Summary
Smart Sensors are key components for the upcoming Green and Digital European era. Recently, novel emerging electronics components – such as high energy-efficient many-core application processors featuring a power consumption of few tens of mWs – have enabled high-accurate on-device inference capabilities, i.e. Deep Learning inference, to extract high-level information from sensor data. However, this technology improvement is not sufficient to ensure robust solutions suitable for consumer and industrial applications. The main issue comes from the wide variety in real-world test conditions and, consequently, the lack at design-time of representative (labelled) sensor data, needed to train DL inference networks. For this reason, the currently used “train-once-and-deploy-everywhere” design process for edge intelligence has proved to be weak, even after an endless cyclic procedure involving data collection, model training and in-field testing.
This limitation is addressed by the SEA2Learn project by developing energy-efficient and real-time mechanisms to adapt the inference capabilities of resource-constrained smart sensors based on the stimulus from the surrounding environment. The proposed strategy, which is unprecedent in this domain, aims at placing in the same training loop multiple smart sensor nodes that interact with a Learning Agent. The latter will leverage a new class of lightweight methods belonging to the Continual Learning (CL) domain operating on unlabelled multi-sensor data. Thanks to the envisioned SEA2Learn framework, the embedded intelligence can adapt over time based on real-world data, making the design process more robust and 10-100x faster than today. To realize this vision, the fellow’s expertise in HW/SW design for embedded machine learning will be complemented by the Continual Learning knowledge of the hosting research group at KU Leuven and enriched by a tight collaboration with an SME that manufactures IoT platforms for edge computing.
This limitation is addressed by the SEA2Learn project by developing energy-efficient and real-time mechanisms to adapt the inference capabilities of resource-constrained smart sensors based on the stimulus from the surrounding environment. The proposed strategy, which is unprecedent in this domain, aims at placing in the same training loop multiple smart sensor nodes that interact with a Learning Agent. The latter will leverage a new class of lightweight methods belonging to the Continual Learning (CL) domain operating on unlabelled multi-sensor data. Thanks to the envisioned SEA2Learn framework, the embedded intelligence can adapt over time based on real-world data, making the design process more robust and 10-100x faster than today. To realize this vision, the fellow’s expertise in HW/SW design for embedded machine learning will be complemented by the Continual Learning knowledge of the hosting research group at KU Leuven and enriched by a tight collaboration with an SME that manufactures IoT platforms for edge computing.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101067475 |
| Start date: | 01-09-2022 |
| End date: | 31-08-2024 |
| Total budget - Public funding: | - 175 920,00 Euro |
Cordis data
Original description
Smart Sensors are key components for the upcoming Green and Digital European era. Recently, novel emerging electronics components – such as high energy-efficient many-core application processors featuring a power consumption of few tens of mWs – have enabled high-accurate on-device inference capabilities, i.e. Deep Learning inference, to extract high-level information from sensor data. However, this technology improvement is not sufficient to ensure robust solutions suitable for consumer and industrial applications. The main issue comes from the wide variety in real-world test conditions and, consequently, the lack at design-time of representative (labelled) sensor data, needed to train DL inference networks. For this reason, the currently used “train-once-and-deploy-everywhere” design process for edge intelligence has proved to be weak, even after an endless cyclic procedure involving data collection, model training and in-field testing.This limitation is addressed by the SEA2Learn project by developing energy-efficient and real-time mechanisms to adapt the inference capabilities of resource-constrained smart sensors based on the stimulus from the surrounding environment. The proposed strategy, which is unprecedent in this domain, aims at placing in the same training loop multiple smart sensor nodes that interact with a Learning Agent. The latter will leverage a new class of lightweight methods belonging to the Continual Learning (CL) domain operating on unlabelled multi-sensor data. Thanks to the envisioned SEA2Learn framework, the embedded intelligence can adapt over time based on real-world data, making the design process more robust and 10-100x faster than today. To realize this vision, the fellow’s expertise in HW/SW design for embedded machine learning will be complemented by the Continual Learning knowledge of the hosting research group at KU Leuven and enriched by a tight collaboration with an SME that manufactures IoT platforms for edge computing.
Status
SIGNEDCall topic
HORIZON-MSCA-2021-PF-01-01Update Date
09-02-2023
Images
No images available.
Geographical location(s)
Search
