Humans are lazy and like to automate. Sometimes it doesn’t matter what the outcome of that automation is as long as we can sit back, stretch our legs and not worry about it. In the last decades we have asked the question wether we can leave extracting important information from data to machines.
The answer to this question is the imperfect automation that really needs no introduction: Machine Learning. Computers, as the information digesting automata they are, are wonderful data hoarders. No human can remember all of Wikipedia so why not task a computer program with the problem? As it turns out, there are reasons why we should look at this critically.
Machine learning is a strange term to me as these algorithms are not performing what we usually consider learning. Calling the computation of a statistical model learning is a weird act of anthropomorphization.1 Statistical models don’t think nor learn. So why do we claim they do?
Learning is a key part of consciousness and is understood as permanent changes in ones behavior based on past experiences. This is an important point. Learning is inherently linked with experiences and is not just a pure task of remembering facts. Also, learning does not end. It is not a process with a finite amount of steps. The result of learning does not have to be knowing the truth. It is very possible that your experiences lead you to infer incorrect information about the world. This distorted world view can be corrected later by more experiences, some contradicting the preconceptions a learner might have. The result of learning is hard to quantify. In education we sometimes try to measure the progress of a learner (students in this case) using exams. However, only a tiny subset of what makes up learning can actually be tested this way. This often comes down to evaluating the ability of reproducing knowledge or certain thinking patterns like being able to prove a mathematical statement or reason about texts with some historical context.