AI and Machine Learning Fundamentals
The nesting you must be able to state
AI is the broad discipline; machine learning (ML) is the subset where models learn patterns from data instead of being explicitly programmed; deep learning (DL) is the subset of ML built on multi-layer neural networks. A scenario describing hand-written rules by a developer is not ML.
Learning types
- Supervised: labeled data. Classification predicts a discrete class (spam / not spam); regression predicts a continuous number (price, demand).
- Unsupervised: no labels - clustering, dimensionality reduction, anomaly detection.
- Reinforcement: an agent learns from rewards in an environment.
Data splits and fit
Training data fits weights, validation data tunes hyperparameters and model choice, test data gives one final unbiased estimate. Overfitting = strong train performance, weak validation; fix with more data, regularization, simpler models, early stopping. Underfitting = poor on both; fix with richer features or a more capable model.
Metrics
Accuracy = correct / total. Precision = TP/(TP+FP) - of the items flagged, how many were right. Recall = TP/(TP+FN) - of the real positives, how many were caught. F1 balances the two. Regression uses MAE, RMSE, R-squared. ROC-AUC measures ranking quality across thresholds.
Worked example
The trap: accuracy is meaningless on imbalanced data. 1,000 transactions, 10 fraudulent; the model flags 50 and 8 of those are truly fraud. TP=8, FP=42, FN=2, TN=958. Accuracy = 96.6%, which sounds excellent, but precision = 8/50 = 16% and recall = 8/10 = 80%. Ask which error is more expensive before choosing a metric.
Rule of thumb
Fraud, disease and defect detection usually prioritize recall; spam filters and content takedowns usually prioritize precision. Monitoring for data drift after deployment is part of the ML lifecycle, not an optional extra.