Pattern Language for Designing Distributed AI Systems

Design of Artificial Intelligence (AI) and Machine Learning (ML) applications, hereafter referred to as AI systems, is often based on a typical ML pipeline. One of the reasons for choosing this approach is its simplicity and modularity. While simple, such an approach tends to be rigid with respect to changing needs, technologies, devices, and algorithms. Recent research on design patterns for ML has introduced best practices for engineering AI systems. We examine a set of these patterns, or a pattern language, where individually selected patterns can build on each other to offer a complete design solution for a distributed AI system. We demonstrate the use of this pattern language to design an AI system for emotion classification of social media content. The result is an AI system that is not only easy to change and reuse in a similar context, for instance emotion classification of image data, but one whose architecture has better performance, usability, maintainability, security, and reliability.