Keras is an open-source deep learning library that provides a high-level, human-friendly API for building and training neural networks. Created by Francois Chollet and first released in March 2015, it became one of the most widely used entry points into deep learning, prized for letting researchers and students define models in a few readable lines of code.
History
Keras began as an independent library able to run on top of multiple backends, initially Theano and later TensorFlow and Microsoft's CNTK. Its popularity grew quickly alongside the post-AlexNet deep learning wave, and in 2017 Google adopted it as the official high-level API of TensorFlow, shipping it as tf.keras. In 2023 the project returned to its multi-backend roots with Keras 3, which runs on TensorFlow, PyTorch and JAX from a single codebase, a deliberate bet against framework lock-in at a time when research had consolidated around PyTorch.
Design philosophy
Chollet described the guiding principle as "deep learning for humans": consistent and simple interfaces, sensible defaults, progressive disclosure of complexity, and clear error messages. A model is assembled from composable layers, compiled with a loss function and optimizer, and trained with a single fit call that handles gradient descent, batching, validation splits and callbacks. This abstraction made Keras a standard teaching tool, featured in influential online courses by Andrew Ng and in Chollet's own widely read textbook Deep Learning with Python.
Reception and influence
Keras has been praised for documentation quality and pedagogical clarity, and criticized in research settings for hiding too much of the training loop; power users often drop to lower-level APIs for custom architectures. Its layer-oriented ergonomics shaped later frameworks and high-level wrappers across the ecosystem, and its error-message philosophy became a reference point in discussions of developer experience for machine learning tooling. While large-scale LLM research today is dominated by PyTorch, Keras remains widely used in industry, education and applied work, and is credited with lowering the barrier to entry during deep learning's formative years.