class of machine learning techniques in which a task is solved based on pseudo-labels which help initialize weights the weight, then the actual task is performed with supervised or unsupervised learning
Discovered by embedding cosine similarity (sentence-transformers MiniLM, 384-dim).