Introduction to Machine Learning

So what is machine learning? And how can machine learning create value for companies of the future?

To explain what machine learning (abbreviated ML) is, it is easiest to think about a simple example.

Imagine that you have been given a dice by a magician and you therefore suspect it to be biased so that it does not land on all sides equally often. However, you are curious in nature so you might want to find out how likely each side of the dice actually is. The typical way you would find such a probability distribution is to throw the dice a lot of times and then count how many times it lands on each side. After throwing many times, you will have a good approximation of the probability distribution.

Such a simple experiment is basic statistics. But in practice, there is no profit in such simple experiments, so let's consider another example.

Instead of assessing a dice with unknown distribution, you might be an actuary at an insurance company and have to assess the risk of a potential customer getting seriously ill within the next year. The simple way would be to try and do as with the dice, just to look at all previous customers and find out how many people got sick on average within the first year. If you have no information at all about the clients, this would typically be the only thing, but imagine that all potential customers for your health insurance will first have a doctor's examination. Then you want to have all sorts of parameters about the customer such as blood pressure, liver count, BMI, etc. If with all those parameters you try to look again at what the average for illness the first year is, by simply counting how many with the same blood pressure, liver count, BMI, etc. gets sick the first year - you will definitely run into a problem. The problem you run into is that with all the parameters, there will actually not be enough customers for you to have a reasonable statistical basis for all combinations of the values of the parameters.

What one typically does to solve that problem is to group the customers in a "sensible" way so that there are still enough customers in each group that you can safely believe in the probabilities of illness for each group.

And now we are finally at the point where I can tell what ML techniques are, because they are ways of making such a sensible grouping, completely automatically.

But how does the ML technique know what a sensible grouping is? It knows this by first training it with data where you already know the answers and deducing a structure from it.

The fact that this is done automatically means that in the example of potential customers for insurance where you have several different parameters from a medical examination - you will first provide data for existing clients where you know the outcome (illness within the first year) to an ML based program , and then the same program can spit out risk percentages on all your potential customers at once.

It gives you the advantage that as long as you trust your ML program to make a sensible treatment of your data, you don't have to do any analysis yourself. So you can spend your time on other parts of your work in the insurance industry.

Whether the ML techniques will do the right thing, depends very much on the details and this is where an ML expert comes into the picture. There are quite a few pitfalls to using ML techniques, so it is important to first analyze what is the best approach. However, it can then run completely automatically.

Getting your work done automatically is one of the reasons why ML techniques are gaining ground in many parts of society. Another reason is that because the ML techniques run in a computer, it allows you to scale them to areas where a person who had to perform the analysis manually would never finish.

An example of a task that people would never want to solve manually is the recognition of objects in an image by doing statistical analysis on millions of images. Even a relatively small 224x224 pixel image has 2 * 224 * 224 = 150528 pixels. If you need to analyze that many parameters across millions of images to find out which combinations, for example. gives the probability that there is a cocker spaniel in a picture - then you will probably never finish.

Nevertheless, it is a fairly common thing to do with ML techniques. And in fact, ML techniques are now just as good as humans for classifying images.

Another reason why ML techniques are so popular is that progress in the field is incredibly fast-paced. This means that the techniques are becoming more and more accurate all the time.

At the same time, efforts are being made to build ML techniques on top of each other in what is called deep learning. Deep learning provides a hierarchical structure of your ML architectire so that, for example, if you need to detect faces in an image, then a deep learning setup can automatically find out at the lowest level eg. to detect an eye, nose or mouth and then at higher levels they use detections to do a better analysis of whether there really is a face in the image there. Namely, it has proved more robust to have a hierarchical structure than if analysis had to be done directly on all pixels. With deep learning, there are pixels at one end but they gradually become larger and larger elements until the algorithm ultimately makes the binary judgment whether there is a face or not.

These deep learning techniques have proven to be very useful. In areas that have traditionally required experts to spend years devising techniques to solve a problem, deep learning can increasingly replace the work of devising the techniques and instead work directly on data and automatically make sensible breakdowns of data.

An example is speech recognition. Traditionally it has required experts in languages who have created computer programs to analyze the sounds, possibly based on knowledge of the shape of the pharynx and the function of the vocal cords. Today, you almost directly output the recorded sound from a microphone to a deep learning technique and then it automatically spits out what words were said. It does, however, require that you first train it with thousands of hours of speech with associated facts about what was said and it is not quite perfect yet - but as the development continues, there is not much doubt that the techniques in this area will soon be as good as human beings.

As more and more things can be automated, it will, as with lego bricks, provide new opportunities to build additional things on top. Eg. one must expect that as cars and trucks become self-driving, it will have a major impact on the transport sector when the trucks can run 24 hours a day and truck drivers will have to find other work.

The example of the transport sector is just one of many areas that will be severely affected in the future, and as a business owner it should be made clear that if you do not keep up with this development then there is a great risk that competitors or 2 entrepreneurs in a garage can automate parts of your business in a way that (like truck drivers) it ends up without a reason for existing.