0.1 Introduction to Machine Learning: Definition, AI vs ML vs DL★★

# Who Wrote the Rules?

• Nobody wrote rules telling your phone what your face looks like.
• Yet it unlocks the moment it sees you.
• It learned your face from examples, the way a child learns.
• That's machine learning, and it changes what computers can do.

# What Is Machine Learning PYQ 1M

Machine learning (ML) is the field in which computers learn patterns from data and improve at a task with experience, instead of following hand-written rules.

Classic definitions

• Arthur Samuel (1959): the field of study that gives computers the ability to learn without being explicitly programmed.
• Tom Mitchell (1997): "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."

Worked example: a spam filter

Training data

• The sample data a model learns from is called training data. The learnt result is the model.

flowchart TD
    A["<b>Traditional code</b><br>data + rules"]:::blue --> B["<b>Output</b>"]:::blue
    C["<b>Machine learning</b><br>data + answers"]:::amber --> D["<b>Rules (model)</b><br>learnt"]:::green
    classDef blue fill:#5b84d424,stroke:#5b84d480
    classDef amber fill:#d6a24424,stroke:#d6a24480
    classDef red fill:#d0707024,stroke:#d0707080
    classDef green fill:#52a86e24,stroke:#52a86e80

• In traditional programming, humans write the rules; in ML, the computer finds the rules from examples.

# Artificial Intelligence PYQ 5M

Artificial intelligence (AI) is the science of making machines perform tasks that normally need human intelligence, such as reasoning, understanding language, seeing and deciding. The term was coined by John McCarthy in 1956.

Goals of AI

• Reasoning and problem-solving · knowledge representation · perception (vision, speech) · natural language understanding · decision-making and planning · learning.

Why we need AI

1. Huge data: humans can't analyse billions of records; AI can.
2. Automation of complex tasks: driving assistance, medical image screening, fraud checks.
3. Speed and scale: decisions in milliseconds, 24×7.
4. Accuracy and consistency: no fatigue or mood swings.
5. Personalisation: recommendations, adaptive learning apps.
6. Dangerous or repetitive work: robots in mines, space, factories.

• Why now: cheap computing power (GPUs, cloud), vast data from the internet and sensors, and better algorithms.

Narrow vs general AI

• Narrow (weak) AI: good at one task: chess engines, voice assistants, face unlock. All AI today is narrow.
• General (strong) AI: human-level intelligence across all tasks; still a research goal.

# How ML Relates to AI PYQ 4M

• ML is a subset of AI: the data-driven way of achieving AI.
• Early AI was mostly rule-based (expert systems): humans wrote "if–then" rules. These broke when situations were too many to list.
• ML lets the system learn the rules from data, so it handles new cases and improves with more data.
• Most modern AI (speech recognition, recommendations, translation) is built with ML.

# AI vs ML vs DL PYQ 6M

Deep learning (DL) is a subset of ML that uses neural networks with many layers to learn features directly from raw data (→ 4.4).

flowchart TD
    subgraph AI["Artificial Intelligence"]
        direction TB
        a1["machines that act<br>intelligently"]:::blue
        subgraph ML["Machine Learning"]
            direction TB
            m1["learns patterns from data"]:::amber
            subgraph DL["Deep Learning"]
                d1["many-layer neural networks"]:::green
            end
        end
    end
    a1 ~~~ m1
    m1 ~~~ d1
    classDef blue fill:#5b84d424,stroke:#5b84d480
    classDef amber fill:#d6a24424,stroke:#d6a24480
    classDef red fill:#d0707024,stroke:#d0707080
    classDef green fill:#52a86e24,stroke:#52a86e80
    style AI fill:#5b84d414,stroke:#5b84d480
    style ML fill:#d6a24414,stroke:#d6a24480
    style DL fill:#52a86e14,stroke:#52a86e80

• Every DL system is ML, and every ML system is AI, but not the other way round.
• The detailed ML vs DL comparison → 4.4.

# Applications of Machine Learning

1. Healthcare: disease prediction from reports, cancer detection in scans.
2. Finance: fraud detection on card transactions, credit scoring.
3. E-commerce: product recommendations ("customers also bought").
4. Computer vision: face unlock, number-plate reading, quality checks in factories.
5. Speech and language: voice assistants, translation, chatbots.
6. Autonomous vehicles: detecting lanes, signs and pedestrians.
7. Spam and malware filtering in e-mail.

# Steps to Develop an ML Application

flowchart TD
    A["<b>1 · Define the problem</b><br>goal, success measure"]:::blue --> B["<b>2 · Collect data</b>"]:::blue
    B --> C["<b>3 · Prepare data</b><br>clean, features"]:::amber
    C --> D["<b>4 · Choose a model</b>"]:::amber
    D --> E["<b>5 · Train</b>"]:::amber
    E --> F["<b>6 · Evaluate and tune</b>"]:::amber
    F -.->|"not good enough"| D
    F --> G["<b>7 · Deploy and monitor</b>"]:::green
    classDef blue fill:#5b84d424,stroke:#5b84d480
    classDef amber fill:#d6a24424,stroke:#d6a24480
    classDef red fill:#d0707024,stroke:#d0707080
    classDef green fill:#52a86e24,stroke:#52a86e80

1. Problem definition: what to predict, and how success is measured (accuracy, error).
2. Data collection: from databases, sensors, web, surveys.
3. Pre-processing: handle missing values, remove noise, encode and scale features (→ 0.3).
4. Feature engineering: choose or create the most useful inputs.
5. Model selection: pick suitable algorithms (tree, regression, neural network…).
6. Training: the algorithm learns from the training data.
7. Evaluation: test on unseen data with suitable metrics (→ 3.1).
8. Tuning: adjust hyperparameters, try other models (→ 3.4).
9. Deployment and monitoring: put it into use; retrain when data changes.

• Example: predicting house prices → collect past sales → clean and encode location → try linear regression and random forest → choose the lower-error model → deploy on a property website.

# Objectives, Design Issues and Approaches

Objectives of ML

• Predict outcomes for new data · automate decisions · discover hidden patterns · improve with experience · generalise from examples to unseen cases.

Designing a learning system (Mitchell's checkers example)

1. Choose the training experience: what data the system learns from (games played against itself).
2. Choose the target function: what exactly to learn (the value of a board position).
3. Choose its representation: how to express it (a weighted sum of board features).
4. Choose the learning algorithm: how to adjust it (update weights to reduce error).

Design issues

• Data quality and quantity: too little or noisy data gives poor models.
• Overfitting vs underfitting (→ 3.3).
• Bias and fairness: biased data gives biased decisions.
• Interpretability: can we explain the prediction?
• Computational cost: training time and memory.
• Choosing the algorithm and features.

Approaches

• By learning type: supervised, unsupervised, reinforcement (→ 0.2).
• By method: tree-based, instance-based (KNN), probabilistic (Naive Bayes), linear models, neural networks.

# Can Every Problem Be Solved by ML?

No. ML fits only some problems.

When ML fits

• A pattern exists but is hard to write as rules (recognising faces).
• Enough data is available.
• Some error is acceptable.
• The situation changes over time and the system must adapt (spam).

When ML does not fit

• Exact rules already exist: tax calculation, sorting numbers, interest formulas.
• No or very little data: a brand-new product with no history.
• 100% accuracy or full explanation needed: legal judgements, safety-critical logic.
• Purely random events: predicting lottery numbers.
• Example: ML suits predicting which customers may leave a telecom company; it does not suit calculating their monthly bill.