Machine Learning Myths

Machine Learning has gone through multiple waves of its adoption. Over the years data availability has increased exponentially. At the same time computer power has multiplied according to Moore’s law creating multitude of opportunities for machine learning.

Machine Learning is going through rapid evolution from basic machine learning ( hard wired techniques which came initially from statistics) to advanced machine learning to deep machine learning and the whole umbrella of algorithms in Artificial Intelligence covering not only learning, but problem representations, complex state-space models, reasoning, perception, thoughts, emotions and all the way up-to theorem proving, problem solving, search to advanced tasks like planning. Every step in right direction has made overall machine learning algorithms as one of the most talked phenomenon only with ML Scientists but practitioners as well as lay people.

I would like to elaborate on some of the myths about Machine Learning.

1. ML can learn anything

Although to a large extent this true, but not in the sense of the way people are trying to use them as blackboxes to throw them at large set of varied problem situations or business use case. In next 20 years or so indeed present day ANI ( Artificial narrow intelligence) to AGI ( Artificial General Intelligence) to ASI ( Ariticial Super Intelligence). The real ASI will surpass all human capabilities and then it can be safely said that ML can learn anything. This statement alone will move through its own Darwinian Evolution.

2. ML is simply curve fitting !

Indeed some of the basic ML algorithms started with curve fitting and basic error minimization. Present day ML algorithms ( not just the state-of-the-art, but even basic) have many sophisticated techniques deployed making them fairly complex. They have lot of parameters which themselves can “learn” depending on the new data, information and situations thrown at them. Training and Reinforcement Learning ( un-supervised) makes algorithm really intelligent in human natural intelligence sense and computer sense( ability to crunch large amount of data and infer from them).

3. ML is simply as intelligent the programmer who writes them, per se ML has no intelligence !

This was true to the GOFAI algorithms and techniques. These GOFAI algorithms were purpose built, in a way hard wired and many case they were non-changing. Implicit in their functioning was assumption that data will come in certain pre-defined format, if it comes that way predictions would be accurate.

Present day ever learning algorithms have inbuilt “Thought Scaffoldings” like human where they will do better every time a new situation is thrown at them. They are almost like “Natural Selection” as proposed by Charles Darwin’s “Origin of Species”, chapter 4. New age ML algorithms do have capability to learn at multiple levels starting from some hardwired rules to well defined stimulus-response actions. They also go the classical Popperian way of “look before you leap”, which means they have their own model of outside environment, they do extensive “Generate-and-test” on broad range of scenarios and then go into real life to come up with flying colors.

4. ML requires large amount of Data !

This statement is true in certain ML Business Use cases, but for ML as general field of study and practice this is not true. Innate ability of certain algorithms like Deep-Q Learning, Temporal Difference Learning and Actor Critic Model for Reinforcement Learning requires minimum amount of data to start with. Every action-response pair transforms the internal model of agent to the next level and makes improvement in the internal “thinking” machine.

5. ML requires accurate data and like computers they are too GIGO ( Garbage in-Garbage Out) !

Accuracy of data is a celebrated set of problems in ML. ML has capability to refine, cleanse and come up with data that is minimal for Reinforcement Learning algorithms. Large proportion of algorithms are designed in a manner that independent of the data accuracy they will still perform reasonably well by taking into account non-number cues like natural language and other non-data information from environment.