From: Itamar Arel To: Jeffrey Epstein <[email protected]> Subject: AGI Research Project Date: Fri, 27 May 2011 03:40:58 +0000 Attachments: GMU Invitation_Arel.pdf Dear Jeffrey, It's been a while since we last corresponded and I hope you're doing well. I'm writing to you with regards to a research project I'd like to dedicate my time to during the coming academic year. I've been invited by the Krasnow Institute at George Mason University (see attached) to spend a sabbatical working in collaboration with researchers there on an ambitious AI research project. In particular, I've proposed to study how deep machine learning (as a model of the mammalian cortex) can be combined with reinforcement learning to yield a path to building truly intelligent machines. I intend to focus on the following two key topics: (1) Feedback mechanisms from the control subsystem to the perception (deep learning) subsystem to help guide improved representation. Current deep learning schemes assume representation is driven purely by regularities in the observations. I argue that while regularity is a strong driver, feedback from the control (i.e. decision-making) subsystem should also guide model learning, since it helps allocate (limited) resource to forming "relevant" representations. I would like to study this in detail, and test the solutions developed on challenging natural images. (2) Exploration vs. Exploitation - the balance between exploiting what is known and exploring the unknown remains a key challenge in reinforcement learning. I propose to have curiosity guide exploration, whereby the gradient of the modeling error will produce internal rewards that motivate exploration. I'd like to study this framework on robotic applications, showing that "boredom" is a driver for active learning. The Krasnow Institute houses several computational neuroscientists, whom I'd like to collaborate with as I develop methods to address the above two challenges. I'm very excit