Most of today’s TinyML solutions carry out the inference at the edge, where data are acquired, while they entirely demand training to external resources. However, empowering edge devices with the ability to learn from local data has very important implications in terms of prediction and privacy as it would allow adapting a pretrained model to specific users or time-changing conditions without sharing data externally.
In this work, we present a framework in C programming language to train CNNs on STM32 microcontrollers. We adopt our framework to successfully personalize a 1D-CNN for Human Activity Recognition and we provide a software tool to estimate the memory and computational resources required to accomplish model personalization.
Schedule
Timezone: PDT
Enabling on-device learning on STM32 microcontrollers
Beatrice ROSSI, Research Scientist
STMicroelectronics
Michele CRAIGHERO, PhD Student
Politecnico di Milano
Beatrice ROSSI, Research Scientist
STMicroelectronics
Beatrice Rossi graduated in Mathematics and Applications at Università degli Studi di Milano Bicocca in 2008. Since then, she has been working in STMicroelectronics, System Research and Applications. Her research interests include Edge AI, Tiny Machine and Deep Learning, and Distributed Ledger Technology for the IoT.
Michele CRAIGHERO, PhD Student
Politecnico di Milano
Michele Craighero graduated in Computer Science and Engineering at Politecnico di Milano in 2022 and he is currently in the first year of his PhD. His research project is titled “Learning and Adaptation in Distributed Environments” and it is a collaboration between Politecnico di Milano and STMicroelectronics. His research interests include Machine Learning techniques for Time Series Classification, Change Detection and Unsupervised Domain Adaptation.
Schedule subject to change without notice.