Introduction to Machine Learning
The Introduction to Machine Learning domain tests whether you can define ML precisely, distinguish learning paradigms, and reason about data and generalization. Questions often ask you to classify a scenario as supervised, unsupervised, or reinforcement learning, or to diagnose overfitting from train/test results.
Core distinctions
- Artificial intelligence is the broad goal of machines performing cognitive tasks; machine learning is the subset where models improve from data rather than explicit rules; deep learning is a subset of ML using multi-layer neural networks.
- Supervised learning uses labelled examples. Classification predicts a discrete label (spam/not spam); regression predicts a continuous value (house price). Unsupervised learning finds structure without labels: clustering, dimensionality reduction, anomaly detection. Reinforcement learning trains an agent through rewards and penalties from an environment.
- A dataset has instances (rows), features (input variables), and, in supervised learning, labels (target). Training data fit the model; validation data tune it; test data estimate final performance once.
- Generalization is performance on unseen data. Overfitting means low training error but high test error (high variance); underfitting means poor performance on both (high bias). More data, simpler models, regularization, and cross-validation reduce overfitting.
Lifecycle
CRISP-DM is iterative: business understanding, data understanding, data preparation, modelling, evaluation, deployment. Data preparation usually consumes most effort. Deployment is not the end; monitoring feeds back into the cycle.
Worked example
A team builds an email spam classifier from 10,000 labelled emails. They split 70% training, 15% validation, 15% test. Training accuracy is 99.2%, validation 74%, test 73%. Reasoning: the gap between training and held-out performance signals overfitting, not a data-split error, because validation and test agree. The team should simplify the model, add regularization, gather more diverse emails, and use k-fold cross-validation on the training set to tune hyperparameters. They must not tune on the test set; it is opened once for the final estimate.
Traps
Do not call unsupervised learning 'unlabelled classification'; it has no labels by definition. Do not use test data for feature selection or scaling. Accuracy is misleading on imbalanced data. CRISP-DM is iterative, not a one-way waterfall. ML is not synonymous with AI or deep learning.