Computers are changing our lives in more ways than we know. They have an idea of the type of person that we are and try to sell to that person. You must manage how the machines see you…. and The Master Algorithm by Pedro Domingos makes you keenly aware of this. It is basically the short history of machine learning. He does a thorough but not overly technical review of the basic types of machine learning. It does seem a little deeper than I like from time to time, but overall it was a good rundown.
The overall summary of purpose is to help people understand these algorithms that they use every day. It’s important to know what type of learner is behind specific things we use so we can identify their possible shortcomings. The man that doesn’t understand the technologies that guide him is like a man with a car and no sense how to drive it.
Machine learning has infiltrated our lives in many ways that people don’t typically realize. From where beer is placed inside Wal-Marts (basic algorithm discover people buy diapers and beer at the same time), to your friends on Facebook, to the types of advertisements you see. It is important to understand how these algorithms see you, and modify your searches and online activity to paint the picture you would like these algorithms to see and what to pitch you.
A major focus of the different types of learning focused on the shortcomings of the current algorithms. I’m not going into each of the different types of learning algorithms; connectionists, evolutionaries, symbolists, Bayesian, anologizers, and self recursive learners. However, many of the models shortcomings fall around Hume’s Problem of Induction in which we can only generalize so far from current observations/information before that information can be falsified. All swans were thought to be white until black swans were found in Australia…. But it seems that it must be this way. We must provide some guidance and generalizations for our current models to work. They key is then that the models must be subjected to the scientific method. Models fail when not tested and tweaked. No models are infallible. Their outputs must be tested and the models modified for better future performance.
This leads to the authors point that computers are just idiot savants. They are very good at limited tasks at this point. They tend to miss generalizations and overfit the data. Feedback and modification is crucial to refining these systems.
The author then concludes with some main points. First, it won’t be man vs. machine. It will be man with knowledge of machines vs. man without knowledge of the machines. I think this is accurate in the near term. I agree with this general premise, but only in the near term. If we do create “the master algorithm” it will be the machine with all of the power…
I do not agree with the author’s second point. He believes that the rise of algorithms and robots will lead to a safer world where human casualties from conflict will go down or be nonexistant. His thoughts are that we are going to have robot on robot fighting only. This seems rediculous to me. There is no reason to believe that robots will battle only other robots or that there will be some sort of mutual determent because of both sides having super weapons.
The final point is that we need to draw from not one type of learner but many, emulating to some extent on what people do. We are not one static type of learner, we shouldn’t expect the master algorithm to have one type of learning to become super intelligent. At the same time, we need to be very humble in what we don’t know. The fact that we seem to have discovered that there are no known physical laws restricting creating a superintelligence or master algorithm means that it will more than likely happen some time in the future if a major catastrophe doesn’t wipe us out first. This is a scary thing. Many smart people are working on what creating the master algorithm might mean. But again, our knowledge is very limited, and I think that like the author most people are far too optimistic about how this may turn out. There are major challenges ahead. Those that are competing in this arms race show no signs of slowing down. More people need to be focused on the crucial issues of ethics, governance, and control. It sounds rather cliche, but life may depend on getting these answers right in advance.
Deep Thoughts 