STARTLE, developed by Mike Hook and colleagues at Roke Manor Research of Romsey in Hampshire, UK, employs an artificial neural network to look out for abnormal or inconsistent data. Once it has been taught what is out of the ordinary, it can recognise dangers in the environment.
For instance, from data fed by a robotic vehicle’s on-board sensors, STARTLE could notice a pothole and pass a warning to the vehicle’s control system to focus more computing resources on that part of the road.
“If it sees something anomalous then investigative processing is cued; this allows us to use computationally expensive algorithms only when needed for assessing possible threats, rather than responding equally to everything,” says Hook.
This design mimics the amygdala, which provides a rapid response to threats. The amygdala helps small animals to deal with complex, fast-changing surroundings, allowing them to ignore most sensory stimuli. “The key is that it’s for spotting anomalous conditions,” says Hook, “not routine ones.”