Artificial Intelligence Network Teaches Itself Math System Invented During Holocaust

Ottawa – Artificial intelligence researchers working at the Royal Polytechnical College recently observed that a machine learning computer network called the Unsupervised Deep Learning System (UDLS) has optimized itself to solve mathematical calculations by using a shortcut system originally conceived over 75 years ago in Nazi concentration camps by the Russian engineer and mathematician Jakow Trachtenberg.

“At first, we assumed that UDLS was somehow seeing patterns that we were unaware of due to its enormous processing power.” said Mark Halput, the team’s lead research scientist and Chair of the Artificial Intelligence department at the college. “However, upon further analysis, we discovered that UDLS had combined its own observations of calculated results with an implementation of this human- devised mathematical system to greatly increase the speed whereby it produces results.”

The Trachtenberg System, as it is now known, was developed over seven years during Trachtenberg’s incarceration as a political prisoner in several concentration camps for his pacifist beliefs and outspoken criticism of Hitler’s regime. Breaking down complex calculations that are typically only able to be solved with pencil and paper, he outlined simple procedures that almost anyone can follow to mentally resolve math problems such as multiplication of two nine-digit numbers. In fact, his system was so successful and easy-to-learn that many European countries such as Switzerland still teach it as part of its basic mathematic curriculum.

While knowledge of this system is commonplace in other parts of the world, not many students are taught it in North America. So, it was surprising to the researchers when UDLS, created in Canada, started making use of this method to improve its computational speed and accuracy.

In terms of measurable progress, the network was able to calculate the digits of pi to the 50 trillionth digit in 198 days, roughly a third faster than previous records of 303 days and did so using less RAM and memory utilization than previously thought possible. Although the calculation of pi to this length is not by itself very useful, calculations such as these often serve as verifiable benchmarks of processing power and progress.

Even more impressive was that this calculation was not completed on a quantum computer. Halput continues, “Almost everyone assumed that we have basically approached the limits of what traditional binary computers can do. UDLS helped us realize that there is still much to be learned and progress to be made in this field – specifically around self-optimization and deep learning that leaves plenty of room for advancement.”

Sarah Wu, a neural network post-doctoral researcher at the University of Radcliffe-Leeds, who was not involved in the project, was impressed with the preliminary results. “As the pace of technology and artificial intelligence continues to accelerate, I think it is important to realize that just as human knowledge is based on previous discoveries, we should not be shocked to learn that computers can, and will, increasingly utilize past knowledge to solve increasingly complex problems. We are all standing on the shoulders of giants.”

A more detailed summary of UDLS and the researchers’ findings will be published in the upcoming August edition of AI Today.