Keras

Keras is a high-level neural network API written in Python and capable of running on top of TensorFlow, Microsoft Cognitive Toolkit, or Theano. Originally developed by Francois Chollet as an independent project, Keras was designed around the principle of user-centric API design: machine learning code should be readable, modular, and extensible, with each layer of abstraction solving exactly one problem. In 2019, Keras was officially integrated into TensorFlow 2.0 as its default high-level interface, making the separation between Keras and TensorFlow largely nominal for most practitioners.

Keras represents a recurrent pattern in software ecosystems: the simplification layer that eventually becomes indistinguishable from the system it simplified. When Keras was independent, it offered genuine portability across backends. As TensorFlow's frontend, it offers syntactic sugar over a single backend — sugar that has become so dominant that many programmers write Keras code without knowing they are using TensorFlow underneath. The abstraction has succeeded by disappearing.

The Keras API has been widely imitated, with PyTorch's nn.Module, JAX's Flax, and even TensorFlow's own Estimator API all converging on similar patterns: sequential or functional model construction, layer composition, and callback-driven training loops. This convergence suggests that Keras did not merely provide a good API — it established the mental model for how neural network code should be structured.