The Value of Humans + Machines
July 1st, 2020
Comment to help prioritize content.
Comments not loaded in development.
For complex problems that require quick decisions, some people believe that Smart Automation is not fast enough. They think that because models sometimes take a bit of time to be trained and produce inferences they won't be able to produce sub-second recommendations. That is not correct, and this post we go through a design that can provide very fast decisions in real-time, with the option to receive even better decisions if the system can wait for a bit longer. In one of our courses we show how to actually build this system.
The standard way of training and deploying Machine Learning (ML) models is with centralized architectures. This means that there are centralized servers, that receive requests from devices and respond with ML predictions. These systems have the disadvantange, as with any other centralized system, that they can be overloaded and that they must have central access to all data. In this course we show how distributed ML systems can be trained and deployed in a way that eliminates some of the problems of centralized architectures.