IRT Nanoelec tackles the topic of artificial intelligence (AI) from the perspective of embedded electronics. The development and expansion of embedded AI applications is triggered by the generation of large amounts of data that require local processing at component and system levels. The aim is to provide these components with the ability to make decisions in a more decentralized, autonomous and reliable manner. Therefore, processing must be integrated close to the sensor in order to optimze the flow of data and to ensure that it remains undamaged and confidential. This allows a good trade off between data flows from the sensor to the user, the related energy foo prints of both the sensor and the central computer handling reduced amounts of data. The teams involved at Nanoelec mainly focus on image sensors and on the security of AI components and systems.
Imaging is one of the three main fields that use AI. The aim is not only to achieve better image quality, but also to extract relevant data from the image, taking account of the environment, the object and the scene (potentially across a range of lighting conditions), and knowledge of the context. The main objective of the
new Nanoelec/Smart Imager program is to evaluate the advantages of using 3D-stack technology to integrate the processing of artificial intelligence into the third layer of an image sensor. The research teams focus
on developing generic AI building blocks and the associated processing as well as on exploring the impact of
these blocks on low-energy imager architectures.
Cyber-security of microcontrollers
Future connected objects will need reliable implementations of embedded AI algorithms and security mechanisms to protect them from potential software and hardware threats. Securing the implementation of AI algorithms within these objects and the communications with the outside world is critical to the deployment and safe use of this technology in embedded systems. Researchers from academic labs as well as industrial teams are working together within the Nanoelec/Pulse program to make the implementation and deployment of AI algorithms, in particular machine-learning algorithms embedded in IoT components. Concerning the safety of autonomous vehicle, program teams are interested in the validation of artificial intelligence (AI) systems for mobility / vehicles. The question of evaluating and validating the effectiveness of these technologies is one of the last central issues delaying the adoption of embedded context capture technologies. It is therefore a major issue for the market access of autonomous vehicles.
Imaging is one of the three main fields that use AI. The aim is not only to achieve better image quality, but also to extract relevant data from the image, taking account of the environment, the object and the scene (potentially across a range of lighting conditions), and knowledge of the context. The main objective of the new Nanoelec/Smart Imager program is to evaluate the advantages of using 3D-stack technology to integrate the processing of artificial intelligence into the third layer of an image sensor. The research teams focus on developing generic AI building blocks and the associated processing as well as on exploring the impact of these blocks on low-energy imager architectures.
Core partners of Nanoelec/Photonic Sensors program are participating in the European Neoteric project, which aims to design and build a cytometry demonstrator using photonic circuits. Cytometry is a technique for the qualitative and quantitative counting and characterization of particles, molecules or biological cells in a fluid. Neoteric consists in demonstrating that a photonic circuit with learning functions (AI) can better perform image recognition of counted particles. Compared to conventional image analysis using deep learning software, silicon photonic technology promises
an increase in frame rate, improved classification performance, and lower energy consumption by several orders of magnitude.
Processing characterization images
The development of deep learning algorithms and the use of artificial intelligence are of great interest for characterizing electronic components
and systems. In particular, these technologies will help improve the effectiveness of experiments conducted with large instruments available under the Nanoelec/PAC-G platform. These experiments generate very large data sets. The pooled processing of data, metadata, simulations and images using AI in the associated images can accelerate convergence towards the development of required materials and aim at reducing the environmental footprint, ensuring supply sovereignty and operational reliability.
Application demonstrators for imagers
Teams involved in the Nanoelec/SystemLab initiative define, develop and test cases using embedded AI technologies that are integrated into or associated with image sensors. The project aims to develop functional demonstrators and explore new applications scenarios. One of these use cases consists in positioning wireless image sensors in variuous natural environment to monitor fires or measure biodiversity, for instance. Another subject uses image sensors capable of analyzing human postures and expressions, with the aim of detecting emotions, to warn against hazardous situations or prevent improper conduct. And a third use case will involve detecting obstacles, such as potholes and patches of ice for vehicles.
Learning AI on open hardware Risc-V
The Nanoelec/Human Capital and Training Design program led to a new training course on hardware-software codesign, thanks to the use of a free processor based on the Risc-V instruction set. This instruction set and the processors implementing it are royaltyfree and are increasingly used by the industry worldwide. In 2020, the teaching team at Grenoble-INP, again within the context of Nanoelec, developed a semi-generic reference platform based on Risc-V, designed a vision system embedded on FPGA and an FPGA for hardware processing of artificial intelligence algorithms. The combined software/hardware design, up to the actual prototype, allows the students to experiment with the respective advantages of these technologies (flexibility of software, efficiency of hardware) and to find a trade off between the two, within the same system.