Traditionally, machine learning has been “supervised”. The machine is shown a picture of an apple, and told that it’s an apple, or is told a certain language is english by showing many example words, or that a certain audio signal is music and some other speech. However, since we want the machine to learn increasing number of concepts in various areas, providing examples of each is too much spoon feeding! More over, machines inherently being dull require a lot of examples, say 50 to 100 to really understand what an apple is, and yet apples are still quite symmetric. Deformable objects like people or animals are even harder.
This has led to so many types of learning, and in the coming months I expect to write a short note on each. The technical depth will obviously increase, and most examples will pertain to images and vision. Below is an expected list:
Supervised learning; Unsupervised learning; Semi-supervised learning; Weakly supervised learning; Multiple label learning; Multiple instance learning; Multi-instance multi-label learning; Ambiguously labeled data learning; Probably Approximately Correct (PAC) learning; Transfer learning; Zero-shot learning; One-shot learning; Curriculum learning; Active learning; Proactive learning; Metric learning; Structured learning; Reinforcement learning; Consistency learning.
I am pretty sure to have missed many more learning techniques out there, let me know!