"Exploring techniques to build efficient and robust TinyML deployments"
Lead Principal Systems Engineer
Data is key to designing effective deep learning applications, where characteristics and availability of data vary from application to application. Edge deployment of deep learning methods requires privacy, low power usage, and robustness against out-of-distribution data. Furthermore, data for training and deployment tasks, often referred to as the training dataset and the calibration dataset, respectively, may not be available in some applications. In this talk, trade-offs between power and performance, given the availability of training data for supervised learning, will be highlighted. In addition, a dynamic fixed-point quantization scheme suitable for edge deployment in absence of sufficient calibration data will be presented, and trade-offs in compute resource for quantization, such as memory and cycles, will be discussed. Finally, edge deployment architecture utilizing deep learning methods to handle out-of-distribution data due to sensor degradation and alien operating conditions will be presented.