On-Device Training of Fully Quantized Deep Neural Networks on Cortex-M Microcontrollers
Talk, Embedded Talk: Deep Learning on Narrow Resources, University Erlangen-Nürnberg, Germany
This talk discusses the potential of, as well as the technical challenges involved in, training fully quantized deep neural networks (DNNs) on Cortex-M microcontroller units (MCUs). First, the talk presents the additional challenges that arise due to the tight resource constraints and limited computing power of MCUs, as well as methods to overcome memory and computing constraints. Second, variational autoencoders are evaluated for their potential to train DNN classifiers unsupervised, i.e., without labels, and initial results are discussed.
More information | Slides