What Is Machine Learning?

Machine learning (ML) is a technology that enables computers to learn patterns from data and make predictions or decisions on new data — without being explicitly programmed.

Traditional Programming:
  Input + Rules → Output

Machine Learning:
  Input + Output → Rules (learned automatically)

This paradigm shift is the essence of ML. Instead of a developer coding every rule, an algorithm discovers the rules from data.

When ML is the right tool:

The rules are too complex or hard to express explicitly (image recognition, natural language understanding)
The environment changes frequently, making manual rule updates impractical (price prediction, recommender systems)
You need to discover hidden patterns in large datasets (anomaly detection, customer segmentation)

The AI / ML / DL Hierarchy

Artificial intelligence, machine learning, and deep learning are often used interchangeably, but they have a clear containment relationship.

┌──────────────────────────────────────────────┐
│  AI (Artificial Intelligence)                 │
│  → Any technology that mimics human cognition │
│  ┌────────────────────────────────────────┐   │
│  │  ML (Machine Learning)                 │   │
│  │  → The subset of AI that learns        │   │
│  │    from data                           │   │
│  │  ┌──────────────────────────────────┐  │   │
│  │  │  DL (Deep Learning)              │  │   │
│  │  │  → The subset of ML that uses    │  │   │
│  │  │    multi-layer neural networks   │  │   │
│  │  └──────────────────────────────────┘  │   │
│  └────────────────────────────────────────┘   │
└──────────────────────────────────────────────┘

AI: includes game AI and rule-based expert systems
ML: statistical / optimization-based learning algorithms (linear regression, random forests, etc.)
DL: multi-layer networks with millions of parameters (CNN, RNN, Transformer)

Types of Machine Learning

Supervised Learning

Trains on data that has labeled answers. The goal is to learn a mapping function from inputs to outputs.

Training data format:
{(x₁, y₁), (x₂, y₂), ..., (xₙ, yₙ)}

x: input features
y: target label (the answer)

Type	Output	Examples
Classification	Discrete class	Email spam / not spam; tumor malignant / benign
Regression	Continuous value	House price; tomorrow’s temperature

Unsupervised Learning

Discovers hidden structure in data without labeled answers.

Training data format:
{x₁, x₂, ..., xₙ}  ← no y (no labels)

Clustering: group similar data points (K-means, DBSCAN)
Dimensionality Reduction: compress high-dimensional data (PCA, t-SNE)
Anomaly Detection: identify data that deviates from normal patterns

Reinforcement Learning

An agent learns a policy that maximizes cumulative reward through interaction with an environment.

Agent → Action → Environment
Environment → State + Reward → Agent

Examples: AlphaGo, game-playing AI, robot control

Linear Regression

Linear regression is the most fundamental regression algorithm in supervised learning. It models a linear relationship between input features and a continuous target variable.

Hypothesis Function

Simple linear regression:
h(x) = θ₀ + θ₁x

where:
h(x): predicted value
θ₀:   y-intercept (bias)
θ₁:   slope (weight / coefficient)
x:    input feature

Intuition: θ₀ is the “baseline value” when x = 0. θ₁ is how much the prediction changes when x increases by 1.

Example: predicting house price from size (sq ft)
h(x) = 50,000 + 200x

Size x = 1,000 sq ft → predicted price = $250,000
Size x = 2,000 sq ft → predicted price = $450,000

Cost Function

The cost function quantifies how wrong the model’s predictions are. The goal of training is to find the θ₀ and θ₁ values that minimize the cost function.

MSE — Mean Squared Error

J(θ₀, θ₁) = (1 / 2m) × Σᵢ₌₁ᵐ (h(xᵢ) − yᵢ)²

where:
m:      number of training samples
h(xᵢ): predicted value for sample i
yᵢ:    actual value for sample i
(h(xᵢ) − yᵢ): residual error

Why square the residuals?

Prevents positive and negative errors from canceling out
Penalizes larger errors more heavily (squares amplify outliers)
Makes the function differentiable — required for optimization

Intuition:
Perfect prediction → J = 0
Larger prediction errors → J increases
Training = finding θ values that minimize J

Gradient Descent

An iterative optimization algorithm that updates parameters repeatedly to minimize the cost function.

