Research Area:  Machine Learning

Abstract:

Class incremental learning and Internet of Things (IoT) are related but have not yet been extensively integrated. With the exponential rise in the number of IoT devices, the total amount of data produced is massive, thus it is infeasible to send all the data directly to the cloud. Training deep learning models in an incremental manner can help saving the training time and reduce the computational load required to train the model. In this work, we develop and demonstrate an incremental class learning system using only features of images to reduce large memory footprints in comparison with using raw images. Our experiments show that the proposed method achieves similar accuracy within a shorter time frame. We also develop a distributed training method of convolutional neural networks (CNNs) between an IoT edge device and a cloud. The IoT edge device carries out feature extraction to reduce the transmission load to the cloud which also eases the training cost. The cloud carries out the classifier training.

Keywords:  

Author(s) Name:  Swaraj Dube, Yee Wan Wong, Jeen Ghee Khor, Hermawan Nugroho

Publisher name:  Springer

DOI:  https://doi.org/10.1007/978-3-030-70917-4_34

Volume Information:  

Paper Link:   https://link.springer.com/chapter/10.1007/978-3-030-70917-4_34