Four principles, one decade
Keras is built on four core principles that have guided its development for a decade — and they aren't a wiki manifesto bolted on after the fact, they're baked into individual API decisions:
- User-friendliness — Designed for humans, not machines. The API minimizes the number of user actions required for common use cases and provides clear error messages. As Chollet puts it: "Keras follows best practices for reducing cognitive load."
- Modularity — A model is a sequence or graph of standalone, fully configurable modules. Layers, optimizers, loss functions, and regularizers are all independent building blocks that can be combined with as few restrictions as possible.
- Extensibility — New modules are easy to add (as new classes and functions), and existing modules provide ample examples. The ability to easily create new components makes Keras suitable for advanced research.
- Minimalism — Just enough to get the job done, no more. Each module is kept small and simple. Source code is readable. No unnecessary complexity lurking beneath the surface.
These four explain why Keras consistently ranks among the most-loved ML frameworks in developer surveys: the API is predictable. Once you learn the patterns, everything else follows logically — and the rest of this quest is largely you watching the same four ideas reappear, in layers, in training loops, in serialization.