III: Small: Opportunistic Learning on Wheels: Peer-wise Training of Machine Learning Models among Connected Vehicles

点击翻译按钮获取中文摘要

Vehicles are becoming computers on wheels, instrumented with rich varieties of sensors producing large amounts of complex data for driving control, road and weather conditions. Machine learning models become indispensable to enable the vehicles to interpret the data, recognize the environments and make control decisions. Training robust, accurate models in vehicular environments pose new challenges due to constant vehicle mobility; low-rate, unreliable, intermittent peer-wise and cloud communication; and extremely diverse road/weather conditions. Recent federated learning methods have started to address constrained cloud communication. However, they do not exploit peer collaboration among resource-rich vehicles, or consider spatial-temporal correlation in data quality, quantity and diversity, thus cannot fully utilize the rich datasets produced. This project specifically exploits unique opportunities in vehicular environments to training robust, accurate models using rich datasets. It will support the education and training of both undergraduate and graduate students in the important discipline of modern machine learning methods, the mentoring of high school students in scientific exploration, and exchange of latest advances with the industry and possible pilot studies for technology transfer.This project will create a novel opportunistic learning (OL) paradigm: peer-wise vicinity vehicles exploit the opportunistic nature of encounters, connectivity and data availability, and use that to orchestrate peer collaboration and cloud assistance to train robust, accurate machine learning models jointly. Vicinity vehicles exchange training updates peer-wise and with the cloud to produce aggregate models. Based on short/long term vehicular mobility and connectivity, vehicles and road-side boxes migrate or pre-fetch training states to handle departure and arrival of vehicles. Incentive mechanisms derived from explore-exploit learning of drivers' cost distributions will motivate vehicle participation, maximize training utility while preserving driver privacy. This project will produce a holistic set of foundational algorithms and systems for training accurate, robust machine learning models in highly dynamic vehicular and edge environments beyond what existing federated learning methods can achieve.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.