Core Intuition

Imagine a hiker descending a mountain:
- At the current position, identify the steepest downhill direction
- Take one step in that direction
- Repeat until reaching the valley (minimum)

Mountain       = cost function J(θ)
Current position = current θ value
Steepest downhill = negative gradient direction (−∇J)
Step size         = learning rate α

Update Rule

θ₀ := θ₀ − α × ∂J/∂θ₀
θ₁ := θ₁ − α × ∂J/∂θ₁

where:
α:         learning rate (0 < α < 1)
∂J/∂θ:    partial derivative (gradient) of cost function

For linear regression:
∂J/∂θ₀ = (1/m) × Σ(h(xᵢ) − yᵢ)
∂J/∂θ₁ = (1/m) × Σ(h(xᵢ) − yᵢ) × xᵢ

Learning Rate α — Why It Matters

α too large:  overshoots the minimum → diverges
α too small:  converges extremely slowly → computationally wasteful
α just right: cost function decreases smoothly and converges

Typical starting values: α = 0.01, 0.001
Verify by plotting cost vs. iteration

Variants of Gradient Descent

Variant	Description	Trade-off
Batch GD	Updates using the full dataset	Stable; slow for large data
SGD	Updates on one sample at a time	Fast; noisy
Mini-Batch GD	Updates on small batches (32–256)	Best balance (most common)

Multiple Linear Regression

Real data involves multiple features simultaneously affecting the target.

h(x) = θ₀ + θ₁x₁ + θ₂x₂ + ... + θₙxₙ

Vector notation:
h(x) = θᵀx

Example: predicting house price
h(x) = θ₀ + θ₁×size + θ₂×bedrooms + θ₃×age + θ₄×floor

Feature Scaling

When features have very different ranges, gradient descent converges slowly.

Example:
x₁ = house size (500 – 5,000 sq ft)
x₂ = bedrooms (1 – 6)

→ x₁ has much larger scale
→ cost function contours become elongated ellipses
→ gradient descent zig-zags inefficiently

Min-Max Normalization

x' = (x − x_min) / (x_max − x_min)

Result: all features scaled to [0, 1]

Standardization (Z-score)

x' = (x − μ) / σ

Result: mean = 0, standard deviation = 1
Advantage: less sensitive to outliers; works well for normally distributed features

For gradient-descent-based algorithms, always apply feature scaling first.

Underfitting vs. Overfitting

Underfitting (High Bias):
→ Model too simple to capture the pattern
→ Both training and test error are high

Overfitting (High Variance):
→ Model memorizes training data including noise
→ Training error low; test error high

Good Fit:
→ Balanced bias and variance
→ Generalizes well to unseen data

Train / Validation / Test Split

from sklearn.model_selection import train_test_split

# 80/20 split
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

Key Concept Cards

Hypothesis h(x) = θ₀ + θ₁x ★★★★★
The core of linear regression. θ₀ is the bias (intercept), θ₁ is the weight (slope). Training = finding the optimal θ₀ and θ₁.

Cost Function MSE and the training goal ★★★★★
J(θ) = (1/2m) × Σ(h(xᵢ) − yᵢ)². Average squared residuals. Minimizing J is the goal of training.

Gradient Descent update rule ★★★★★
θ := θ − α×(∂J/∂θ). Move in the direction opposite the gradient by step size α. Repeat until convergence.

AI ⊃ ML ⊃ DL ★★★★☆
Deep learning is a subset of ML; ML is a subset of AI. They are not synonyms.

Practice Quiz

Q1. Given h(x) = 2 + 3x, what is the prediction at x = 5? What do θ₀ and θ₁ represent?

Prediction = 2 + 3×5 = 17. θ₀ = 2 is the y-intercept (baseline prediction when x = 0). θ₁ = 3 is the slope (each unit increase in x raises the prediction by 3).

Q2. What goes wrong if the learning rate α is too large or too small?

If α is too large, gradient descent overshoots the minimum and the cost function diverges (fails to converge). If α is too small, updates are tiny, requiring a very large number of iterations — training becomes impractically slow. Try values like 0.001, 0.01, 0.1 and monitor the cost-vs-iteration curve.

Q3. What is the key difference between supervised and unsupervised learning?

Supervised learning trains on data with labeled outputs (y) — it learns an input-to-output mapping. Unsupervised learning has no labels — it discovers inherent structures (clusters, patterns) within the data itself.

Ch1. Machine Learning — Supervised Learning and Linear Regression