Lab 3: The Machine Learning Taxonomy

Objective

Map the landscape of machine learning approaches. By the end you will understand:

• The three main ML paradigms and when to use each

• Key algorithms in each category

• Real-world applications of each approach

• How self-supervised learning powers modern LLMs

---

The Three Main Paradigms

                        MACHINE LEARNING
                              │
          ┌───────────────────┼───────────────────┐
          │                   │                   │
    Supervised          Unsupervised         Reinforcement
    Learning              Learning             Learning
    (labelled data)    (no labels needed)  (learn from rewards)
          │                   │                   │
    ┌─────┴─────┐       ┌─────┴─────┐       ┌─────┴─────┐
Classification Regression Clustering Generative  Policy
  (cat/dog)  (house price) (segments) (GANs, VAE) (games, robots)

---

1. Supervised Learning

The most common paradigm. Every training example has an input and a known correct output (label). The model learns the mapping.

Two subtypes:

Classification — "Which category?"

Output is a discrete class label.

# Binary classification: email spam detection
from sklearn.linear_model import LogisticRegression

# X: feature vectors (word counts, etc.)
# y: 0=not spam, 1=spam
model = LogisticRegression()
model.fit(X_train, y_train)

probability = model.predict_proba(["Buy now!!"])[0][1]
# → 0.94  (94% probability of spam)

Examples:

• Email spam / not spam

• Medical image: tumour / benign

• Credit application: approve / deny

• Sentiment: positive / neutral / negative

Regression — "How much?"

Output is a continuous number.

# House price prediction
from sklearn.ensemble import GradientBoostingRegressor

# Features: bedrooms, location, size, age
model = GradientBoostingRegressor(n_estimators=200)
model.fit(X_train, y_prices)

predicted_price = model.predict([[4, 'London', 120, 5]])
# → [£485,000]

Examples:

• House price prediction

• Stock price forecasting

• Demand forecasting

• Temperature prediction

Common Supervised Algorithms

Algorithm	Best For	Interpretable?
Linear/Logistic Regression	Baseline, simple relationships	✅ Yes
Decision Trees	Categorical features, non-linear	✅ Yes
Random Forest	Tabular data, robust	Partially
Gradient Boosting (XGBoost)	Competitions, tabular data	Partially
Support Vector Machine	High-dimensional, small data	Partially
Neural Networks	Images, text, complex patterns	❌ No

---

2. Unsupervised Learning

No labels. The model must find structure in the data on its own.

Clustering — "Which group?"

Find natural groupings in data.

from sklearn.cluster import KMeans

# Customer segmentation: no labels, just features
kmeans = KMeans(n_clusters=4, random_state=42)
kmeans.fit(customer_features)  # purchase history, demographics

segments = kmeans.labels_
# → [0, 2, 1, 3, 0, 2, ...]
# Segment 0: high-value frequent buyers
# Segment 1: price-sensitive occasional buyers
# ...

Examples:

• Customer segmentation

• Document clustering (group similar articles)

• Anomaly detection (the point that doesn't fit any cluster)

• Gene expression analysis

Dimensionality Reduction — "What's essential?"

Compress high-dimensional data while preserving structure.

from sklearn.decomposition import PCA

# 500-dimensional word embeddings → 2D for visualisation
pca = PCA(n_components=2)
embeddings_2d = pca.fit_transform(word_embeddings_500d)

# Now plot — similar words appear near each other
# "king" near "queen", "London" near "Paris"

Examples:

• Visualising high-dimensional data (t-SNE, UMAP)

• Feature compression before training

• Noise removal from signals

• Recommendation: compress user preferences to latent factors

Generative Models — "Create new data"

Learn the underlying distribution of the training data to generate new samples.

GANs (Generative Adversarial Networks):

Generator          Discriminator
(creates fake) → (real or fake?) → feedback → Generator improves
      ↑                                              │
      └──────────────── adversarial game ────────────┘

VAEs (Variational Autoencoders):

Input → Encoder → Latent space (z) → Decoder → Reconstructed input
                     │
                  sample z → generate new variation

---

3. Reinforcement Learning

The model (agent) learns by taking actions in an environment and receiving rewards or penalties. No labelled data — just trial and error.

┌──────────┐  action  ┌─────────────┐
│  AGENT   │ ───────▶ │ ENVIRONMENT │
│          │ ◀─────── │             │
└──────────┘  reward  └─────────────┘
             + state

# Q-learning: conceptual (cartpole balancing)
import gymnasium as gym

env = gym.make("CartPole-v1")
state, _ = env.reset()

# Agent chooses action based on Q-values (learned value of each action)
action = agent.select_action(state)   # 0=left, 1=right
next_state, reward, done, _, _ = env.step(action)

# Update Q-values based on reward received
agent.update(state, action, reward, next_state)

Famous RL successes:

• AlphaGo / AlphaZero — board games

• OpenAI Five — Dota 2 at world championship level

• AlphaStar — StarCraft II

• ChatGPT — RLHF (humans rate responses; model learns to maximise ratings)

• Robotics — Boston Dynamics locomotion

• Data centre cooling — DeepMind reduced Google's cooling costs 40%

RL is hard because:

• Sparse rewards — agent may do 10,000 actions before getting any feedback

• Reward hacking — agent finds unintended ways to maximise reward

• Sample inefficiency — requires millions of episodes to learn

---

4. Self-Supervised Learning — The Modern Paradigm

The paradigm that powers LLMs. Labels come from the data itself — no human annotation needed.

Masked Language Modelling (BERT):

Input:  "The [MASK] sat on the mat"
Target: predict [MASK] = "cat"

Next Token Prediction (GPT):

Input:  "The cat sat on the"
Target: predict next token = "mat"

By training on trillions of tokens from the internet using these self-supervised tasks, the model is forced to develop deep understanding of language, facts, reasoning patterns, and world knowledge — without a single human-labelled example.

This is the key insight behind modern LLMs: language modelling at scale = general intelligence.

---

Choosing the Right Approach

Do you have labelled data?
├── YES → Supervised Learning
│         ├── Output is a category? → Classification
│         └── Output is a number?   → Regression
└── NO  → Do you know the reward signal?
          ├── YES → Reinforcement Learning (games, robotics, RLHF)
          └── NO  → Unsupervised Learning
                    ├── Find groups?     → Clustering
                    ├── Compress data?  → Dimensionality Reduction
                    └── Generate data?  → GANs, VAEs, Diffusion

---

Summary

Paradigm	Signal	Typical Use	Example Models
Supervised	Labelled data	Classification, regression	XGBoost, ResNet, BERT
Unsupervised	None	Clustering, generation	K-Means, GAN, VAE
Reinforcement	Reward signal	Games, robotics, alignment	PPO, AlphaZero
Self-supervised	Labels from data	LLMs, vision-language	GPT, CLIP, DALL-E

---

Further Reading

• ML Cheatsheet — Supervised vs Unsupervised

• Spinning Up in Deep RL — OpenAI

• The Illustrated BERT — Jay Alammar

• Kaggle Learn — Machine Learning