An Efficient Machine Learning System for Connected Vehicles

Abstract

In online machine learning systems, a computer gathers sets of training data from some data sources and trains a machine learning model every it gets a set. In the cases that the time required to gather a set is long such as the case of the data gathering from connected vehicles, the delay for reflecting the observed environmental values included in the training data to the model lengthens. In this paper, we propose a system to reduce the delay for reflecting observed environmental values to the models suppressing the increase of the validation loss. Our proposed system cyclically broadcasts the parameters of the machine learning model to the data sources and the data sources calculate the result of the loss function for their observed training data. We evaluated the proposed system assuming a machine learning system for connected vehicles. The vehicles of that training data give a larger value to the result than a given threshold send the training data to the computer for training the machine learning model. Our experimental evaluation revealed that our proposed methods can achieve lower validation loss values than a conventional method.