The Need for Dedicated Programming Languages in Machine Learning

This article analyzes the intersection of machine learning (ML) and programming language design, questioning whether a purpose-built language for ML is necessary. It critiques current approaches where Python APIs mask underlying domain-specific languages, such as TensorFlow’s graph-based system, which creates complexity through metaprogramming and dual semantics. While libraries like TensorFlow offer performance optimizations, they often lack flexibility for complex models involving recursion or control flow. Conversely, define-by-run frameworks like Chainer improve usability but struggle with Python’s inherent performance limitations. The text highlights hybrid solutions like MXNet’s Gluon but notes their architectural ambiguities. The authors argue that future ML systems require languages that balance generality and performance, supporting features like automatic differentiation, parallelism, and probabilistic computation within a unified syntax. Ideally, these languages would utilize programmable semantics and flexible abstract syntax trees to handle both static and dynamic behaviors, bridging the gap between high-level research flexibility and low-level deployment efficiency.