Artificial intelligence (AI) usually requires a lot of computing power and energy, which can be a problem for small devices like sensors used in the Internet of Things (IoT).
These devices often have limited memory and small batteries, making it difficult to run complex AI models directly on them.
But a new research project called E-MINDS is changing that.
Led by researchers from Pro2Future, Graz University of Technology (TU Graz), and the University of St. Gallen, the E-MINDS project has developed smart ways to run AI efficiently on tiny devices—without needing outside help from bigger computers.
For example, the team was able to run AI models on a device with just 4 kilobytes of memory. That’s less memory than a simple text file!
This tiny device could still analyze signals to figure out where interference was coming from, using location data.
The secret lies in breaking down the AI models into smaller, more focused pieces. Instead of using one big model, the system uses several small ones, each designed to handle specific problems.
For instance, one model might handle interference caused by metal walls, another by people, and another by shelves.
A controller chip on the device figures out what kind of interference is present and loads the correct model from a server in just 100 milliseconds—fast enough for use in places like warehouses.
Another method the researchers used is called subspace configurable networks (SCNs), which are flexible AI models that adjust to the type of data they’re given. These models were tested on tasks like identifying objects or fruits in images.
Even though they were small and used less power, they were still fast—up to 7.8 times faster than traditional setups that rely on external computers.
The team also used techniques called quantization and pruning to shrink the AI models even more. Quantization simplifies the math used in the model, switching from complex numbers to simpler ones, which saves power and memory.
Pruning trims away the parts of the model that aren’t necessary for the task, keeping only what’s essential while maintaining accuracy.
Beyond just shrinking the models, the researchers also made sure they could be quickly and efficiently loaded onto the devices.
Although the main focus of the project was using AI to improve indoor positioning systems—like helping drones or robots find their way in busy warehouses—the team sees many other uses. For example, smarter car key systems that can tell if the key is really nearby, or home gadgets that last longer on a single battery charge.
According to project leader Michael Krisper, the E-MINDS team brought together experts in hardware, AI model optimization, and indoor positioning to lay the groundwork for smarter, more efficient devices in the future.
