The AI revolution in Science
Today 40 monkeys had the privilege to listen to Dr. Abbey Waldron talk about her experiences from Machine Learning in the Science world. It was an information packed presentation where we learned about things like Quarks, Leptons, Force carriers and the famous Higgs Boson and that scientists used beer to spot them. I guess we must try that at our next social event the 12:th of December. We also learn that matter can appear in five states and that the Bose-Einstein Condensate is one state used to catch small particles.
It turns out Machine Learning has been used almost as long as there have been computers to crunch the numbers in science but it got more widely used after 1989. But then in simpler forms. Because science being science and all, it is for some reason important to prove that what you find is real and you also want to make sure you do not miss anything. So, scientists have refused to use neural nets because their many flaws. If you train a GAN to improve images it will learn to create a clear image out of a blurry one. But there is a extremely low chance that the image you get is actually the correct one.
Of course, an image classifier can look at tones of data to find interesting information that can be studied further by scientists, where the usual student assistants fall asleep after a couple of hours. But the neural nets might also miss a lot of cases because it was not properly trained, or the scale or rotation of the object did not match the trained data. Even though Geoffrey Hinton and his colleagues published a new paper that might solve some of these issues with a Capsule Network. (Learn more about the Capsule Network by Siraj or listen to Hinton talk about it himself).
Did you think that python and R are the only languages out there for writing Machine Learning. Think again. Python is popular because google as adopted it and there are many good libraries that makes it easy to build models. But you can write machine learning in any language. Abbey says that she mostly uses C++ because it is a lot faster because it compiles to byte code instead of running on some script interpreter. If you are a beginner in programing or machine learning though you might want to begin with python since there are so many resources and libraries to support you and most of the latest research comes out in python code as well.
When having after works, we have had some discussions about machines taking over or AI used for bad purposes. I did not get more optimistic over this when I heard that the reason for us having nuclear power is that it was a nice way to promote the production of enriched plutonium. But the real intention and the money behind the campaign was focused on nuclear weapons. It makes you wonder what forces lurks around the AI revolution. I guess we will know that in about 50 years...
- Particle Physics, Machine Learning, Science