Machine Learning vs Predictive Modeling

Is Machine Learning different from Predictive Modeling

The above question seems to haunt most people who have been doing statistical predictive modeling before the term machine learning came into play.

Nowadays it seems whoever have run a classification problem with any of the advanced algorithms like Neural Network, Support Vector Machine, etc. calls themselves a machine learning expert. But is this machine learning? We have this statistical/mathematical models from the early 60's .  And the only reason not all of them was popular back then is because it was too advanced for the computing power available at those times.

So what is truly machine learning? Remember Skynet from Terminator, Terminator himself/herself, Vicky from iRobots,etc. well those are the achievements machine learning scientists often want to achieve however bad they are depicted in the movies. A machine that can learn on its own, which will have a feedback mechanism for it to upgrade it's knowledge is the pinnacle of machine learning.For this reason Microsoft's chat bot Tay was such a good example of machine learning. The chat bot was exposed to the real world and was given a chance to learn on it's own. However because of some group of people who targeted the chat bot to learn more offensive staff than usual, its personality became something unexpected, in less than 24 hrs. This also shows the other side of machine learning, (which by the way almost every movie showcasing an AI ends up with). A machine can learn on it's own, but like a human child it need human parenting to differentiate the right from the wrong, and then, and only then it'll grow to become a truly independent machine who can think on its own.

But where does predictive analytics fits in. Well it fits right at the very beginning. The path to modern machine learning started when Ordinary Least Squares were discovered and we found that based on known/historical data we can predict the unknown. But everyone working in the field knows that you cannot just build a model and be happy about it, without being sure that it's working. Hence statisticians/data science professionals developed several techniques to validate that their models work. And post validation once the models were put to use they were continuously monitored and modified when shifts in pattern happen in the data. This is and always have been what we called predictive analytics. And machine learning uses this entire manual effort as a stepping stone and let a machine do the above tasks. With the emergence of ensemble models and various other techniques as well as computing power, it is now possible to create feedback loops, that will continuously update a machine that whether the decisions it is making based on history is right or wrong, and if it's going down a wrong path, then can it rectify itself or it needs to inform it's human parents. So when we build and deploy a system that uses historical data be it actual data, or human behavioral pattern, etc. and expects a machine to work on its own making the above right or wrong decision we enter the domain of machine learning.

And when we can deploy a machine which uses multiple such algorithms which is taking decisions to do different jobs, we are really building machines that is way closer to where machine learning should be. Take google car for example, I do not know the intricate details, but from what I read, they have multiple sensors on their car that collect data, and replicates safe human drivers driving pattern and then with help of feedback loop improves upon it. That is where machine learning should be, inching closer and closer to AI.