WikiMath

This presentation of automatic learning techniques is complemented by three other types of learning.

By Reinforcement

Reinforcement learning focuses on optimization issues. It consists of applying rules to improve the current best score.

The learning system (the agent) observes the environment, selects and performs actions, and in return obtains rewards (or penalties). He must then learn for himself what is the best (political) strategy to obtain the most rewards possible.

Examples:

• Q-learning, Monte-Carlo, SARS, and neural networks.
• robots that learn to walk.
• AlphaGo who, after analyzing millions of games, learned by playing against himself.

By transfer

Transfer learning is the reuse of previously acquired knowledge in order to reapply it in another environment.

ex: Use an RNN that recognizes French handwriting, for Japanese.

Semi-supervised

It is a learning on a lot of data without a label, and some with it. Most of the time, these are combinations of unsupervised and supervised algorithms.

For example, Deep Belief Networks (DBN) rely on unsupervised components (Boltzmann Machines) stacked on top of each other. These machines are trained sequentially in unsupervised mode, then the complete set is precisely adjusted using supervised learning techniques. Or, Google Photo clusters individuals and recognizes them from a few names (with redundancy).