Using Cached-DFL, the researchers developed a quasi-social network where cars can view each other's discovery profiles. (Image credit: Hiroshi Watanabe/Getty Images)
Scientists have found a way to allow self-driving cars to exchange information while driving without having to establish direct connections.
“Cached Decentralized Federated Learning” (Cached-DFL) is a framework for sharing artificial intelligence (AI) models between autonomous vehicles, enabling them to overtake each other and share accurate and up-to-date information. This sharing includes the latest navigation techniques, traffic patterns, road conditions, and road signs and signals.
Typically, cars must stay in virtual proximity to each other and grant permission to share driving information collected during their journeys. But with Cached-DFL, the researchers created a quasi-social network in which cars can view each other’s driving profiles — all without sharing any driver information or driving patterns.
Currently, self-driving cars rely on data stored in a single central repository, which also increases the risk of major data breaches. The Cached-DFL system allows vehicles to integrate data into trained AI models that contain information about driving conditions and scenarios.
“Think of it as creating a shared experience network for self-driving cars,” said Dr. Yun Liu, the project’s principal investigator and a professor of engineering at NYU’s Tandon School of Engineering. “A car that has only driven in Manhattan can now learn about road conditions in Brooklyn from other vehicles, even if it has never been there itself.”
Cars can share information about how they handle scenarios similar to those in Brooklyn that may occur on roads in other boroughs. For example, if Brooklyn has oval potholes, cars can share how to handle them no matter where they are.
The scientists uploaded their research to the arXiv preprint database on August 26, 2024, and presented their findings at the Association for the Advancement of Artificial Intelligence conference on February 27.
The Key to Better Self-Driving Cars
In a series of experiments, scientists found that fast and frequent communication between self-driving cars improves the efficiency and accuracy of driving data.
The scientists created 100 virtual self-driving cars in a simulated version of Manhattan and had them “drive” in a semi-random pattern. Each car had 10 AI models that were updated every 120 seconds, and this is where the caching part of the experiment came in. The cars store data and wait to share it until a suitable vehicle-to-vehicle (V2V) connection is established. This is different from traditional self-driving car data sharing models, which are instantaneous and do not store or cache information.
The scientists plotted the speed at which the cars learned and tested whether Cached-DFL outperformed the centralized data systems common in today’s self-driving cars. They found that as long as the cars were within 100 meters (328 feet) of each other, they were able to view and share information. The vehicles didn’t need to know about each other to share data.
“Scalability is one of the key benefits of decentralized FL,” Dr. Jie Xu, an assistant professor of electrical and computer engineering at the University of Florida, told Live Science. “Instead of every car communicating with a central server or every other car, each car only exchanges model updates with the cars it communicates with. This approach to localized data exchange prevents communication overhead from growing exponentially as more cars are connected to the network.”
The researchers suggest that Cached-DFL will make self-driving technology more accessible by reducing the need for computing power, since the computing load is distributed across many vehicles rather than concentrated on a single server.
Sourse: www.livescience.com
