
…them with corresponding emotions. The data that engineers commonly use to train algorithms on how to make those connections include images of people posing expressions that have been annotated with emotion labels like “happy” and “sad” by a trained third party. To create training sets for the algorithms, companies have annotators review a collection of images and label them as “happy,” “fear,” “anger,” etc. Amazon declined to comment on the record about how Rekognition’s algorithm has been trained, citing proprietary concerns.