Talk for TWIMLCon 2022. Abstract It’s hard enough to train and deploy a machine learning model to make real-time predictions. By the time a model’s out the door, most of us would rather move on to the next model. And maybe that is what most of us do, until a couple months or years pass and the original model’s performance has steadily decayed over time. The simplest way to maintain a model’s performance is to retrain the model on fresh data, but automating this process is nontrivial.
