Title of the lecture
“How to make androids dream of electric sheep?”
A puzzling question, but one to take seriously if we ever want to build machines that think like you and me. The field of artificial intelligence experiences a lot of attention, mostly due to advances in computational infrastructure, finally allowing us to train deep neural networks effectively. But these machines are mostly stateless classifiers, trained beforehand, and lack the ability to learn over their lifetime. What I am looking for instead are machines that first observe the world, then create proper internal models, and finally use their experiences to improve their future decision making.
Neuroevolution is a method that allows us to optimise complex computational substrates (Markov Brains) to perform in various test environments – simply because evolution is the only known method that ever produced the intelligence we want to mimic in a machine. Instead of reverse-engineering the brain, I reverse engineer the process that made them. I present a sample of my work dealing with our initial question, how to make machines that have memory and rich internal representations.