Machine Learning (ML): Types of learning paradigms
1. Definition
Machine Learning (ML) is the branch of AI in which a system learns patterns from data and improves its performance with experience, instead of being given explicit rules for every case. Tom Mitchell's definition: "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."
TRADITIONAL PROGRAMMING MACHINE LEARNING
data + RULES → answers data + ANSWERS → rules (the model)
(a human writes the logic) (the algorithm finds the logic)
ML is the right tool when the rule is too complex to write by hand (recognise a face, predict demand) but plenty of labelled or observable examples exist. It is the wrong tool when the rule is simple, exact and auditable.
2. The Learning Paradigms (the core of this topic)
| Paradigm | Training data | Goal | Typical tasks | Algorithms |
|---|---|---|---|---|
| Supervised learning | Labelled (input → known output) | Learn a mapping from input to output | Classification, regression | Linear/logistic regression, KNN, SVM, naive Bayes, decision trees, random forest, gradient boosting, neural networks |
| Unsupervised learning | Unlabelled (inputs only) | Discover inherent structure | Clustering, dimensionality reduction, anomaly detection, association rules | K-means, hierarchical clustering, DBSCAN, PCA, t-SNE, Apriori, autoencoders |
| Semi-supervised learning | A few labelled + many unlabelled | Use cheap unlabelled data to boost a small labelled set | Classification with expensive labels | Self-training, co-training, label propagation |
| Reinforcement learning (RL) | No dataset — feedback from an environment | Learn a policy that maximises long-term reward | Game playing, robotics, control, scheduling | Q-learning, SARSA, DQN, policy gradients, PPO |
| Self-supervised learning | Unlabelled data with labels generated from the data itself | Learn general representations | Language modelling, image pretraining | Masked-language/next-token prediction, contrastive learning (modern LLMs) |
3. Supervised Learning in Detail
- Classification — output is a class: spam/not-spam, disease/no-disease, digit 0-9, sentiment positive/negative/neutral, loan default yes/no.
- Regression — output is a continuous number: house price, temperature, demand, delivery time.
- Workflow: collect and clean labelled data → feature engineering → split into training / validation / test sets (e.g. 70/15/15) → train the model by minimising a loss function → tune hyperparameters on validation → evaluate once on the untouched test set → deploy → monitor.
- Evaluation metrics: accuracy, precision, recall, F1 (important with imbalanced classes), confusion matrix, ROC-AUC for classifiers; RMSE/MAE/R² for regression.
- Pitfalls: overfitting (memorising the training set — great on train, poor on test), underfitting (too simple), class imbalance, data leakage, and non-representative samples. Cures: more/cleaner data, regularisation, cross-validation, simpler models, augmentation.
4. Unsupervised Learning in Detail
- Clustering — group similar items: customer segmentation, document grouping, image compression, market basket grouping.
- Dimensionality reduction — compress many features into few while preserving structure (PCA, t-SNE) for visualisation and speed.
- Anomaly detection — flag the unusual: network intrusion, machine fault, fraudulent claim.
- Association rule mining — "customers who bought X also bought Y" (Apriori), used in cross-selling and recommendation.
There is no "correct answer" to compare against, so evaluation is harder — silhouette scores, business usefulness and stability of clusters are used instead.
5. Reinforcement Learning in Detail
An agent interacts with an environment: it observes a state, takes an action, receives a reward and moves to the next state, learning a policy that maximises cumulative reward. Key terms: state, action, reward, policy, value function, discount factor, episode; the trade-off between exploration (try new actions) and exploitation (use what works); the problem is modelled as a Markov Decision Process (MDP). Rewards may be sparse and delayed — the credit-assignment problem.
Applications: game playing (AlphaGo, Atari), robot walking and grasping, traffic-signal control, energy/cooling optimisation, ad bidding, recommendation, and LLM alignment through RLHF (reinforcement learning from human feedback).
6. Other Useful Concepts
Batch vs online learning (train on all data at once vs continuously update) · instance-based vs model-based (memorise examples, e.g. KNN, vs learn parameters) · transfer learning (reuse a model trained on a big dataset for a new task — the norm in vision and NLP) · few-shot / zero-shot learning · federated learning (train across devices without moving raw data — privacy-friendly) · ensemble methods (bagging/boosting, random forests) · MLOps (versioning data and models, deployment, monitoring for drift, retraining).
7. Supervised vs Unsupervised vs Reinforcement — quick contrast
| Basis | Supervised | Unsupervised | Reinforcement |
|---|---|---|---|
| Labels | Required | Not required | None — reward signal instead |
| Feedback | Immediate per example | None | Delayed, after actions |
| Main tasks | Classification, regression | Clustering, reduction, anomalies | Sequential decision-making, control |
| Data need | Large labelled dataset | Any dataset | Environment/simulator to interact with |
| Example | Spam filter | Customer segmentation | Self-driving control, game AI |
8. Exam Points
- Define machine learning; differentiate it from traditional programming — 5 marks, frequent.
- Explain supervised, unsupervised and reinforcement learning with examples — 7 marks, almost certain.
- Differentiate classification and regression — 3-5 marks.
- Explain RL with agent, state, action, reward and policy; give two applications — 5 marks, frequent.
- What is overfitting? How is it detected and prevented? — 5 marks.
- Explain the ML pipeline / workflow steps — 5 marks.
- Define semi-supervised and self-supervised learning — 3-5 marks.
- Which paradigm suits a given problem (segmentation, spam filter, robot walking)? — 3-5 marks.