Accurate and scalable approximate nearest neighbor search (anns)-based training of extreme classifiers
Topics: Search Intent, Shopping
This Microsoft patent describes a method for training “extreme classifiers,” which are machine learning models designed to categorize data into millions or even hundreds of millions of different labels. The system improves training speed and accuracy by using a two-stage approach: it begins by using simple random labels to teach the model basics, then introduces “hard negative labels” identified through an Approximate Nearest Neighbor Search (ANNS) index to fine-tune the model’s ability to distinguish between very similar items.
