Interview
François Chollet: History of Keras and TensorFlow | AI Podcast Clips
- Deep learning tooling is expected to mature as the practitioner base expands from under 10,000 in 2015, with recurrent neural networks predicted to gain prominence around 2016 following the March 2015 release of Keras.
- Plans include developing a Python-based library using code-defined models for improved usability comparable to scikit-learn, featuring a refactored codebase completed in December 2015 that supports both TensorFlow and Theano backends.
- Theano is projected to serve as the initial default due to its ease of use, stability, and speed, though it is anticipated to eventually be superseded by TensorFlow.
- A transition from a side project to a primary role is scheduled to occur after October 2016, facilitating tight integration of Keras into TensorFlow core.
- A sprint toward the conclusion of TensorFlow 2.0 is anticipated, though its specific duration remains uncertain.
- TensorFlow 2.1 is expected to incorporate eager execution to simplify framework usage, while version 2.0 aims to resolve the previous trade-off between usability and flexibility.
- The updated framework is designed to unify higher-level features with lower-level interfaces, offering a workflow spectrum ranging from high-level Keras usage to low-level custom training loops suitable for both researchers and machine learning engineers.