Machine learning is building programs that get their behaviour from data instead of hand-written rules: you show examples, a method fits a model, and the model makes predictions on cases it has not seen. It matters wherever the rules are too many or too fuzzy to write down, such as spam, prices, faults or recommendations.
Key concepts
- Supervised learning: learn a mapping from inputs to known labels, for classification or regression.
- Unsupervised learning: find structure without labels, such as clusters, low-dimensional views or anomalies.
- Generalisation: a model is good if it works on new data; overfitting is memorising the training set.
- Train, validation and test splits: tune on one, report on a test set you touched once.
- Features and leakage: what the model sees, and the trap of letting information from the future or the label slip in.
Vocabulary
- Loss function: the number the training procedure tries to make small.
- Hyperparameter: a setting chosen before training, such as tree depth or regularisation strength.
- Regularisation: a penalty or constraint that keeps a model from fitting noise.
- Cross-validation: repeating the train and validate split over several folds to estimate error.
What to study, in order
- Linear algebra, probability and a little calculus: vectors, distributions, gradients.
- Linear and logistic regression by hand, with gradient descent in plain NumPy.
- Decision trees and ensembles of trees with scikit-learn on a small tabular set; always run a baseline first.
- Evaluation done properly: confusion matrix, precision and recall, calibration, cross-validation.
- Build an end-to-end project: clean data, split it, model, error analysis, and a write-up of what failed.
Related tags
#probability #data-science #deep-learning #ai-evaluation #optimization