MACHINE LEARNING – MODULE 4

Complete Teacher-Style Mind Map / Cheat Sheet

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

BAYESIAN LEARNING

Definition

Bayesian Learning uses probability to make predictions under uncertainty.

Main Idea:

Learn from prior knowledge

Update beliefs using new evidence

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

PROBABILITY-BASED LEARNING

Features

Uses probabilistic models

Handles uncertainty & randomness
Learns from data
Predicts future events

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

BAYES THEOREM

Purpose

Calculates posterior probability using prior knowledge and evidence.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Bayes Theorem Formula

P(H|E)=\frac{P(E|H)P(H)}{P(E)}

Terms:
 P(H) → Prior Probability
P(E|H) → Likelihood Probability
P(E) → Marginal Probability
P(H|E) → Posterior Probability

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

TYPES OF PROBABILITIES

Prior Probability

Probability before observing evidence.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Likelihood Probability

Probability of evidence given hypothesis.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Posterior Probability

Updated probability after observing evidence.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

CLASSIFICATION USING BAYES MODEL

Goal

Find most probable hypothesis.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

MAP Hypothesis

Full Form:

Maximum A Posteriori Hypothesis

Rule:

Choose hypothesis with highest posterior probability.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

MAP Formula

h_{MAP}=\arg\max_h P(E|h)P(h)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

MAXIMUM LIKELIHOOD HYPOTHESIS

Definition

Chooses hypothesis with maximum likelihood.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ML Formula

h_{ML}=\arg\max_h P(E|h)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

NAÏVE BAYES ALGORITHM

Definition

Probabilistic classifier assuming features are independent.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Features

Simple & fast

Works well with large datasets
Handles text classification
Uses conditional probability

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

APPLICATIONS OF NAÏVE BAYES

 Spam detection
Recommendation systems
Face recognition
Text classification

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

STEPS IN NAÏVE BAYES

1. Compute prior probability

2. Compute likelihood probability

3. Apply Bayes theorem

4. Compute posterior probability

5. Choose maximum probability class

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

CONDITIONAL PROBABILITY

Formula

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ZERO PROBABILITY ERROR

Problem

If any likelihood becomes zero:

Entire probability becomes zero

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Solution

Laplace Correction / Smoothing

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

LAPLACE CORRECTION
Purpose

Avoid zero probability issue.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Method

Add 1 to every frequency count.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

BRUTE FORCE BAYES ALGORITHM

Definition

Computes posterior probability for all hypotheses.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Problem

Computationally expensive

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

BAYES OPTIMAL CLASSIFIER

Definition

Chooses class with maximum combined posterior probability.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Formula

v_{MAP}=\arg\max_{v_j}\sum_i P(v_j|h_i)P(h_i|T)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

GIBBS ALGORITHM

Definition
Randomly selects one hypothesis according to posterior probability.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Advantage

Reduces prediction error

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

NAÏVE BAYES FOR CONTINUOUS ATTRIBUTES

Problem

Continuous attributes cannot directly use frequency tables.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Solution

Gaussian Distribution

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

GAUSSIAN NAÏVE BAYES

Assumption

Continuous features follow normal distribution.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Gaussian Probability Formula

P(x_i|C_j)=\frac{1}{\sigma_j\sqrt{2\pi}}e^{-\frac{(x_i-\mu_j)^2}{2\sigma_j^2}}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Terms

μ → Mean
σ → Standard deviation
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ARTIFICIAL NEURAL NETWORKS (ANN)

Definition

Computational model inspired by biological brain neurons.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Applications

Pattern recognition
Image processing
Speech recognition
Medical diagnosis
Forecasting

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

HUMAN NERVOUS SYSTEM

Components

Central Nervous System (CNS)

Peripheral Nervous System (PNS)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

BIOLOGICAL NEURON

Parts

Dendrites
Soma / Cell Body
Axon
Synapse

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Working
Input → Processing → Output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ARTIFICIAL NEURON

Definition

Mathematical model of biological neuron.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Components

Inputs
Weights
Summation Function
Activation Function
Output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

