Efficient Training on Edge Devices Using Online Quantization
Michale H.Ostertag, Sarah Al-Doweesh and Tajana Rosing
ABSTRACT
Sensor-specific Calibration function offer superior perfomance over global models and single-step calibration procedure but require prohibitive levels of sampling in the input feature space. Sensor self-calibraation by gathering training data through collaborative calibration or self-analyzing preditive results allows these sensors to gather sufficient information. Resource-Constrained edge devices are then stuck between high communition costs for transmitting training data to a centralized sever and high memory requirments for storing data locally. We propose online dataset quantization that maximizes the diversity of input features maintaining a representative set of data from a larger stream of training data points. We test the effetiveness of online dataset quntization on two real-world datasets:air qualily calibration and power prediction modeling. Online Dataset Quantization out preforms reservoir sampling and performs eqully to ofline methods.
