Featured
As complicated as computational theory may seem, some of its most fundamental elements mimic the same phenomena you’d find in nature.
Look at neural networks, which were designed by computer scientists in the late 1950s. They function similarly to the neural pathways of the human brain and have been foundational to many artificial intelligence-based technologies out in the world today.
“Nature is pretty good at solving problems,” says Mauricio Tedeschi, a fourth-year computer science and physics student at Northeastern University. “You can see this in a lot of biological structures — the human brain is one of the most powerful machines in the universe. Emulating the human brain’s function seems like an obvious way to make much better technology.”
During his co-op at Forschungszentrum Jülich, Tedeschi applied those principles as he helped the German interdisciplinary research center develop and test algorithms for dedicated “brain-inspired hardware” for high-performance computing.
Tedeschi worked in The Peter Grünberg Institute, which is one of 14 institutes within Forschungszentrum Jülich. The institute focuses its research on a few specific subjects: quantum materials, quantum computing, software systems and the area Tedeschi worked on — neuromorphic computing.
“Neuromorphic computing really just means brain-like computational devices that are inspired by the structure and function of the human brain,” Tedeschi says.
Neuromorphic computing is a major area of study in particular as researchers strive to make computer chips more efficient and cost effective. By mimicking the neurons of the human brain, how can researchers better optimize these chips while bringing down their computation costs?
Tapping into his expertise in physics, mathematics and computer science, Tedeschi assisted in the development of code designed to do just that.
“What this institute is trying to do is utilize specialized hardware inspired by the laws of physics to perform certain computational tasks that can improve efficiency of computation and make workflows a lot faster and solve problems a lot better,” Tedeschi says.
Tedeschi was working with chips based on a brain-inspired computing model known as Hopfield neural network, which was created by John J. Hopfield in 1982.
As opposed to deep learning neural networks that are used in services like ChatGPT, Hopfield networks are useful for recreating and saving patterns, which is especially useful in solving specific optimization problems in computation.
In practice these technologies can help in solving what is known in the computational world as the “Traveling Salesman Problem,” a term used to describe the issue of finding the most optimal routes in a series of different options.
“There are multiple different problems in supply chain and business operations that can readily be solved with these kinds of chips,” says Tedeschi. “You have all this scheduling data or this geographical data. What is the most optimal schedule to lay out all these different tasks?”
While at the co-op, Tedeschi worked hand in hand with scientists at the institute and contributed to research that will be published in the 2024 IEEE International Conference on Rebooting Computing (ICRC) proceeding into how these chips can handle these highly complicated tasks more efficiently.
“We want to produce hardware that is much cheaper for algorithms to be operated on,” he says. “By using these hardware architectures, you’re able to reduce the amount of data bottlenecks at lower energy levels that the human brain uses instead of these massive data centers that churn out so much energy.”