Using a contest to learn how to help a machine learn is at the heart of a course, Deep Learning for Self-Driving Cars, that Lex Fridman conceived of for MIT students.  As Fridman says, the point of the course is to “make deep learning accessible to beginners.” The resources and lectures are available for all, even non-MIT students to view, at selfdrivingcars.mit.edu.

What makes Fridman excited about machine learning is,

“The idea that you can teach a machine to automate things; to do the boring, hard things for you, while you get to do the fun things.”

The work is applicable to more than just driverless vehicles, as the virtual cars are analogous to packets in a broadband network, so it isn’t too difficult to imagine creating similar models for dynamic, machine-driven, bandwidth management.

Fridman points out that it is relatively easy to begin programming one’s first neural network, as all that is needed is a web browser. He explains that understanding the complex math behind the neural network isn’t necessary to master this course. Students learn the basics of reinforcement learning – a Pavlov-type approach of rewards and penalties the neural network receives to help it teach itself and train it to achieve the end goal.

And the goal of Fridman’s contest is for the entrant to build a neural network to drive a virtual car as close as possible to the speed limit on a high-traffic road. To receive credit for the course, students’ virtual cars must drive a minimum of 65 mph. The first contest drew over 12,000 entries from people ranging from high school students to middle-aged managers from companies involved in autonomous vehicle development.

The latest competition, which runs through June/July, is expected to draw even more entries. There are multiple prizes, but the biggest prize may be in developing a neural network that enables the virtual car to drive a minimum speed of 65 mph. This could show potential employers that the contest participant is capable of doing MIT-level work, without necessarily attending MIT; and this accessibility to learning and measurement of learning may be the biggest impact of all from this innovative course.