A team of researchers from Graz University of Technology (Tu Graz), Pro²-Future, and St. Gallen University are trying to lift Tiny Machine Learning (TINYML) to its absolute limits.
“Of course, these small devices run models that are not large-scale language models, but rather models with very specific tasks to estimate distances, for example,” explains Michael Krisper, scientist at Tu Graz's Institute of Technical Information. “But you need to make these models a little smaller first. This requires a few tricks and it's exactly these tricks that we've been working on as part of our project.”
Researchers came up with a method to run multiple AI models on resource-constrained devices (TOP) to correct noise from interference (above). (📷: Michael Zedick)
Machine learning is not a new technology, but the explosion of interest in large-scale language models (LLMS) (technology that drives chatbots like Openai's ChatGPT and Google's Gemini) has spilled into other lighter implementations. For these researchers, it is TINYML. Embedded machine learning designed for resource-constrained devices, including microcontrollers.
In this project, “specialized AI models” run on-device on an Ultra Wideband (UWB) location tracker with just 4kb of memory to identify and correct the source of location data interference. Secret: Split tasks into application-specific models rather than a single model that does everything. For example, if there is interference from a metal wall, it will be handled in one model. Interference from another shelf. Interference from people is handled by the third model.
Determining the model to use was the work of the orchestration system that the team discovered, determining the type of interference present and loading the required models within 100 ms. The same approach can be used elsewhere with the car's key fob, for example, to block keyless relay attacks or extend battery life with smart home remote control systems, according to the team.
More information about the project is available On the Tu Graz website.
