September 16, 2016 12:00:00
Posted By Peter Bentley
|
Today we have a story in The Economist describing the work of one of my PhD students Haitham Baomar. The research focuses on the use of multiple neural networks for the control of aircraft. Unlike methods such as Deep Learning which are more suitable for vision-based classification tasks, this approach breaks the problem of flying aircraft under difficult conditions into smaller components, with a different neural circuit focussing on each aspect. It's a very effective method that learns the skills directly from human pilots. It's very fast to learn, and it's also much more suitable for safety-critical applications compared to black box methods such as Deep Learning, for it can be shown very clearly how each neural circuit controls each component. I'm always amazed how well the intelligent autopilot works. I have fun turning it off and putting the (simulated) aircraft into a nosedive in a storm with an engine fire - then activiting the intelligent autopilot, which immediately brings the plane under control and performs all necessary procedures. I'd trust this system long before I'd trust an autonomous car today. Check out the article here. And Haitham's research page here.
|