Artificial intelligence: Inspired by the human brain

Microsoft wants to bring Three Mile Island back online. It was a headline that really made people sit up and take notice. The decommissioned US nuclear power plant is best known for its partial meltdown in 1979. The project shows the lengths the tech giant is prepared to go to to cover the enormous energy requirements of its applications. And Microsoft is far from alone. AI applications are driving up the power requirements of data centers worldwide. The amount of energy consumed to train GPT-4, for example, would have been enough to power a small town. Moore’s Law, which states that the number of transistors in an integrated circuit doubles approximately every two years, leading to an increase in computing power, may be slowing down, but computers will continue to become more powerful. The ever-growing density of computer chips is part of the energy problem. Because the processor and memory are separate, data is constantly in motion.

The core disciplines of AI – pattern recognition, analysis and inference – require a particularly enormous volume of data transfer. This back and forth comes at the cost of latency and requires a great deal of energy. In order to resolve this classic bottleneck problem, current research approaches in microelectronics are focusing on “in-memory computing.” The idea is that computing operations should run in close proximity to the memory. The block-based architecture of storage and processor units is eliminated and energy-intensive data traffic is minimized. Nature shows us how information processing works in energy-saving mode. Our brain manages with a power of around 20 watts, which is less than some light bulbs. Inspired by the human brain, scientists are developing hardware structures that are based on biological neurons. The aim of neuromorphic computing is to create an AI infrastructure that is energy-efficient, fast and autonomous enough to move intelligence to the edge, i.e. as close as possible to the data source or to the end devices. Battery-powered sensors or wearables could benefit from low energy consumption. Instead of draining the battery in hours or days, edge AI applications could run for months or years, depending on the operating mode. Since the detour via the cloud would be eliminated, the data would remain local. This would reduce latency and improve robustness. The system would remain operational even without a network connection.

It would be useful if robot lawnmowers no longer stopped working as soon as the internet went down. However, there are also more important benefits than greater convenience for consumers. “The obvious application is in medical technology,” explains Prof. Regina Dittmann, Head of the Institute of Electronic Materials at Forschungszentrum Jülich. “We are thinking, for example, of warning systems that monitor vital functions. Devices that can run for weeks without a battery change would be a major step forward.” Dr. Markus Eppel, Group Manager Advanced Analog Circuits at the Fraunhofer Institute for Integrated Circuits (IIS), sees another urgent area of application in the “monitoring of critical infrastructure.” This is an issue that was brought to public attention by the partial collapse of Dresden’s Carola Bridge in September 2024. Dr. Eppel cites an example from his own development work. In 2022, Fraunhofer presented a smart screw called “Q-BO.” It uses microelectronics to record both the contact pressure and vibrations. This allows it to monitor the stability of the connections. Bearing damage in generators or motors can also be detected. The anomaly detection in such systems still mostly runs on classic microcontrollers. “Neuromorphic hardware can achieve this much more efficiently, i.e. with longer maintenance intervals or faster response times,” says Eppel. The automotive industry is among those banking on this potential. This is where the “holy grail” of fully autonomous driving beckons. Manufacturers hope that neuromorphic hardware will advance machine vision without putting a strain on the traction battery.

Neuromorphic microelectronics is still a long way from widespread use. “We are seeing the first commercial components,” says Eppel, but “validation under real operating conditions” still awaits in many cases. In Jülich, Regina Dittmann’s team is integrating novel components into CMOS chips that demonstrate “the basic functions of neuromorphic computing.” System tests have so far been carried out in simulations. The search for competitive network architectures is leading to different technical approaches. With its “Adelia” accelerator, Fraunhofer IIS is pursuing a mixed-signal approach that combines analog and digital circuits to execute neural networks. “In tests, our model for voice activity recognition consumed up to 90 percent less energy than purely digital approaches – with a loss of accuracy of no more than 3 percent,” explains Eppel.

Another approach can be found in the “Senna” accelerator. It executes so-called spiking neural networks with deterministic timing and extremely short response times of less than 20 microseconds. The aim is to ensure that AI systems make a decision in every case – regardless of memory bottlenecks that would otherwise lead to delays or errors. To do this, “Senna” does not process continuous numerical values, but short pulses known as spikes. Spiking neural networks are therefore based on the efficiency miracle that is the brain; a neuron only fires when something important happens. Voice control, for example, would no longer be forced to constantly analyze ambient noise. It would only react when the keyword was spoken. In order for this reaction pattern to function permanently, the learned connections must be conserved with as little energy as possible. Components from Jülich inspired by brain cells are designed to boost the memory capacity of neural networks. Memristors – a portmanteau of memory and resistor – are essentially resistors with memory. They maintain their conductance even when the power is switched off. Made from hafnium or tantalum oxides and 2D materials, they are designed to mimic the way synapses work.

To break away from conventional computing, Forschungszentrum Jülich is cooperating with RWTH Aachen University on the NEUROTEC project. “Our aim is to map the entire value chain – from material development to chip and system architecture to software,” is how project manager Regina Dittmann sums it up. Industry partners are also on board. At the same time, the NeuroSys Future Cluster is working on creating an “ecosystem of neuromorphic technologies.” In line with the motto “from coal to AI”, the region is to become a leading location for neuromorphic electronics. NEUROTEC is already in its second project phase. By the end of 2026, funding amounting to 36 million euros will have been provided. Project manager Dittmann points to several start-ups as an indicator of success. However, the region does not aim to be entirely self-sufficient; closer cooperation with the “Silicon Saxony” high-tech network is planned in and around Dresden, the center of the German semiconductor industry. While it will probably be the end of the decade before chips with memristive components from Jülich enter series production, Dittmann believes that neuromorphic chips based on conventional technologies could be ready for market in the nearer future. Markus Eppel is even more specific. He sees the “widespread use of neuromorphic systems in leading technology clusters within the next two to three years.”