About the DN ANT project

Although the Cloud AI infrastructure remains the primary choice for training and executing Machine Learning/Deep Learning (ML/DL) models, its reliance on significant computational resources, dedicated hardware, large and diverse training data, network bandwidth, and associated issues with latency, efficiency, privacy, and security, limit their application to a handful of (high-end) use cases. As generative AI grows at a record-breaking speed, Large Language Models (LLMs) with billions of parameters place significant demands on computing infrastructure. The inference costs of large models increase correspondingly with the rising number of active users and their frequency of use. Running inference exclusively in the cloud is pricey and not sustainable for scaling. To achieve scalable, cost-effective, and pervasive intelligence in emerging AI applications, AI capabilities should be extended from the cloud to the devices, enabling billions of autonomous devices ranging from edge servers to small-scale microcontrollers. Embedded AI will revolutionise the way we provide automation, environmental monitoring, healthcare, agriculture, etc.

Integrating AI into embedded devices presents a major challenge because it demands fundamental changes to current AI ecosystems. While Cloud AI is well studied, transforming advanced Cloud AI models into highly-optimised, low-footprint models running on low-power, low-memory embedded devices interconnected in large-scale networks is largely unexplored, and the fundamental limits of distributed Embedded AI are not yet well understood. As the first Doctoral Network (DN) on Embedded AI, the primary goal of ANT is to train Doctoral Candidates (DCs) to co-create novel solutions overcoming the fundamental challenges of Embedded AI. Specifically, the ANT DCs design high performance distributed Embedded AI algorithms and systems under stringent resource constraints as well as adaptivity, scalability, robustness, privacy, security and explainability requirements. By focusing on tuning Embedded AI towards the particular mix of AI vertical requirements (energy efficiency, accuracy, security, privacy, reliability, scalability, low latency, etc.), the DCs enable unprecedented capabilities for intelligent networks to provide extremely high-accuracy, low-latency, low-energy and trustworthy Embedded AI solutions.

The EU Green Deal initiative and the European partnership on AI, Data and Robotics 8 have raised the urgent need for Embedded AI innovations. A skilled R&D workforce is the defining factor in winning the global race in high-profit markets. This is particularly critical for the EU where the industry is already suffering from a lack of highly skilled professionals9 . Embedded AI is an inherently multi-disciplinary topic whose mastery demands knowledge of computer science, electrical engineering, mechanical engineering, information technology, as well as domain knowledge of the relevant vertical industries such as robotics, environmental monitoring, digital health, agriculture technology. Unfortunately, such a multi-disciplinary training is currently unavailable in Europe.