by Ilan Shaviv, Ph.D. – Imagry CTO
In the dynamic world of autonomous driving, the choice between supervised and unsupervised learning methods plays a pivotal role in shaping the capabilities and reliability of self-driving vehicles. At Imagry, our commitment to pushing the boundaries of innovation has led us to adopt a supervised learning approach, setting us apart from the unsupervised methods employed by some other players in the industry.
Let’s delve into the nuances of these two approaches and explore why Imagry’s emphasis on supervised learning is a game-changer.
Supervised vs. Unsupervised Learning: A Fundamental Distinction
The fundamental difference between supervised and unsupervised learning lies in the utilization of labeled training data. In an unsupervised learning approach, the vehicle relies on actual driving experiences by human drivers and perception of the world, deriving patterns and relationships automatically. A prominent example of this is Tesla’s “unsupervised learning” method, where data from participating vehicles is collected over the cloud to continuously train their self-driving option. However, this approach may lead to imperfect learning, as demonstrated by instances of less-than-ideal driving behaviors observed in the unfiltered data (such as rolling “stops” at stop signs, and driving 60 in a 55 mph zone.)
In stark contrast, Imagry employs a powerful supervised learning methodology, utilizing multiple neural networks to emulate the cognitive processes and decision-making of skilled and law-abiding human drivers. This is based on extensive visual data collected continuously from 2019 (today Imagry is collecting data from demo vehicles driving autonomously on public roads in the U.S., Germany, Japan, and Israel), and annotating them consistently with a predefined policy using proprietary in-house developed software and human oversight. The neural networks learn by imitating the driving path solutions generated by human drivers and filtered for non-desirable behaviors by Imagry personnel, ensuring that the Imagry autonomous driving system aligns seamlessly with the highest driving standards.
Other Differentiators in the AI Learning Approaches
Typically, autonomous driving solutions that employ unsupervised learning techniques are based on a monolithic neural network that handles all tasks in a single, large network. This approach simplifies training and reduces effort, but it creates a big “black box” with low explainability and high maintenance complexity. The Imagry solution, on the other hand, which is comprised of two steps (perception and motion planning), uses multiple small neural networks to understand and respond to the visual cues. This modular approach, where different tasks are handled by separate networks, offers better traceability, easier maintenance, and adaptability.
Ultimately, the perception and motion planning development processes used in the Imagry solution incorporate a semi-automated approach, whereby proprietary in-house tools are used to automate and therefore speed up the labeling process. However, all data undergoes thorough human review before being integrated into the neural networks’ learning process. This emphasis on supervised learning reinforces Imagry’s dedication to precision and reliability.