A framework for human-AI integration in the enterprise

Description

Recent developments in ML and DL have provided remarkable advances in the predictive capabilities of AI. However, the black box nature of the modern ML and DL models creates challenges for enterprises looking to adopt these techniques and has raised awareness toward human-AI hybrid systems. Three factors, namely regulations (e.g., GDPR), risk (e.g., business criticality), and cost, determine if and how an enterprise product should integrate human and machine intelligence.

Bahman Bahmani details the shortcomings of existing machine learning interpretability approaches and provides architectural blueprints (e.g., parallel versus serial patterns) for human-AI integration. Bahman also introduces the notion of modifiability, a particular form of debuggability, where human operators can not only audit but also modify the behavior of a machine learning system, irrespective of how (e.g., due to which combination of input features) the system decided on its course of behavior. You’ll learn three design ideals for modifiability in critical applications, namely immediacy, determinism, and simplicity (of the modifications), as well as the role of user interfaces (for the humans in the loop) in achieving these design goals.

You’ll see examples and case studies from real-world, operational, large-scale AI systems for applications such as product categorization, information extraction from unstructured documents, and online ad auctioning, and you’ll gain an actionable framework for how to approach human-AI integration for your own enterprise workflows, products, and services.