What is human-in-the-loop machine learning? Better data, better models
Machine learning models are often far from perfect. When using model predictions for purposes that affect people’s lives, such as loan approval classification, it’s advisable for a human to review at least some of the predictions: those that have low confidence, those that are out of range, and a random sample for quality control.In addition, the lack of good tagged (annotated) data often makes supervised learning hard to bootstrap (unless you’re a professor with idle grad students, as the joke goes). One way to implement semi-supervised learning from untagged data is to have humans tag some data to seed a model, apply the high-confidence predictions of an interim model (or a transfer-learning model) to tag more data (auto-labeling), and send low-confidence predictions for human review (active learning). This process can be iterated, and in practice tends to improve from pass to pass.To read this article in full, please click here
Machine learning models are often far from perfect. When using model predictions for purposes that affect people’s lives, such as loan approval classification, it’s advisable for a human to review at least some of the predictions: those that have low confidence, those that are out of range, and a random sample for quality control.
In addition, the lack of good tagged (annotated) data often makes supervised learning hard to bootstrap (unless you’re a professor with idle grad students, as the joke goes). One way to implement semi-supervised learning from untagged data is to have humans tag some data to seed a model, apply the high-confidence predictions of an interim model (or a transfer-learning model) to tag more data (auto-labeling), and send low-confidence predictions for human review (active learning). This process can be iterated, and in practice tends to improve from pass to pass.