Machine Learning is probably one of the most commonly used words in most industry sectors these days. But, the notion of machine learning has been around for a while (longer than my current lifetime). So what is it?
Wikipedia defines it as
a branch of artificial intelligence, concerns the construction and study of systems that can learn from data.
Probably not a suitable definition according to most academics since Wikipedia is not a legitimate source to them but should be good for the rest of the world. A more academic definition: a field of study that gives computers the ability to learn without being explicitly programmed (Samuel, 1959).
Machine Learning is used in a multitude of industries from the information technology to healthcare and financial markets to space and robotics. It initially grew out of Artificial Intelligence work which gave computers a new capability. Current examples include database mining of large datasets from web click data on social networks and search engines, predicting correlations between a disease and symptomatic factors using medical records, recognising handwriting on an app like the new Z Launcher app by Nokia, autonomous drones, computer vision algorithms to help computers see using vision based sensors, natural language processing (speech recognition), self-customising programs seen through product recommendations on on-line stores and social networks, monitoring spacecraft health [ an exciting area — trust me ;-) ] to predict its performance and maintain it at a certain threshold, from space radiation and possibly the most hyped out of them all from science fiction movies understanding human learning (giving machines the ability to learn).
Teaching a person how to learn something can be done in two main ways. The first way would be to tell the person what the right answer is to a question and from then on, they search for answers similar to that for similar challenges. And the other way is to not tell them anything and let them figure it out themselves. Similarly, in machine learning there are two main types of which algorithms are derived from namely, supervised and unsupervised learning. Supervised is like the former where the right answer is provided and the machine is left to learn in future similar challenges. Unsupervised is like the latter, where the machine is left to figure out the answer by itself. In both cases, data is required. Larger data sets provide better awareness of a possible solution. There are others to these two main types (reinforcement learning, recommender systems, developmental learning, etc.) but they have not been mentioned here since, this is a brief intro.
A property in learning algorithms is the category of challenge encountered based on predicting the output value. Huh? What I’m trying to say is that predicting answers for machines can either be continuous (taking into account time or have some flow) or discrete (more like 1s and 0s discrete). For example, as a child I used to fight for my parents attention with my twin brothers. I noticed that the price of my parents attention was trying to be like my brothers in every way possible. I had to be and act like a child (drinking from the bottle, child prams at the age of four — embarrassing) taking into account all factors that made me like them. And within that year I managed to get my parents attention to the level they did. This is similar to predicting continuous output because I'm trying to predict the price of my parents’ attention based on factors that classify me as being cute, needy, etc. etc. that kids are all about. This is a regressive form of predicting or Regression. A more discrete output prediction is like trying to predict whether I would get beaten up or not by my brothers if they figured out I was turning them against each other, muahahahaha !!! This form of predicting output values is called Classification, since you’re classifying your outputs as a yes or no (like a Boolean: True or False — but can be more than two classifications). Not sure if I did these examples well so I'm always open to suggestions.
Anyway, I think that should do it for a brief introduction. More to come about machine learning algorithms in the future. And I promise more pictures ‘cause I know hard it can be without illustrations to understand anything in life.
This is a very old article originally written in 2012 during my research project which involved developing a supervised learning (off-line and online) platform to perform predictive analytics on spacecraft telemetry.