● training flower■ test flower◯ ring = misclassified— line = decision boundary
Fisher's 1936 Iris dataset: 150 flowers, 50 from each of three species, with four measurements each. Pick two measurements, and each classifier learns its own way to divide the plane into species regions. The shaded map shows what the model would predict for a flower at any point; stronger color means more confidence. Circles are training flowers, squares are held-out test flowers, and a ring marks a mistake.
Each model trained with its default settings on the same split and features. Linear models cut with straight lines, trees with axis-aligned steps, neighbors with jagged local patches, Naive Bayes with curves from per-species bell curves, and neural networks with smooth bends. Click one to open it.