In this talk, we look at the deliberation process that helped explain “Why Swift for TensorFlow?”. The choice was guided by the goals of the project, which imposed specific technical requirements which we cover in this talk. We defined goals around the properties that are important to maintain and improve in the system: Expressiveness, Performance Predictability, Fast Iteration Time, Debuggability and Introspection, Flexible Deployment, Fast Deployment, Best-of-class Automatic Differentiation (AD), Embrace TensorFlow Graph Ecosystem.
A cornerstone of the design is an algorithm that is called Graph Program Extraction, which allows you to write in an eager execution-style programming model while retaining all of the benefits of graphs. The design also includes support for advanced automatic differentiation built directly into Swift.
Following the theoretical section, we will do a Swift model training walkthrough.
By the end of the talk, the audience will have a good overview of what is Machine Learning, how TensorFlow works and why Swift for TensorFlow is the best way to train a model.
This talk is targeted at Swift programmers with no or little machine learning experience.