📘 Artificial Intelligence · Chapter 10

Machine Learning
A Complete Study Guide

From Rote Memorization to Complex Intelligence — everything you need to ace your seminar on Sections 10.1 – 10.3.8.

8Learning Methods
3Real-World Examples
10+Key Definitions

Introduction to Machine Learning

📖
Definition — Machine Learning (ML): The field concerned with algorithms that improve their performance at a given task through experience. An ML agent gets better the more data it processes, without being explicitly reprogrammed.
🎯

Goals of Machine Learning

  • Improve the performance of an intelligent agent over time
  • Enable agents to adapt to new, unseen scenarios
  • Automate complex decision-making without hand-coding every rule
  • Extract actionable knowledge from raw data
⚠️

Core Challenges

  • Handling large, dynamic datasets that shift over time
  • Identifying relevant features amid noise and irrelevant data
  • Avoiding overfitting — learning the noise, not the signal
  • Scaling algorithms to real-world complexity

Concept & Scope of ML

Machine Learning sits at the intersection of computer science, statistics, and domain knowledge. Its scope spans from simple linear models to deep neural networks — covering perception (recognising images, speech), prediction (credit scores, demand forecasting), and decision-making (game-playing agents, robotics).

Real-World AI Applications
Machine Learning Algorithms
Core: Data + Experience
💡
Pro-Tip: Think of ML as a feedback loop. The agent observes data → makes a prediction → receives feedback → updates its internal model. Over thousands of cycles, even a simple algorithm can achieve superhuman performance.

Learning Paradigms

Intelligence is not monolithic. A truly intelligent agent must learn, search, and problem-solve — often simultaneously. These three pillars form the backbone of intelligent behaviour.

📚

Learning

Acquiring new knowledge or skills from experience — through examples, feedback, or exploration of the environment.

Core Paradigm
🧩

Problem Solving

Formulating goals and executing action sequences to reach them — blending planning, reasoning, and learned heuristics.

Core Paradigm
🔗
Key Insight: These paradigms are interdependent. Learning improves search efficiency; better search unlocks harder problem-solving; solving problems yields new training data that fuels further learning. This cycle is what drives modern AI breakthroughs like AlphaZero.

Observation Phase

The agent passively collects data from the environment — sensor readings, user behaviour, game moves.

Induction Phase

Patterns are extracted from the raw data and distilled into a generalisable internal model or rule set.

Inference/Action Phase

The trained model is applied to new inputs — making predictions, taking actions, or solving novel tasks.

Feedback & Adaptation

The outcome is evaluated; the model is updated to correct errors and reinforce successful strategies.

Learning Concepts, Methods & Models

Machine learning encompasses a rich family of algorithms, each suited to different data types, task goals, and feedback mechanisms. Below we explore each method in depth.

🗃️ Rote
👁️ Observations
🏷️ Supervised
🔓 Unsupervised
⚖️ Semi-supervised
🎭 Ensemble
⛏️ Discovery
🧩 Problem Solving
10.3.1

Rote Learning

Learning through direct memorisation

📖
Definition: Rote Learning is the simplest form of machine learning — the direct storage of input-output pairs in memory, retrieved verbatim when an identical input is encountered again. No generalisation occurs.

How It Works

An agent stores every experienced situation and its corresponding optimal action in a lookup table. When the same situation recurs, the stored answer is returned instantly without any recomputation.

📥
Input Data
🗃️
Store in Memory
🔍
Exact Match Lookup
📤
Retrieve & Return

Characteristics

  • ✔ Strength Perfect recall for previously seen inputs
  • ✔ Strength Zero recomputation cost at inference time
  • ✘ Weakness Cannot generalise to unseen situations
  • ✘ Weakness Memory grows linearly with experience
  • ✘ Weakness Fails when data is dynamic or context-dependent
🌍
Real-World Example: A chess engine precomputing and caching every opening move sequence. When an opponent plays exactly that sequence, the engine retrieves the optimal response in O(1) time — but if the opponent deviates, the engine must fall back to search.
10.3.2

Learning from Observations

Extracting patterns from recorded facts

📖
Definition: The agent observes the environment, records facts (observations), and extracts general patterns or rules from the accumulated record — without a teacher providing explicit correct answers.

