Pedro Domingos
Every machine learning breakthrough today belongs to one of five distinct tribes, but the ultimate prize is a single unified algorithm capable of deriving all past, present, and future knowledge from data.
Machine learning is divided into Symbolists, Connectionists, Evolutionaries, Bayesians, and Analogizers, each utilizing different fundamental mechanisms like logic, neural networks, genetic search, probabilistic inference, and similarity matching.
A true Master Algorithm will successfully integrate the unique strengths of all five tribes to create a universal learner capable of deriving any knowledge from data without explicit human programming.
The greatest obstacle in algorithmic learning is overfitting, which occurs when a system hallucinates nonexistent patterns from training information and consequently fails to operate accurately on new problems.