Wednesday, August 10, 2016

Machine Learning: The Master Algorithm


Most of the knowledge in the world in the future is going to be extracted by machines and will reside in machines.
– Yann Le Cun, Director of AI Research, Facebook

MACHINE LEARNING is the automated discovery of Knowledge. With the Internet and sensors everywhere flooding the world with data, the ground for machine learning is extremely fertile. Machine Learning is done by learning algorithms.

Pedro Domingos’ book is a discourse on the quest for the Master Algorithm which is able to learn anything and everything anytime.

However, the quest for a Master Algorithm is by default a quest to know the nature of the learning process, and there are five schools of thought here on just what happens when we are learning so that machines can emulate them. (Dominos call’s them the five tribes of machine learning). 

Each of the tribes have their strong and valid points and any master algorithm would have to take a bit from each tribe.  them. Below, the five tribes of machine learning, and what they do:

1.    Symbolists fill in gaps in existing knowledge and use deduction to synthesize new knowledge. Example tools: Algebra, Set Theory, Boolean Logic.

2.    Connectionists strive to emulate the brain. Example tools: Back propagation artificial neural networks, Deep-Learning neural networks.

3.    Evolutionaries simulate evolution and natural selection to search the data space and evolve generations of virtual chromosomes of knowledge using a fitness function as a selection criterion. Example tools: Genetic Algorithm, Particle Swarm and Ant Colony Optimization.

4.    Bayesians systematically reduce uncertainty through changes in the probability of constantly updated data. Example tools: Markov Chain Monte Carlo simulation, Kalman Filter.

5.    Analogizers strive to find patterns and similarities in data to classify them and reduce noise. Example tools: Support Vector Machine, Nearest Neighbour algorithm and Radial Basis functions.

Each of the above is suitable for certain applications in widely different fields and certainly, it would be a challenge to find a Master Algorithm.

No comments:

Post a Comment