The Observation Cycle

Observe
Record
Analyse
Generalise
Predict

Key Principles

  • Observations are recorded as a set of attribute-value pairs
  • Patterns are discovered via inductive reasoning
  • No explicit target label is provided during observation
  • Works well for time-series and sensor data
💡
Pro-Tip: Learning from observations is the computational equivalent of a scientist formulating a hypothesis. The quality of learning depends heavily on the richness and representativeness of the observations collected.
10.3.3

Supervised Learning

Learning with labeled data and explicit feedback

📖
Definition: The learning algorithm is trained on a labeled dataset — each example consists of an input and a corresponding correct output. The model learns a mapping f(X) → Y that generalises to unseen inputs.

Classification

Predicts a discrete label from a finite set of classes.

  • 📧 Spam vs. Not-Spam email detection
  • 🩺 Disease diagnosis (positive / negative)
  • 💳 Credit default prediction (default / no default)

Common Algorithms: Decision Trees, Naive Bayes, SVM, Logistic Regression, k-NN

The Supervised Learning Pipeline

1Collect & Label Data
2Train Model
3Validate & Tune
4Test on Held-out Data
5Deploy
10.3.4

Unsupervised Learning

Finding hidden structure without labels

📖
Definition: The algorithm is presented with unlabeled data and must discover hidden structure, patterns, or compressed representations on its own — without any notion of a "correct" answer.
🪣

Clustering

Group similar data points together. e.g. k-Means, DBSCAN, Hierarchical Clustering.

📉

Dimensionality Reduction

Compress high-dimensional data while preserving structure. e.g. PCA, t-SNE, Autoencoders.

🔗

Association Rule Mining

Discover co-occurrence rules in transactional data. e.g. Apriori algorithm, market basket analysis.

🎨

Generative Modelling

Learn the data distribution to generate new samples. e.g. GANs, Variational Autoencoders.

💡
Pro-Tip: Unsupervised learning is especially powerful when labeling is expensive or impossible — which is true for most real-world data. Much of the modern AI revolution (LLMs, diffusion models) is driven by unsupervised pre-training.
10.3.5

Semi-supervised Learning

A powerful hybrid of labeled and unlabeled data

📖
Definition: A hybrid paradigm that uses a small amount of labeled data combined with a large pool of unlabeled data during training. The unlabeled data improves the model's understanding of the input distribution.
Supervised All labeled
←   Semi-supervised   →
Unsupervised All unlabeled

Why it matters

In practice, collecting labeled data is expensive and time-consuming (think: manually annotating thousands of medical images). Semi-supervised learning leverages the cheap unlabeled data to significantly boost accuracy over a purely supervised approach.

Common Techniques

  • Self-training — iteratively label confident predictions
  • Co-training — two models teach each other
  • Graph-based methods — label propagation
  • Generative models — learn joint distribution

Real-World Applications

  • 🏥 Medical image classification
  • 📰 Text document categorisation
  • 🗣️ Speech recognition
  • 🌐 Web content classification
10.3.6

Ensemble Learning

Many weak learners, one strong result

📖
Definition: An ensemble combines the predictions of multiple individually trained models (called base learners) to produce a final prediction that is more accurate and robust than any single model alone.
🌲

Bagging

Train multiple models on random subsets of data, then aggregate (e.g. majority vote). Reduces variance.

e.g. Random Forest
🚀

Boosting

Train models sequentially; each focuses on the errors of its predecessor. Reduces bias.

e.g. XGBoost, AdaBoost
🗳️

Stacking

Use a meta-learner to learn how to best combine the outputs of base learners.

e.g. Blended ensembles
💡
Pro-Tip: "Diversity is strength." Ensembles work best when the base learners make different kinds of errors. Combining three models that all fail on the same examples gives little benefit; combining models that fail on different examples creates a powerful safety net.
10.3.7

Discovery-based Learning

Inductively finding previously unknown, meaningful information

📖
Definition: Also called Knowledge Discovery in Databases (KDD) or Data Mining, this paradigm inductively extracts previously unknown, potentially useful, and ultimately understandable patterns from large datasets.

The KDD Process

🗄️
Raw Data
🧽
Preprocessing
🔄
Transformation
⛏️
Data Mining
💡
Knowledge

Key Properties of Discovered Knowledge

  • Previously Unknown — not an obvious or trivial relationship
  • Valid — holds true with high confidence on the data
  • Useful — actionable in a business or scientific context
  • Understandable — interpretable by a human expert
🌍
Classic Example: "Customers who buy diapers on Friday evenings also tend to buy beer." — A supermarket discovered this pattern in transaction logs. This counterintuitive association rule was only possible through automated data mining of millions of receipts.
10.3.8

Learning by Problem Solving

Improving performance through task execution and optimisation

📖
Definition: The agent actively attempts to solve real tasks. Through repeated trial-and-error and optimisation of its action sequences, it improves its problem-solving strategy over time — leveraging experience from past attempts.

