º£½ÇÉçÇø

Researchers from º£½ÇÉçÇø’s Department of Physics have secured funding for an ambitious European project exploring a radically different approach to artificial intelligence.

The four-year project, KAN-DO-IT, has been awarded €4.2 million through the European Innovation Council (EIC) Pathfinder programme.  

Led by Professor Marco Fanciulli at the University of Torino, the consortium brings together scientists from Imperiaand across Europe. º£½ÇÉçÇø’s Department of Physics will receive approximately €660,000 as part of the award. 

Dr Jack Gartside, who leads º£½ÇÉçÇø’s Neuromorphic Metamaterials group and will run º£½ÇÉçÇø's contributions to the project, said “We’re thrilled to have the support of this EIC Pathfinder Grant, it’s one of the very few ways to support an international team working together across Europe, and they’re hard to get – just 43 were funded out of over 2000 submitted. 

The big excitement for us is the chance to work together long term with our friends across Italy, Sheffield, Germany and the Netherlands. Everyone brings something unique to the table, and what we can do as a team would just not be possible to achieve alone.”  

KAN-DO-IT will investigate how nonlinear physical systems can be used to perform machine-learning tasks directly, potentially offering a more energy-efficient alternative to conventional AI.  

Rethinking how AI learns 

Artificial intelligence is increasingly woven into everyday life, from image recognition and language translation to healthcare and autonomous systems. However, these advances come with a growing demand for energy and computing resources.  

Most modern AI systems are designed to run on Graphics Processing Units (GPUs), which excel at performing vast numbers of mathematical operations simultaneously. As a result, many AI architectures have evolved to suit conventional computing hardware.  

The challenge is that these architectures were developed for digital processors, not the nonlinear physical systems that could one day offer faster and more energy-efficient alternatives.  

For the past six years, the Neuromorphic Metamaterials group has been investigating whether entirely new forms of AI could be designed around the strengths and natural behaviour of physical systems themselves, and how we might train them. 

Recently, they demonstrated a new approach based on so-called Kolmogorov-Arnold Networks (KANs), which allow physical systems to perform much of the computation themselves. 

The team showed that this approach could dramatically reduce the size of neural networks while maintaining performance. Their work demonstrated reductions of up to 600-fold compared with conventional software approaches, alongside projected energy savings of up to a thousand times for suitable machine learning tasks.  

The new project aims to build on these results by developing larger networks, smaller devices, faster operation and more advanced learning algorithms.  

Learning from physics 

At º£½ÇÉçÇø, researchers will develop physics-aware training methods that avoid the need for conventional backpropagation techniques while also implementing and testing physical learning networks experimentally. 

The team hopes this work will help establish a new class of learning systems capable of operating in situations where power, connectivity and computing resources are limited.  

The technology and approach at the heart of this project represents an exciting step towards turning exotic nonlinear physical dynamics into an industrially-scaled technology that can benefit a lot of people. Dr Gartside, º£½ÇÉçÇø

Dr Gartside added “The technology and approach at the heart of this project represents an exciting step towards turning exotic nonlinear physical dynamics into an industrially-scaled technology that can benefit a lot of people.  

We really believe in the core science, and we think that we’re just scratching the surface of how much nonlinear physics can offer neural networks beyond what GPUs and conventional AI are capable of.” 

Potential applications include adaptive health-monitoring wearables, neuroprosthetic devices, autonomous robotic systems and data-efficient computer vision technologies. 

Bringing AI closer to the user 

Unlike many modern systems, which rely on powerful remote servers and data centres, these devices could process information locally, helping to improve privacy, reduce energy consumption and give users greater control over their data. 

If successful, KAN-DO-IT could establish a new class of physics-based learning systems designed to operate far from data centres on extremely small energy budgets. These systems could learn continuously from scarce local data and adapt throughout their lifetime.  

Jack said “We want to free the benefits of AI from being owned and controlled by a tiny number of big tech companies. We want to keep data, privacy and autonomy in the hands of the end-user, rather than being shipped off to massive remote data centres. 

We’re really excited to get started. We’d like to thank the EIC and all our collaborators across Europe for their support and partnership in making it happen.”