DeepMind breaks 50-year-old math record – for a week
Researchers from Google subsidiary DeepMind have developed an AI system that has broken a 50-year-old mathematics record. However, the new record lasted just a week, as two researchers from the Johannes Kepler University in Linz have apparently already topped it.
Computer systems are becoming more and more intelligent. Artificial intelligence (AI) already supports us in everyday life. Anyone who owns a smartphone, for example, can no longer avoid an AI such as Google’s Assistant or Apple’s Siri. A further development of the system by Google subsidiary DeepMind is now attracting attention again.
The keyword here is: matrix multiplication. This type of multiplication plays an important role in the creation of systems for speech or image recognition, as well as for the display of graphics. For example, video games would not be possible without matrix multiplication. The calculation is usually done on the graphics card of a computer.
DeepMind: AI system reduces necessary computing capacity
In normal operation, the graphics card does nothing other than multiply matrices with each other. The more arithmetic operations are necessary for this, the more resources a computer needs. In 1969, the German mathematician Volker Strassen therefore developed a method to reduce the number of arithmetic operations.
While a 4 × 4 matrix previously required 64 multiplications, it was only 49 afterwards. DeepMind’s AlphaTensor network also follows this approach. An advanced system has now succeeded in reducing the multiplications to 47. This was made possible by a fictional math game that the researchers used to train the AI.
Researchers from Linz beat the record after a week
AlphaTensor subsequently also managed to reduce the number of multiplications to 96 for 5×5 matrices. But as soon as DeepMind set the record, came two researchers from the Johannes Kepler University in Linz and grabbed the crown. They achieved the same results with only 95 multiplications.
However, both developments should pave the way for future graphics cards, as they are dependent on the calculations. With every reduction in complexity, computer systems are created that can do significantly more than their predecessors and enable even more realistic simulations.
Also interesting:
