Swift: Google's bet on differentiable programming

differentiable-programmingswifttensorflowgoogledeep-learning

Abstraction: Google's Swift for TensorFlow project integrating native differentiable programming

Key points:

Swift for TensorFlow (S4TF) adds first-class, language-level gradient(of:) function — no library imports required, applicable to arbitrary Swift code
Swift chosen over Julia, Go, Rust, C++, and a new language because it balances Python-like ergonomics with near-C performance (25× faster than Python; achievable C speed with pointers)
@derivative attribute allows specifying derivatives for external C library functions, enabling seamless backprop through non-Swift code
valueWithGradient computes both function output and gradient in one call; callAsFunction enables PyTorch/Keras-style model invocation
Python interoperability via PythonObject type lets Swift call matplotlib, NumPy, etc. directly
Project remained experimental as of 2020: TF Eager backend caused slowness; key contributors (Chris Lattner, others) departed; Apple and Google aligning on broader Swift ecosystem goals

Connections: Google · Tensorflow · Andrej Karpathy · Fast AI · Differentiable Programming · Automatic Differentiation