on 19 December 18
Machine learning technologies have advanced tremendously in recent years. One of the commonly cited examples is IBM's Watson, which became famous when it appeared on the hit TV game show Jeopardy! in 2011 and won. After that it was used in the healthcare industry with great success, and more recently, it became the underlying brain behind IBM's predictive analytics solution, Watson Analytics. Machine learning is now past the peak of expectations in Gartner's Hype Cycle for 2015.
So what is machine learning and why is its application growing so fast? Machine learning falls within the ambit of the larger field of artificial intelligence. Artificial intelligence, in general, deals with carving out specific patterns or conclusions and making decisions based on a broad set of fuzzy inputs or data for which context is not very clearly presented. It represents science's best efforts to mimic the way human intelligence works, although it's still a long way off from completely achieving the same level of sophistication. Traditional computer science works with discrete logic where everything is either true or false, black and white. Artificial intelligence attempts to enter the gray areas, and machine learning is the branch of artificial learning that is used to help navigate these gray areas by driving decision making when there's not enough clarity to begin with, and improving on successive decisions by learning from the previous ones.
Computer scientists have been working on improving artificial intelligence and machine learning technologies for decades, and now there are a number of reasonably mature machine learning platforms available commercially. It's good timing, too, because the amount of data that is now available from devices, social media and Big Data is vast and machine learning has become one way of making sense of some of it. When enough data is available, whether past or current, machine learning can be applied either to make decisions by making predictions or by uncovering patterns that are not readily identifiable when there are a lot of variables.
Learning is a complex process, and consists of a wide range of techniques and influences on which an assimilation of significant understanding is based. For example, consider a person walking through a city. How would that person know when he/she might be approaching a dangerous or crime-ridden area? The easy way is by looking at past statistics of crimes, types of crimes and specific locations where they occurred, and using that as a guide to be more alert when passing through those areas. In the absence of such data, ie, if placed in a new city for the first time, and without any data whatsoever, or any understanding of what a crime-ridden area might look like, the person might just walk and make observations. They may notice that people were injured when gunshots were heard at the same time, or that screams happen at the same time as a negative incident. Or that screams, gunshots and injuries happening together were much more indicative of an incident of crime than any one or two of them alone. When these observations are made, some may be validated, some may be invalidated, and that forms a learning, or a basis for safer actions and reactions as the person continues to walk.
Machine learning works in a similar way. The first kind, in which previously analysed past data is available to learn from is called supervised learning, and the second, in which there is data, but associations have to be made by identifying patterns, is called unsupervised learning. Supervised learning algorithms are used primarily to make predictions when knowledge is available from the past. Unsupervised learning is more useful when there is no starting guide, or to uncover new patterns or insights that were not apparent before. The underlying point in both of these approaches is that decision making in the absence of pre-programmed rules has to be done on the basis of inferences that are based on the analysis of data.
Machine learning may seem similar to predictive analytics, but they are different in that the latter is purely a statistical technique based on chosen variables used to validate a hypothesis or predict the future based on the past, whereas machine learning actually provides a recommendation or makes a decision using logic formed from observation of all kinds of variables without any starting biases. Going by this perspective, machine learning may even be used in place of statistical analysis as an implementation of predictive analytics.