To tell cats from dogs, one can learn what separates them or learn what each is. The first draws a boundary; the second models the things themselves. The choice shapes what a system can and cannot do.
The intuition
A discriminative model learns the boundary between categories directly. A generative model learns how each category’s data is produced. One asks which class; the other asks how the data came to be.
The structure
The mechanism differs accordingly. Discriminative methods focus effort on the decision, often winning at classification. Generative methods capture more, able to synthesize and to reason about missing parts.
The subtlety
The turn is that modeling more can cost sharpness. Learning the full data distribution is harder than learning a boundary, so generative models may classify less accurately even as they understand more. Breadth trades against focus.
The price
Each shines in its place. Where the task is decision, discrimination suffices; where the task is imagination, explanation, or handling the incomplete, generation is needed. Purpose selects the approach.
The boundary
The implication is that understanding and deciding are different aims. To act one may need only a boundary; to comprehend one needs a model of the world. The richer knowledge is not always the useful one.
The larger point
Discriminative models learn boundaries, generative models learn how data arise, and the choice trades classification sharpness against explanatory breadth. Deciding and understanding are distinct aims. The richer model is not always the more useful. Seen this way, the concept is less a fact to be filed than a lens through which other facts arrange themselves. Its worth lies not in any single application but in the pattern of thought it makes available. To hold it clearly is to see a whole family of problems as variations on one theme, and to recognize the same shape recurring where it was not expected.
