While the potential of Machine Learning (ML) and Artificial Intelligence (AI) is widely recognized in various sectors (health, industry, commerce, etc.), regretfully many ML/AI projects do not make it past the Proof of Concept (PoC) stage. In this talk I will share a number of my own experiences with “failed” AI projects (i.e., projects that easily passed the PoC stage, but never made it into production), and I will examine the root causes of these failed AI projects. To do so, I will have to provide a bit more background regarding the various types of AI models/projects that exists, explain how AI works in some detail, and discuss the common production/deployment patterns that companies use in their attempts to scale their ML/AI activities. Effectively, I will describe the AI deployment process from data collection, to AI model development, to model evaluation, and finally towards large scale model deployment. At each of these stages I will highlight the challenges involved and the common points of failure.

Next, I will turn my talk to potential solutions: although it is hard to find a uniform solution for scaling every possible ML/AI solution in the book, for a large class of applied AI/ML models efficient and effective deployment methods have recently been developed. I will explain how deploying AI models on the edge (i.e., not in the cloud) solves a number of common AI deployment problems. Furthermore, I will explain how modern technological advances enable the effective deployment of trained ML/AI models on edge devices despite the diversity in device types (e.g., different hardware, different computational constraints). Finally, I will argue that deployment on the edge makes applied AI more scalable, reduces the energy footprint of AI, improves user privacy, and reduces operational costs of many AI applications.