Neuromorphic hardware: Paving the way for Edge AI

March 9, 2021 | Robotic lawn mowers have only limited “intelligence”: up to now they have needed a boundary cable to prevent them from accidentally buzzing the neighbor's grass as well. Edge AI will make it possible to equip lawn mowers, sensors or other devices with artificial intelligence – without having to send the data gathered for the intelligent task to the cloud. Instead, the “intelligence” will be located in the device itself. But in many cases, this can work only with the help of special, extremely energy-efficient chips; powerful microcontroller chips are rarely suitable because they consume too much energy and are too expensive.

Neuromorphic hardware: Inspired by the human brain

Enter neuromorphic hardware: energy-efficient chips designed especially for Edge AI will serve as the foundation for this technology. The structure of this hardware (morphology) is inspired by biological neural systems, whose complexity ranges from the simple nervous systems of mollusks to those of the human brain and can ultimately teach us a great deal about efficiency. Researchers examine and analyze these neural systems and then use mathematical models to express the way they communicate with each other. Although the communication pathways in the brain and other neural systems cannot be directly translated into electronic circuits, these mathematical models provide the basis for the required energy-efficient circuits.

Neuromorphic hardware: Different ways to the same destination

There are different design choices to be made when developing neuromorphic hardware. One is to design the hardware as an analog system – bio-inspired  – or to imitate the structures and processes digitally. Another is to use deep learning or spiking neural networks. Deep learning is already widely used in artificial intelligence, especially in image recognition. In contrast to the typical neural networks used in deep learning, spiking neural networks (SNNs) imitate the way biological neurons communicate by firing only short pulses (“spikes”) to convey information to other nerve cells. This results in faster and more energy -efficient data transmission and processing. While the average PC relies on a central processor, neuromorphic chips profit from simple computing units working in parallel.

How the different approaches can be combined – such as digital deep learning or analog spiking neural networks – depends on the application in question. Fraunhofer IIS researchers offer a breadth and depth of expertise in the majority of these neuromorphic hardware paradigms, which paves the way for incorporating AI directly into devices for a wide variety of applications, such as manufacturing and mobility, without requiring a connection to the cloud.



Your questions and suggestions are always welcome!

Please do not hesitate to email us.

Stay connected

The newsletter

Sign up for the the Fraunhofer IIS Magazine newsletter and get the hotlist topics in one email.

Home page

Back to home page of Fraunhofer IIS Magazine