Sunday, October 2, 2011
I watched the first Stanford Machine Learning video lecture this afternoon. The pace and professionalism were impressive. Dr. Ng is well-motivated and clearly focused on the subject matter--he never even introduced himself or the course! He does a great job of motivating you with stories of ML applications like web search, OCR for mail sorting, and robotics, including a video of a helicopter doing cartwheels a flopping itself upside down for a while. However, I felt like Dr. Ng was stretching a bit when he said "we couldn't figure out how to write a control algorithm for this helicopter so we implemented a learning algorithm." It is VERY impressive that they were able to implement an ML system for flying a helicopter--especially considering how "learning" mistakes could have ended the experiment prematurely. The grad students must've spent a lot of time in the lab rebuilding trashed RC helicopters. Its random acrobatics clearly show the "unintended consequences" and "control space exploration" that often emerge from an ML system. However, Dr. Ng and his students had trouble finding a conventional control law for helicopters he'd have been smart to check with the professors at my alma-mater, Georgia Tech, where students were implementing autonomous helicopter flying algorithms back in the 90's. They were even doing precision retrieval and placement of payloads, autonomously. And a friend at UC Bolder was doing work with quadracopters and other popeller-driven craft to autonomously plan and execute optimal navigation around obstacles and constraints. These were not ML algorithms, just plain old optimal control. But Dr. Ng accomplished his goal, he got my attention and I'm excited about the class. His enthusiasm for Machine Learning and robotics is contagious.