In his view, a Master Algorithm could be defined as a “unified theory” of Machine Learning. In other words, it would be a model composed of key features of the main Machine Learning algorithmic families, capable of solving any classic Machine Learning task as well as unsolved problems (“from building domestic robots to curing cancer”) with a single method.
It is worth noting that creating a single algorithm that seamlessly integrates features from the Machine Learning “tribes” identified by Domingos (symbolic methods, neural network-based techniques, evolutionary algorithms, Bayesian methods and analogy ones) would be an arduous task (if possible at all), due to the fundamental differences between these different groups of methods. Even if it was feasible to create a hybrid algorithm from the five families, there is no guarantee that the results obtained with it would overperform the sum of its parts: using each specialized method for the tasks where it is more suitable might be more effective than always applying the hybrid technique.
Since the Master Algorithm definition proposed by Pedro Domingos is deeply based on Machine Learning components and holds the goal of solving Machine Learning problems, we believe that it is accurate to call it a Machine Learning Master Algorithm. Such a specification can help to distinguish it from other proposed methods that claim to solve a wide variety of problems with a single method, but not necessarily Machine Learning-focused tasks.
We believe that the kind of Master Algorithm that researchers should try to define and implement is one capable of performing genuine learning, understanding and inference; a method whose level of generality is only paired with that of humans, among other animals. We name such a technique as the General Intelligence Master Algorithm. We believe that combining Machine Learning techniques is not the appropriate way to reach the goal of generalism, since many of the Machine Learning methods are based on procedures and disciplines (such as statistics) that might seem intelligent by solving narrow tasks, but (up to now) have failed to prove being intelligent in a robust, general way.
Natural neuron-based intelligence is the only general intelligence that we know of. Hence, we believe that mimicking the precortical and neocortical mechanisms of the brain is a stepping stone to build a General Intelligence Master Algorithm.
We think that understanding the core principles of the brain and implementing a computational analogue of them, even if approximate (without exact neuro-simulation), is the most promising path to reach the ambitious goal of synthesizing general intelligence.
