Incremental Learning of Image Classification Models Over Low Data-Rate Networks

In the era of pervasive computing and the Internet of Things (IoT), the need for efficient Machine Learning (ML) model deployment and continuous learning in resourceconstrained environments has become increasingly critical. This paper emphasizes the significance of transferring ML models over low data-rate networks to facilitate Incremental Learning (IL) tasks. We propose a novel approach that leverages last-layer training and model compression techniques to reduce the size of deep learning models without significantly compromising their performance. By optimizing the transfer of these compressed models, we can enhance the efficiency of IL processes, allowing for real-time updates and adaptability to dynamic data distributions. Experiment results, using the popular MobileNet v2 and MCUNet models trained on an agriculture dataset, demonstrate an up to 5.5 percentage points improvement in F1-score and 6.3 percentage points improvement in Accuracy when the proposed IL approach is employed. At the same time, the size of the incremental part is only up to 456 bytes for MobileNet v2 and 158 bytes for MCUNet. By applying Huffman compression before model transfer, an extra 84% reduction is achieved, which makes the transfer over low data rate radio technologies such as LoRa more feasible. We also propose a communication protocol for model transfer, which is validated by practical experiments with IoT devices using LoRa transceivers. Experiment results show that the transfer can be completed and the new model can run on those devices within 5.7 to 30.1 seconds, depending on the LoRa settings.