:
o Inputs: It receives multiple numerical inputs.
o Weights: Each input is multiplied by a corresponding weight.
o Summation Function: The weighted inputs are summed together,
often with an added bias term.
o Activation Function: The sum is then passed through a step
function (or threshold function). If the sum exceeds a certain
threshold, the perceptron outputs 1; otherwise, it outputs 0 (or -1).
 Functionality: Perceptrons are designed for binary classification. They
learn a linear decision boundary to separate data points into two classes.
 Learning Rule (Perceptron Learning Rule): The weights of the
perceptron are adjusted iteratively. If the perceptron misclassifies a
data point, its weights are updated to reduce the error. This rule
guarantees convergence if the data is linearly separable.
 Limitations: The most significant limitation of a single perceptron is its
inability to solve non-linearly separable problems. For example, it cannot
learn the XOR logical function, which requires a non-linear decision
boundary. This limitation led to a period known as the "AI winter" for
neural networks until the development of multi-layer networks and the
backpropagation algorithm.

Universal Approximation Theorem (Statement Only)

 Core Statement: The Universal Approximation Theorem is a profound

theoretical result in the field of neural networks. It states that a
feedforward neural network with a single hidden layer, containing a finite
number of neurons (assuming appropriate non-linear activation functions
like sigmoid, tanh, or ReLU are used in the hidden layer), is capable of
approximating any continuous function to an arbitrary degree of accuracy,
provided the input space is a compact subset of real numbers.
 Implications:
o Theoretical Power: This theorem provides a strong theoretical
foundation for the power and versatility of ANNs. It implies that
ANNs are "universal function approximators," meaning they can
learn and represent virtually any complex relationship between
inputs and outputs, no matter how intricate, given enough neurons
in the hidden layer and sufficient training.
o Justification for Hidden Layers: It highlights the critical role of
hidden layers and non-linear activation functions. Without non-
linearity, a multi-layer network would simply behave like a single-
layer network, only capable of learning linear relationships.
 o What it DOESN'T say: It does not specify how many neurons are
needed in the hidden layer, nor does it provide a method for
learning the optimal weights. It only guarantees that such a
network exists. Finding the right architecture and training it
effectively remains a practical challenge.

Multi-Layer Perceptron (MLP)

 Definition: A Multi-Layer Perceptron (MLP) is a type of feedforward

Artificial Neural Network that overcomes the limitations of the single
perceptron by incorporating one or more "hidden layers" between the
input and output layers. Each neuron in the hidden and output layers uses
a non-linear activation function.
 Architecture:
o Input Layer: Receives the initial data.
o One or More Hidden Layers: These layers are where the network
learns complex representations of the input data. Each neuron in a
hidden layer is fully connected to all neurons in the previous layer.
The non-linear activation functions in these layers enable the MLP
to learn and model highly non-linear relationships.
o Output Layer: Produces the final result, which can be a continuous
value (for regression) or probabilities/scores for different classes
(for classification).
 Learning Mechanism (Backpropagation): MLPs are typically trained using
the backpropagation algorithm. Backpropagation is an efficient method
for computing the gradient of the loss function with respect to the
network's weights and biases. It works by calculating the error at the
output layer and then propagating this error backward through the
hidden layers, adjusting the weights at each layer to minimize the overall
error. This allows MLPs to learn from their mistakes and improve their
performance iteratively.
 Capabilities: With the addition of hidden layers and non-linear activation
functions, MLPs are universal function approximators and can learn
arbitrary complex decision boundaries, making them suitable for a wide
range of tasks where single perceptrons fail (e.g., XOR problem).

