Fundamentals of AI and ML
Start with the hierarchy
Artificial intelligence is the broad goal of machines performing tasks that normally require human intelligence. Machine learning is one approach to that goal: algorithms that learn patterns from data instead of following hand-written rules. Deep learning is a subset of ML built on multi-layer neural networks, and generative AI is a subset of deep learning that creates new content rather than only predicting a label.
Learning paradigms
- Supervised learning uses labelled data (features plus a target). Regression predicts a continuous value such as price; classification predicts a discrete class such as spam or not-spam.
- Unsupervised learning uses unlabelled data to find structure: clustering for customer segments, dimensionality reduction for compression and visualization.
- Reinforcement learning trains an agent through reward and penalty signals by trial and error.
- Self-supervised learning manufactures labels from the data itself and underpins large language model pretraining.
Training versus inference
Training fits parameters on historical data. Inference applies a trained model to new inputs. Exam items often describe a workload and ask which phase it is: scheduled batch scoring of last month's transactions is inference, not training.
Fit and evaluation
Overfitting means strong training performance but weak performance on unseen data; underfitting means poor performance on both. Remedies include more or better data, regularization, simpler models, and cross-validation. Choose metrics by problem shape: accuracy misleads on imbalanced data; precision is the share of predicted positives that are truly positive; recall is the share of actual positives that were found; F1 balances the two; RMSE and MAE apply to regression.
Worked example and rule of thumb
A fraud model reviews transactions where 1 percent are fraudulent. A model predicting never-fraud scores 99 percent accuracy with zero recall, which is useless. Because a missed fraud costs far more than a false alarm, optimize recall or F1 rather than accuracy. Rule of thumb: when the cost of a false negative dominates, check recall first.