3D glasses for the computer
Recognizing cats in images – self-learning algorithms have been mastering this for a long time. But what about more complicated objects, such as proteins? Using methods derived from physics, computer scientists are endowing computers with spatial perception, so that they can now detect complex patterns even in curved and multidimensional structures.
In recent years, computer scientists have made significant advances in the field of artificial intelligence (AI). The machines can now drive cars, they beat human world champions at board games like chess or Go and even write prose texts. Most such achievements are based on powerful neural networks, the structure of which is modeled on the visual cortex of mammals. This category includes, among others, so-called convolutional neural networks (CNNs), which are amazingly adept at recognizing patterns in two-dimensional data - especially when it comes to deciphering handwritten words or processing objects in digital images.
It behaves completely differently if you let loose such adaptive algorithms on data sets that do not originate from a planar geometry. For example, if you want to study irregular shapes like those used in three-dimensional computer animations, or if you want to process the amount of data generated by self-driving cars, which the vehicles use to map their surroundings, convolutional neural networks reach their limits…
