Edge artificial intelligence (AI) is enabling real-time data processing and decision-making directly on devices, reducing latency and bandwidth usage while enhancing privacy by minimizing data transmission to the cloud. However, its widespread adoption has faced significant technical obstacles. The DIME project, led by a research team at IMDEA Networks and coordinated by Joerg Widmer, Research Director, and Jaya Champati, Research Assistant Professor, concluded this year, achieving a critical breakthrough in overcoming these challenges. Its outcomes pave the way for faster, safer, and more sustainable edge artificial intelligence applications, with significant impact in key sectors such as healthcare, transportation, and smart cities.

One of the major challenges of deploying deep learning (DL) models on devices such as microcontrollers or smartphones has been understanding their impact on performance in terms of energy consumption, latency, and accuracy. The DIME project conducted an exhaustive study measuring these variables on five IoT devices, ranging from basic microcontrollers to advanced single-board computers like the Jetson Orin Nano. This analysis identified strategies to optimize performance without compromising decision quality.

Additionally, DIME tackled a crucial issue in critical applications such as medical devices and IoT infrastructure: the lack of mechanisms to verify AI model accuracy in real time. The team developed an innovative online learning algorithm that intermittently compares device-based model decisions with those of more robust models on edge servers. This Hierarchical Inference approach corrects errors in real time and offloads data to the server only when absolutely necessary, improving efficiency and reducing data transmission costs.

“The results of the DIME project are laying the foundation for the widespread adoption of edge AI,” says Jaya Champati. This is particularly relevant in remote or underserved areas with limited internet connectivity.

The research has garnered international interest, inspiring collaborations with universities in Canada, India, and Europe and achieving visibility at high-profile conferences such as the ACM Symposium on Edge Computing and IEEE INFOCOM. The data and tools developed during DIME have been made available to the scientific community via GitHub and will be proposed as a tinyML benchmark standard through MLCommons, an organization dedicated to accelerating innovation in machine learning.

The impact of DIME goes beyond technical achievements. By democratizing access to low-cost AI solutions, the project has the potential to empower small businesses, improve people’s quality of life, and foster greater digital equity. “DIME not only offers technological solutions but also creates opportunities for new applications and jobs in emerging AI sectors,” highlights the lead researcher.

With ground-breaking advancements in deep learning inference and a user-focused vision, DIME stands as an example of how technological research can transform industries and societies, driving a more efficient, inclusive, and sustainable future.