I have recently ordered a lot of Neil Gailman books from a popular e-commerce website.  To my surprise, some delightful solar-powered lights were added to my order.  I was astounded, surprised, and amazed that secretly I always wanted those types of light for my Christmas tree, but had yet to act. Then, they were just delivered!

It is magic

If you have yet to be familiar with Neil Gailman, then you will discover the magical and quirky worlds he depicts in his novels. You will find this website http://journal.neilgaiman.com .

 

Now, the magical worlds based on Northern gods, contemporary culture, and other observations have some elements of modern fairy tales.  It is feasible to imagine the people who enjoy these types of novels may like the effect of the lights especially, if they lit by themselves as soon as it is dark.  It may bring some magic into their lives. Is it suitable to make the assumptions, that every Neil Gailman readers are likely to like these lights?

Statistics and probabilities

I do not see thing such abuse of statistics and probabilities should be made by any machine learning systems to promote their products. The use of a large amount of data to create some recommender should only be used without the knowledge of the  of the purchasers. Then, it is imaginable our purchasing habit may start reducing our access to health services or even some political views. This suggestion is a step closer than reality in China, for example. This article is quite enlightening (here).

Fairy machine learning

If we go back in history, fairies and magical powers were just a tool to explain the unexplainable.  Whether machine learning is deep, it has become tools that perhaps not many of its users do not fully understand the concepts behind. For that reason,  the validation of the outcome relies on the interpretation of the users.   Not only the users’ bias can change the conclusion, but also their choice of training data too.  The black box system has magically provided some answers, so it must be true.

Fairy books and fairy lights are reasonably harmless. However, what about suggestions related to health and engineering structure. Like any tools those need to be explained how they work and why they are suitable tools to use.

Validation and explanations

In my research, I was kindly given some NP-hard libraries. The latter had some specific operators and instances.  During the validation phase, I was satisfied with my results, as they were comparable with some published work with that library.

When I validated my results against the state-of-the-art, my world collapsed. For the first problem, the results were many thousand units away from the known minima.  I checked over and over again, my machine learning techniques (i.e., iterative CGP). Then, I further tested the problem library and found the source of my problem. Some of the code would not run at all and minimize the solution enough. The implementation of well-known operations was utterly incorrect.

For the second problem, it became clear the machine learning technique was not suitable. So, I had to develop auto-constructive CGP to learn some mutation operator to find some appropriate solutions and problem-specific solvers.

Without some well-defined tests and mathematical verification, I would not have discovered these issues. I would not have been able to explain why they were happening, either.  So, I had to debunk the magic to provide some concrete explanations and validations methods.

Let’s show us the light…

With the rise of machine learning techniques, there is a call that these applications of statistics and probabilities need to be explained. More generative machine learning is being used. Some humans should validate the generated models. Otherwise, it could only be gibberish and some biased solvers. They should be mathematically analyzed and proven too.

It has become even more important to be able to test every piece of code so that the outcomes can be trusted. The world is slowly moving towards intelligence augmentation. So, humans need to understand clearly how computers have obtained and computed their results. Human intelligence needs to be able to read, decipher, and make sense of these results. Otherwise, we may be making some decisions on spurious results.

In the opinion of the author, http://www.no-free-lunch.org/   brings some compelling arguments.

To conclude

Not every Neil Gailman readers may like fairy lights. Not every fairy lights lovers may like Neil Gailman stories.  We should resit and evolve from finding these spurious results.