Deep Neural Network (DNN)

 Definition: A Deep Neural Network (DNN) is essentially a Multi-Layer

Perceptron with multiple hidden layers. The term "deep" in deep learning
directly refers to the depth of the network architecture—i.e., having
 more than one hidden layer, often many tens or even hundreds of hidden
layers.
 Hierarchical Feature Learning: The key power of DNNs lies in their
ability to learn hierarchical representations of data. Each successive
hidden layer learns increasingly abstract and complex features from the
output of the previous layer.
o For example, in image recognition, the first hidden layer might
learn to detect simple edges. The second layer might combine
these edges to recognize basic shapes or textures. Subsequent
layers might combine these shapes to detect parts of objects (e.g.,
eyes, noses, wheels), and the final layers might assemble these
parts into complete objects (e.g., faces, cars). This automatic
feature extraction is a significant advantage over traditional ML,
where feature engineering is manual.
 Increased Complexity and Capacity: More layers and neurons allow
DNNs to model extremely intricate and high-dimensional relationships
within data.
 Common Architectures: While MLP is a foundational DNN, many
specialized DNN architectures have emerged for specific data types and
tasks:
o Convolutional Neural Networks (CNNs): Excellent for image and
video processing.
o Recurrent Neural Networks (RNNs): Suitable for sequential data
like text and time series.
o Transformers: State-of-the-art for natural language processing.
 Computational Requirements: The training of DNNs often requires
massive datasets and significant computational resources (GPUs, TPUs)
due to the large number of parameters involved.

Demonstration of Regression and Classification Problems using MLP

Multi-Layer Perceptrons are highly versatile and can be effectively configured

to tackle both regression and classification problems by making specific choices
for their output layer and the loss function used during training:

1. Regression using MLP:

o Objective: To predict a continuous numerical value.
o Output Layer Configuration: For regression tasks, the MLP's
output layer typically consists of a single neuron. This neuron is
generally equipped with a linear activation function (or sometimes
no explicit activation function, simply outputting the raw weighted
 sum). A linear activation allows the neuron to output any real
number, which is necessary for predicting continuous values.
o Loss Function: Common loss functions for regression problems
quantify the difference between the predicted continuous value
and the actual target value. Examples include:
 Mean Squared Error (MSE): Calculates the average of the
squared differences between predicted and actual values. It
heavily penalizes larger errors.
 Mean Absolute Error (MAE): Calculates the average of the
absolute differences between predicted and actual values. It
is less sensitive to outliers than MSE.
o Example: An MLP could be trained to predict the precise kilowatt-
hours (kWh) of electricity consumption for a household based on
features like time of day, temperature, number of occupants, and
appliance usage. The single output neuron would directly provide
the estimated kWh value.
2. Classification using MLP:
o Objective: To predict a categorical label or class.
o Binary Classification (Two Classes):
 Output Layer Configuration: For problems with only two
classes (e.g., "yes/no," "true/false," "spam/not spam"), the
output layer usually has a single neuron employing a sigmoid
(logistic) activation function. The sigmoid function squashes
its input into a value between 0 and 1, which is interpreted as
the probability of the input belonging to the "positive" class.
A threshold (e.g., 0.5) is then applied to convert this
probability into a binary class prediction.
 Loss Function: Binary Cross-Entropy is the standard loss
function for binary classification. It measures the
dissimilarity between the predicted probability distribution
and the true binary distribution.
 Example: An MLP could be trained to predict whether a
credit card transaction is fraudulent or legitimate. The
single output neuron, with a sigmoid activation, would output
a probability (e.g., 0.98 for fraudulent, meaning a 98%
chance of being fraudulent).
o Multi-Class Classification (More Than Two Classes):
 Output Layer Configuration: For problems with more than
two classes (e.g., identifying different types of animals,
classifying handwritten digits), the output layer will have
multiple neurons, typically one neuron for each possible class.
 These output neurons use a softmax activation function.
Softmax takes a vector of arbitrary real values and
transforms them into a probability distribution, where each
value is between 0 and 1, and all values sum to 1. The class
with the highest probability is typically chosen as the
prediction.
 Loss Function: Categorical Cross-Entropy (or Sparse
Categorical Cross-Entropy if labels are integer-encoded) is
the go-to loss function for multi-class classification. It
quantifies the difference between the predicted probability
distribution over the classes and the true one-hot encoded
class label.
 Example: An MLP trained to classify images of different
types of flowers (e.g., rose, tulip, daisy). If there are three
types, the output layer would have three neurons, and their
softmax outputs might be [0.1, 0.8, 0.1], indicating an 80%
probability of being a tulip.

By thoughtfully designing the output layer and choosing the appropriate loss
function, MLPs, and by extension, Deep Neural Networks, can be adapted to
solve a vast array of prediction and classification challenges across various
domains.