Pioneers of a New Computing Era
Physicists John Paul Strachan and Emre Neftci conduct research into computer systems that are based on the functioning of the human brain. In July 2021, each of them was appointed head of a newly established research division at Forschungszentrum Jülich. The two top international scientists will contribute to strengthening comprehensive research at Jülich and to putting neuromorphic systems into practice.
Isn’t it fascinating? In a gigantic network, 86 billion nerve cells are connected to each other via synapses and perform astonishing mental activities. John Paul Strachan and Emre Neftci both get excited when talking about the human brain and its incredible complexity, extraordinary capabilities, and unprecedented efficiency. It hardly consumes more energy than a simple light bulb, while conventional supercomputers sometimes need over a thousand times more energy to perform certain tasks.
The perfection of the human brain inspired Strachan and Neftci in their youth and made them who they are today – developers of computer systems that perform calculations based on the model of the human brain. Although such neuromorphic model circuits already exist, none can be used for meaningful applications. By coming to Jülich and establishing two new divisions at the Peter Grünberg Institute, Strachan and Neftci want to help change this.
The two physicists gained extensive experience in neuromorphic computing working in the high-tech areas of California. Strachan, who was born in Costa Rica and grew up mostly in the USA, most recently conducted research at Hewlett Packard Labs in Silicon Valley. The company is a pioneer in the field of neuromorphic computing. Strachan headed a team there conducting research into neuromorphic hardware. He holds more than 50 patents, and completed his degrees at two leading American universities, MIT and Stanford.
Swiss-American Emre Neftci, meanwhile, focuses on the software needed for neuromorphic chips. He most recently worked as an assistant professor at the University of California in Irvine. Before moving to California, Neftci completed his doctorate on the topic of neuroinformatics at ETH Zurich (Switzerland). One thing he likes about Europe is the visionary and yet capable approach to neuromorphic computing that is being pursued there.. “Jülich has experts from different fields who we need for this kind of interdisciplinary research,” he says. There are hardware and software engineers, semiconductor experts, and theoreticians, but also neuroscientists. Both Neftci and Strachan are enthusiastic about Forschungszentrum Jülich’s “package”, which is probably unique in the world. For their new jobs at Jülich, they will gladly trade in their surfboards for a mountain bike or try out some new outdoor activities.
Experimental space for brain research and neuromorphic chips
Strachan envisions a “playground for new computing systems” where Jülich researchers can experiment to their hearts’ content. He and his team want to bring completely new chip concepts to the table. “Neuroscientists can use these concepts to test their models of neural network functions and gain a better understanding of the human brain,” says Strachan. Their ideas and experiences will then in turn help him and his team to further improve the hardware systems. In this context, he speaks of a “feedback loop” where the different disciplines can benefit from each other’s expertise.
When it comes to simulating complex learning processes in the human brain, conventional computers – even supercomputers – quickly reach their limits. In nature, these processes often take days or weeks, sometimes even years. Simulations on supercomputers run even more slowly and so far can only simulate small parts of the network in the human brain. Neuromorphic systems present valuable opportunities to expand and significantly accelerate the necessary calculations.
Human versus supercomputer
About ten years ago, IBM’s Watson computer beat all its human opponents on the US quiz show Jeopardy! However, with regard to energy saving, its flesh-and-blood opponents were the clear winners since the supercomputer required 200 kW or 200,000 watts to achieve its victory. The human brain, on the other hand, delivers a similar performance at just 20 watts.
(source: Forschungszentrum Jülich/NEUROTEC project).
Economical circuits with artificial synapses
In Strachan’s opinion, the biggest goal in neuromorphic computing is to “solve problems the way biological systems do”. He explains that biological systems discover efficient computation by modifying and integrating all levels of abstraction at the same time. By contrast, current computers manage different tasks in a highly segregated way, both in terms of time and space – separate working memory and processor, for example.
At Jülich, scientists have been working for some time on memory and devices that have similar properties to the synapses in the human brain. These researchers were among the first to start developing artificial synapses from memristors about ten years ago. Jülich experts are viewed at the forefront of research when it comes to understanding these switching elements on the microscopic level.
Memristor
The memristor is a so-called portmanteau and made up of the words memory and resistor. Memristive devices have similar properties to synapses in the brain. Jülich researchers want to use these elements to develop economical chips based on models in nature.
Memristors are storage and computing units combined and change their properties based on incoming signals. Their state thus depends on the history of the signals that have been received. Similarly, in our brains, neural pathways also grow or shrink due to increased or decreased signal transduction.
Concept for efficient AI
Both Strachan and Neftci emphasize the benefits that artificial intelligence (AI) technology could gain from neuromorphic approaches. Current AI concepts merely imitate the self-learning mechanisms of neural networks by means of large-scale software and conventional hardware. Neuromorphic systems, however, are constructed in a similar way to the biological networks they emulate. This allows the signal exchange to be recreated much more directly and efficiently than is possible using conventional processors.
According to software expert Neftci, in order to improve the self-learning hardware concept, it is particularly important to close the gap between machine learning and physical systems. “How can the functioning of the human brain be mapped onto chip technology? Machine learning combined with neuroscience holds many answers”, he says. With his team, he wants to develop the right software and algorithms to answer this question.
“For neuromorphic computers, different principles and restrictions apply than for conventional ones,” he says, describing the challenge from a programmer’s perspective. Machine learning – the data-driven learning of computer programs – must be developed further by incorporating findings from brain research. In this way, the different disciplines can benefit from each other and move closer to attaining their common goal.
Future technologies for supercomputing
t another Jülich institute, researchers are already working towards using neuromorphic computers to accelerate calculations on supercomputers in the future. Together with partners from research and industry, the Jülich Supercomputing Centre (JSC) has developed a modular supercomputing architecture that makes it possible to combine different types of computing modules into a unified whole according to the building-block principle.
JSC operates and develops supercomputers, which are among the most powerful in the world, for different scientific purposes. The JUWELS and JURECA supercomputers are already constructed modularly. They each comprise a cluster module and a booster module which are optimized for different types of tasks.
Future technologies – such as quantum computers or indeed neuromorphic modules combined with conventional supercomputing architectures – can also be integrated according to the same principle. It’s a fascinating idea that is exactly to the taste of Jülich’s two new experts for neuromorphic computing.
Janosch Deeg