MCCULLOCH & PITTS NEURON MODEL

Features

First mathematical neuron model

Uses threshold activation
Fixed weights & threshold

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Net Sum Formula

Net=\sum_{i=1}^{n}x_iw_i

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Step Activation Function

f(x)=\begin{cases}1,&f(x)\geq\theta\0,&f(x)<\theta\end{cases}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ANN STRUCTURE

Layers

Input Layer
Hidden Layer
Output Layer

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Working

1. Inputs enter network

2. Weighted summation

3. Activation applied

4. Output generated

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ACTIVATION FUNCTIONS

Purpose

Introduce non-linearity.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

TYPES OF ACTIVATION FUNCTIONS

Identity Function

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Binary Step Function

f(x)=\begin{cases}1,&x\geq0\0,&x<0\end{cases}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━
 Bipolar Step Function

f(x)=\begin{cases}1,&x\geq0\-1,&x<0\end{cases}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Sigmoid Function

\sigma(x)=\frac{1}{1+e^{-x}}

Features

Output between 0 and 1

S-shaped curve

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Bipolar Sigmoid Function

\sigma(x)=\frac{2}{1+e^{-x}}-1

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Ramp Function

f(x)=\begin{cases}1,&x>1\x,&0\leq x\leq1\0,&x<0\end{cases}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Tanh Function

tanh(x)=\frac{2}{1+e^{-2x}}-1

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ReLU Function

f(x)=max(0,x)

Advantage

Reduces vanishing gradient problem

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━
 Softmax Function

Used for multiclass classification.

s(x_i)=\frac{e^{x_i}}{\sum_{j=1}^{n}e^{x_j}}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

PERCEPTRON MODEL

Definition

Single-layer neural network for binary classification.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Components

Inputs
Weights
Bias
Net Sum
Activation Function

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Perceptron Formula

Net=\sum_{i=1}^{n}x_iw_i+bias

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

PERCEPTRON LEARNING ALGORITHM

Steps

1. Initialize weights

2. Compute net sum

3. Apply activation

4. Calculate error

5. Update weights
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Weight Update Rule

\Delta w_i=\alpha(t-o)x_i

Terms

α → Learning rate
t → Target output
o → Observed output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

XOR PROBLEM

Problem

Single perceptron cannot solve XOR.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Reason

XOR is not linearly separable

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

DELTA LEARNING RULE

Purpose

Minimize error during training.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Training Error Formula

E=\frac{1}{2}\sum_{i=1}^{n}(O_{desired}-O_{estimated})^2

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

TYPES OF ARTIFICIAL NEURAL NETWORKS

 Feed Forward Neural Network

Features

Information flows one direction

Simple architecture

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Fully Connected Neural Network

Features

Every neuron connected to next layer neurons

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Multi-Layer Perceptron (MLP)

Features

Multiple hidden layers

Uses backpropagation
Handles complex problems

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Feedback Neural Network

Features

Feedback connections present

Dynamic learning

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

APPLICATIONS OF ANN

Face recognition
Self-driving cars
Stock prediction
Fraud detection
Medical diagnosis
 Robotics
Forecasting

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

ADVANTAGES OF ANN

Solves complex non-linear problems

Learns patterns automatically
Parallel processing
Handles noisy/incomplete data
High prediction accuracy

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

LIMITATIONS OF ANN

High computational cost

Requires large datasets
Difficult to interpret
Complex model design

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

CHALLENGES OF ANN

Major Problems

Overfitting
Underfitting
Difficult parameter tuning
Real-world generalization issues

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

MOST IMPORTANT EXAM TOPICS

Bayes Theorem
Naïve Bayes Algorithm
Laplace Correction
Gaussian Naïve Bayes
 Artificial Neuron
Activation Functions
Perceptron Model
XOR Problem
Delta Learning Rule
ANN Types

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

FAST REVISION PLAN

First 2 Hours

Learn:

• Bayes Theorem

• Naïve Bayes

• MAP & ML Hypothesis

• Gaussian Naïve Bayes

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Next 2 Hours

Learn:

• Artificial Neuron

• ANN Structure

• Activation Functions

• Perceptron

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━

Final 2 Hours

Learn:

• Delta Rule

• XOR Problem
 • ANN Types

• Applications & Limitations

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
━