The Learn-by-Doing Loop

🎯
Problem Encountered
🔄
Apply Stored Strategy
📊
Evaluate Outcome
✏️
Update / Optimise Strategy

Connection to Reinforcement Learning

Learning by problem solving is the conceptual foundation of Reinforcement Learning (RL). The agent receives a reward signal after each action and uses it to refine its policy — the mapping from states to actions.

🌍 Environment
🤖 Agent
📍 State
🎮 Action
🏆 Reward
📋 Policy
💡
Pro-Tip: AlphaGo Zero learned to play Go at superhuman level purely through self-play — learning by problem solving with zero human game data. Starting from random play, it discovered strategies that human players had never conceived.

Knowledge Discovery vs. Machine Learning

These two fields are often conflated, but they have distinct goals and processes. Here is a definitive side-by-side comparison.

Dimension 📂 Knowledge Discovery 🤖 Machine Learning
Primary Goal Extract previously unknown patterns from data Improve predictive performance on a task
Output Understandable rules and associations A trained model (can be a black box)
Human Interpretability ⭐⭐⭐⭐⭐ Highest priority ⭐⭐ Often secondary to accuracy
Feedback Required ❌ Typically unsupervised ✅ Usually requires labels
Data Volume Very large databases (petabyte scale) Varies — can be small (few-shot learning)
Primary Techniques Association Rules, Clustering, Anomaly Detection Neural Networks, SVMs, Decision Trees
Classic Example Diaper-beer association in supermarkets Image recognition via CNNs
Metrics Support, Confidence, Lift Accuracy, F1-score, AUC-ROC
🔗
Overlap: The two fields are complementary, not competing. ML algorithms are often used as the data-mining engine within a KDD pipeline. The distinction lies in the intent — discovering new knowledge vs. optimising a prediction function.

Bank Credit Risk: The Ram Example

This case study illustrates how ML predictions are made in a high-stakes financial setting.

👤

The Scenario

Ram applies for a bank loan. The bank's ML system must decide: will Ram default on his loan repayment? The system doesn't know Ram personally — it must infer from patterns seen in thousands of historical loan applications.

Input Parameters (Features)

📅
Age
32 years
💼
Employment Status
Employed
💰
Annual Income
₹6,00,000
📊
Credit Score
580 (Fair)
🏦
Existing Debts
High
📋
Credit History
Patchy (2 late payments)
ML Model Output
Default Risk: 67% — High Risk
⚠️ Loan Application: Declined
🧠
Key Insight: The model didn't "know" Ram would default. It recognised that Ram's feature profile closely matches historical profiles of customers who did default. This is the essence of supervised classification: learning from labeled historical data to make informed predictions on new applicants.

Multi-perspective Integrated Intelligence (MPI)

MPI is a theoretical framework that describes how a truly intelligent system integrates multiple cognitive perspectives — learning, reasoning, and problem-solving — rather than relying on a single, monolithic approach.

💼 The Job Application Analogy

📄
Candidate applies for a job

Raw input data — CV, cover letter, interview performance.

🔍
HR evaluates the candidate

The system observes and scores multiple dimensions: skills, experience, cultural fit.

📩
Candidate receives feedback

A rejection letter arrives — but it contains specific, actionable feedback on what was lacking.

📚
Candidate learns & adapts

The candidate improves their skills, refines their CV, and applies again — now with higher success probability.

Eventual success through integrated intelligence

The iterative feedback loop — observe, reason, learn, act — mirrors how MPI agents improve.

The Three MPI Pillars

👓

Perceptual Intelligence

The ability to observe and interpret raw data inputs from the environment with high fidelity.

+
🧠

Cognitive Intelligence

Reasoning, pattern recognition, and knowledge integration across domains and time steps.

+

Adaptive Intelligence

Learning from feedback and dynamically updating strategies to improve future performance.

= Multi-perspective Integrated Intelligence (MPI)
💡
Pro-Tip: MPI is not just a theoretical construct — it is the design philosophy behind modern AI systems. A large language model demonstrates MPI by simultaneously leveraging perceptual pattern matching (token prediction), cognitive reasoning (chain-of-thought), and adaptation (RLHF fine-tuning).

Test Your Understanding

Challenge yourself with these seminar-style questions. Click an answer to see the explanation.