On-Device Training of Deep Learning Models on Edge Microcontrollers
Authors:
- 1 August 2022
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 62-69
Abstract
Recent advancements in Artificial Intelligence (AI) together with the increase of device computational power catalyzed the widespread diffusion of Intelligent Cyber Physical Systems (ICPSs) as a novel solution to run smart applications where the inference is performed on the Micro Controller Unit (MCU). Unfortunately, the hardware constraints of these devices limit the tasks that can be accomplished, for this reason, most of the times hybrid Cloud/Edge approaches have been proposed to enable the offloading of those operations (e.g., the training of deep learning models) that would not fit the computation, memory, and energy requirements. However, when security, latency, and connection stability aspects are relevant, it is clear that the use of such an approach is no longer suitable. This becomes even more evident in those contexts where it is asked the detection of new data patterns that can emerge over time. Focusing our attention on the analysis of time series data, in this paper we propose a novel Echo State Network (ESN) model that enables the on-device training directly on the MCUs of the STM32 family. Starting from a custom layer implementation on Keras, we extended the functionalities of the X-CUBE-AI tool provided by STMicroelectronics to recognize this type of network and enable the automatic deployment on the smart boards. To test the proposed solution, we realized a testbed and conducted a set of experiments that demonstrate the feasibility of our MCU on-device training approach which reached a good level of precision, recall, and F1-score.Keywords
This publication has 12 references indexed in Scilit:
- Trimming Feature Extraction and Inference for MCU-Based Edge NILM: A Systematic ApproachIEEE Transactions on Industrial Informatics, 2021
- Intelligent Fault Diagnosis for Large-Scale Rotating Machines Using Binarized Deep Neural Networks and Random ForestsIEEE Transactions on Automation Science and Engineering, 2021
- Leveraging Stack4Things for Federated Learning in Intelligent Cyber Physical SystemsJournal of Sensor and Actuator Networks, 2020
- Quantitative Analysis of Deep Leaf: a Plant Disease Detector on the Smart Edge2020
- A Low-Delay Lightweight Recurrent Neural Network (LLRNN) for Rotating Machinery Fault DiagnosisSensors, 2019
- The Rise of Intelligent Cyber-Physical SystemsComputer, 2017
- Echo State Property of Deep Reservoir Computing NetworksCognitive Computation, 2017
- Detecting changes at the sensor level in cyber-physical systems: Methodology and technological implementation2017
- A Unified Framework for Reservoir Computing and Extreme Learning Machines based on a Single Time-delayed NeuronScientific Reports, 2015
- Optimization and applications of echo state networks with leaky- integrator neuronsNeural Networks, 2007