Periodic Reporting for period 1 - EdgeAI (Edge AI Technologies for Optimised Performance Embedded Processing)
Berichtszeitraum: 2022-12-01 bis 2023-12-31
Work on advanced research topics, published as a SoA - Paper: Narges Norouzi, Sveta Orlova, Daan de Geus, and Gijs Dubbelman, "ALGM: Adaptive Local-then-Global Token Merging for Efficient Semantic Segmentation with Plain Vision Transformers", accepted for publication in the proceedings of IEEE/CvF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, US, 17-21 June 2024.
Beyond SoA activities include novel AI computing architectures based on new neuromorphic approaches, including techniques to boost the micro-architecture's energy efficiency without sacrificing attack/fault robustness. Investigate Sparse Stream Semantic Registers for their applicability in accelerating the operations necessary for Spiking Neural Networks (SNNs). Exploration of hybrid digital/analogue concepts based on Resistive Sum, tailored for executing MAC operations within Fully Connected Neural Networks.
Two novel wireless node designs based on two types of LoRa solutions (sub-gigahertz and mesh 2.4GHz) including designing and developing a hybrid AI-based HW platform combining several processing units, mesh connectivity and multi-sensing IIoT devices.
Analysing explainability methods for resource-constrained hardware and explaining results for model-agnostic and model-specific techniques. Determining the main challenges in distributed hardware for explainable and interpretable decisions.
Digital twin development for synthetic data generator and train AI on edge. Simulate events based on sensor input and inference results generated at the edge.
Usage of Generative AI to optimise and augment datasets.
Use different Neural Architecture Searches (NAS) to improve the deployment of neural networks on different edge AI platforms.
Development of mechanisms for secure transfer of ML models between different devices in a federated learning context.