AI learns the language of mathematics
Previously, neural networks could only add and multiply simple expressions. Computer scientists have now presented a self-learning algorithm that solves differential equations and calculates antiderivatives - outperforming all previous methods.
In some disciplines, one encounters complicated mathematical expressions: be it a differential equation in biology that models the spread of a viral disease, or an integral in physics that calculates the length of a particular curve. In order to master such tasks, you often have to dig deep into your bag of tricks, for example by laboriously reformulating a few terms or partially integrating an equation. However, there is often no way around looking it up in collections of formulas. In everyday life, researchers mostly use algebraic computer programs such as Mathematica, Maple or Matlab, which can handle mathematical expressions.
Their algorithms aren't perfect though. They don't always find the correct result, and they also need an extremely long time for some problems. The French computer scientists Guillaume Lample and Francois Charton, both at Facebook AI Research in Paris, have now developed an algorithm that masters this task better and faster than previous software. This was the first time the researchers were able to teach an AI the symbolic language of mathematics …
