University of Twente Advances Neuromorphic Computing with Hardware-Based Learning Method Free of Backpropagation

University of Twente Advances Neuromorphic Computing with Hardware-Based Learning Method Free of Backpropagation

(IN BRIEF) The University of Twente’s BRAINS Center for Brain-Inspired Computing has developed a groundbreaking hardware-based learning method that enables electronic materials to adapt without using backpropagation or digital optimisation software. Detailed in Nature Communications, the technique — homodyne gradient extraction — allows physical neural networks to find optimal states directly in hardware, offering a significant step toward low-energy neuromorphic devices. This innovation could pave the way for adaptive sensors and highly efficient brain-inspired computing systems.

(PRESS RELEASE) ENSCHEDE, 24-Nov-2025 — /EuropaWire/ — University of Twente researchers from the BRAINS Center for Brain-Inspired Computing have unveiled a pioneering approach that allows electronic materials to adapt in a manner similar to machine learning systems. Their study, published in Nature Communications, introduces a physical learning method that operates entirely without software-based optimisation techniques such as backpropagation. Backpropagation — made widely known in the 1980s through the work of Nobel Prize laureate Geoffrey Hinton and colleagues — remains the fundamental mechanism driving today’s artificial intelligence advancements.

Modern AI depends heavily on backpropagation executed on powerful digital hardware, achieving impressive results but at a substantial energy cost. The human brain, however, performs comparable tasks using only the energy consumption of a standard light bulb. Although neuromorphic hardware promises far greater efficiency by mirroring biological neural systems, it has traditionally struggled because backpropagation cannot be easily translated into physical systems.

The Twente team addresses this bottleneck with an innovative method known as homodyne gradient extraction (HGE). This technique allows physical neural networks to locate their optimal operating point directly within the device itself, without relying on digital computation or software algorithms. While external perturbations are applied to guide the system, the actual optimisation process occurs inside the material, enabling a new form of stand-alone, hardware-based learning.

“This opens the door to stand-alone optimisation of physical neural networks, offering a path towards energy-efficient, adaptive hardware,” says Prof. Wilfred van der Wiel, co-director of the BRAINS Center. The researchers foresee wide-ranging potential applications, including smart sensors capable of on-site adaptation and brain-inspired computing platforms designed for sustainable, low-energy information processing.

The team’s open-access paper, Gradient descent in materia through homodyne gradient extraction, is now available via Nature Communications.

Media contact:

K.W. Wesselink – Schram MSc (Kees)
Science Communication Officer (available Mon-Fri)
+31 53 489 9311
k.w.wesselink@utwente.nl
Building: Spiegel Tuin

SOURCE: University of Twente