Embedded Artificial Intelligence

Thanks to the generalization and adaptation of Artificial Intelligence (AI), this field has grown rapidly in recent years. AI is currently limited by the need for massive data centers and centralized architectures, as well as the need to move this data to algorithms. To overcome this key limitation, AI will evolve from today’s highly structured, controlled, and centralized architecture to a more flexible, adaptive, and distributed network of devices. Mitigating from cloud-based computation to edge computation (which, in the case of AI, is called embedded AI or eAI) has many advantages, including but not limited to privacy, low power consumption and increased durability, low latency and fast response, and independence from the network connection.
Grundfos and VELUX, two leading companies in their respective industries, are taking their first steps in this journey. They produce millions of products, and it is anticipated that hundreds of thousands of them will soon be affected by eAI. Grundfos produces a variety of pumps for different applications, while VELUX focuses mainly on roof windows. For example, one potential scenario for eAI implementation is to add intelligence to the control systems of radiators (which use Grundfos pumps) and roof windows. This intelligent control system can make these different products work seamlessly together to optimize comfort, ventilation, and energy efficiency for home residents.

Despite the advantages eAI brings, it also introduces many challenges. The resource-hungry nature of AI models demands significant computing power and memory, making them unsuitable for small end devices. However, researchers found that it is possible to optimize the models in different aspects, making them reasonably small. Furthermore, the implementation of these models on microcontrollers requires a lot of consideration. This problem is mainly rooted in device heterogeneity and the diversity of the utilities and constraints each microcontroller offers.

Due to the increasing significance of eAI, numerous tools and toolchains have been developed by researchers and engineers. However, these tools are limited in their ability to support various model types and hardware architectures, leaving some essential features out of the equation. Thus, Grundfos and VELUX couldn’t find a suitable solution and are eager to develop a more general easy-to-use toolchain, answering their requirements.
We are keen to develop such a toolchain that not only supports the necessary models and hardware but also goes further and enables features such as incremental learning and resource estimation. The next step is to develop an eAI engine to encapsulate the AI task and allow it to function seamlessly with the rest of the microcontroller’s tasks.
We firmly believe that the successful completion of this project will allow these companies to easily integrate AI models into their products and open the door to a world of new possibilities. Our hope is that the end result will leave many of us happy as customers.