The human brain surpasses any computer in both power and energy efficiency. Researchers are mimicking its functioning to create enhanced computer chips and manage the ever-increasing data produced daily.
By Tom Cassauwers
To safeguard smart home devices from hacking, scientists are designing extremely rapid, energy-efficient brain-inspired chips capable of identifying threats in real time directly on our devices.
From smart refrigerators and televisions to internet-enabled toothbrushes, a growing number of household gadgets are now integrated into the Internet of Things. This facilitates the analysis of usage data or the installation of remote updates, but simultaneously poses a security risk.
These intelligent devices are often targets for hackers aiming to form so-called botnets – networks of compromised devices that can facilitate large-scale cyber-attacks.
Computing on the edge
In response to this issue, one approach is to gather all data transmitted through a device and relay it to a data centre, where AI algorithms can detect suspicious activities across millions of connected devices. However, this process is time-consuming and necessitates transferring vast quantities of data.
Consequently, scientists strive to execute these calculations locally – on the fridge or toothbrush itself.
Yet, this concept of edge computing, where processing occurs on the network’s periphery, presents its own set of challenges. Numerous intricate calculations must be executed swiftly on diminutive chips that use minimal electricity.
“If you’re generating these vast amounts of data, processing it instantaneously is quite demanding,” explained Dr. Matěj Hejda, a research scientist specializing in advanced computing and photonics. Hejda is part of an EU-funded project known as NEUROPULS, which is addressing this challenge directly.
Hejda and fellow researchers on the NEUROPULS team are working on a small chip, or processor, capable of performing rapid AI calculations while using hardly any energy.
“In the event of a cyber-attack, delays are unacceptable. We depend on AI to make swift decisions based on extensive data. That’s the function of our chip,” he stated.
Brain power
Their innovation draws inspiration from the human brain, which executes intricate tasks while consuming significantly less energy than present-day conventional computers. By grounding their efforts in the fundamental characteristics of neural processing, the team aspires to provide intelligent, low-power computing for various real-world applications.
“The circuits imitate the functions of the brain,” noted Dr. Fabio Pavanello, a lead researcher from the French National Centre for Scientific Research at the Centre for Radiofrequencies, Optic and Micro-nanoelectronics in the Alps. Pavanello coordinates the NEUROPULS research.
This new amalgamation of neuroscience and advanced technology is termed neuromorphic computing, and its importance is rapidly rising.
“There are numerous methods to achieve this. We opted for photonics, which entails using light beams as opposed to electrical signals for computations,” stated Pavanello.
Merging memory and processing
Some research is being conducted at Hewlett Packard Enterprise labs in Belgium, where Hejda is stationed. The researchers are striving to resolve a significant bottleneck in contemporary AI computing: memory.
“We have a method to circumvent that barrier,” Pavanello said. In traditional computers, memory is distinct from the central processing unit where calculations are performed. The processor handles calculations, while the data utilized in those calculations resides in the memory unit.
This data constantly needs to be transferred between the memory and the processor, typically via some electrical circuit. This creates a bottleneck for AI as the connection between the processor and memory struggles to manage such enormous data flows.
This bottleneck results in slower calculations and increased energy consumption. However, the researchers may have uncovered a solution.
“Our goal is to integrate memory and calculations into a single location,” remarked Hejda. “Interestingly, this is how it functions in our brain. In nature, memories and cognitive processes appear to coexist.”
Light waves
Another advancement proposed by the NEUROPULS chip is ultra-low-power photonic computing. Instead of relying on electrical signals for calculations, it utilizes specialized chips where light travels through tiny pathways known as waveguides.
Employing light offers several benefits, including negligible signal loss, extremely low latencies or delays when sending and receiving data, and high data rates.
“It’s also simpler to conduct numerous parallel calculations by employing different wavelengths of light,” Pavanello explained.
“With these systems, you can deploy more sensors and collect additional data. This allows for better-informed decisions with lower energy costs.”
Another advantage of photonic technology is the possibility of constructing more secure safeguards for these chips to enhance their operation and data security. “This is a crucial requirement for their safe integration into systems and networks,” Pavanello added.
Boost for self-driving cars
The NEUROPULS research team intends