Most machine learning is based on using large volumes of human-produced data. This is even true of the way most unsupervised learning is used, as the data comes from previous human activity or data collection, such as medical records. However, as models get ever larger, the available training data becomes a major bottleneck.. Self-learning systems seek to go beyond the avaiable data in soem cases the AI system gathers its own data through experimentation, as in reinforcement learning, or where AI sysems are pitted against each other, as found in self play and adversarial learning. These processes may be bootstrapped using initial human-created data (as in AlphaGo or work entirely from scratch using some form of specification of the problems (as in AlphaZero).
Defined on page 157
Used on page 157