The driverless startup Cruise today presented a self-developed tool – the Continuous Learning Machine – that does forecasting tasks on the go. Cruise claims that the Continuous Learning Machine, which automatically tags and breaks down training data, enables some AI models to guide Cruise’s self-driving cars to predict things like whether bikes will hit traffic or kids run on the streets.
One of the challenges of autonomous vehicles is predicting intent. People don’t always obey the rules of the road, and even when they do, they tend to bend those rules. According to the US National Highway Traffic Safety Administration, 94% of serious accidents are due to driver errors or dangerous decisions.
For this reason, Cruise built a continuous learning machine. Using a technique called active learning, errors in perceptual models running on cruise vehicles are automatically identified and only scenarios with a significant difference between prediction and reality are added to the training datasets. According to Cruise, this enables highly targeted data mining and minimizes the number of “simple”
The Continuous Learning Machine also autonomously labels data using model predictions as the “basic truth” for all scenarios. In essence, the framework observes what a person or vehicle might do in the future and compares that to what they are actually doing. The final step is to train a new model, test it, and use it on the road while ensuring that it performs better than the previous model.
According to Cruise, the continuous learning machine has enabled her to make highly accurate predictions for a number of rare scenarios that her models encounter in the real world. These include turns, which Cruise’s cars see an average of less than 100 times a day, and cuts when people change their trajectory to avoid slowdowns or stationary objects. Another example are K-curves – three-point curves in which the driver has to maneuver forwards and backwards. According to Cruise, these are about half as common as U-turns.
“Our machine learning prediction system needs to generalize to both entirely novel events and events it very rarely sees,” wrote Sean Harris, Cruise’s senior engineering manager, in a blog post. “We need to understand the intent of other agents on the street as well as the reasons behind the sequence and interaction between different agents and how they evolve over time. The complexity of this problem is a separate research area. This is another reason why autonomous vehicles are the greatest technical challenge of our generation. “