Crowd-sourced machine learning model update through connected vehicles
Autores:
Fecha: 01.05.2023
Abstract
The refinement of production machine learning models (ML) is a huge challenge, as it requires
gathering data about outlier or corner cases that make the model underperform or even fail. In the
concrete case of Connected, Cooperative and Automated Mobility (CCAM), the impact of these corner
cases can be critical as they can cause safety issues. Current approaches for dataset generation are
mainly based on large recording campaigns by dedicated vehicles recording massive multi-sensor data.
With the advent of edge and 5G technologies, sensor data can be filtered and processed closer to the
source. This paper presents a novel concept where vehicles, instead of storing locally or streaming all
the sensor data, select the data that make the ML model perform below an established threshold and
send it to the Cloud through the MEC. In the MEC the data is anonymized and prepared for sharing
with third parties.
BIB_text
title = {Crowd-sourced machine learning model update through connected vehicles},
pages = {10 },
keywds = {
5G, MEC, Cloud, CCAM, Machine Learning
}
abstract = {
The refinement of production machine learning models (ML) is a huge challenge, as it requires
gathering data about outlier or corner cases that make the model underperform or even fail. In the
concrete case of Connected, Cooperative and Automated Mobility (CCAM), the impact of these corner
cases can be critical as they can cause safety issues. Current approaches for dataset generation are
mainly based on large recording campaigns by dedicated vehicles recording massive multi-sensor data.
With the advent of edge and 5G technologies, sensor data can be filtered and processed closer to the
source. This paper presents a novel concept where vehicles, instead of storing locally or streaming all
the sensor data, select the data that make the ML model perform below an established threshold and
send it to the Cloud through the MEC. In the MEC the data is anonymized and prepared for sharing
with third parties.
}
date = {2023-05-01},
}
