Lecture, Webinar
MIT AGI: Cognitive Architecture (Nate Derbinsky)
Lex FridmanNate Derbinsky, Chris Leisman, John Laird, Paul Rosenblum, Alan Newell, Herb Simon, John Anderson, Christian, Bonnie John, Edwin Olsen, Shivali Mohan, Brian
- Cognitive architectures are predicted to evolve toward systems capable of self-teaching via human input requests and hybrid models combining non-symbolic perceptual processing with symbolic logic for planning.
- Future systems will likely operate under bounded rationality constraints, requiring "satisficing" rather than optimal solutions due to limitations in computation, time, and memory.
- Specific architectures such as ACT-R, Sigma, and SOAR are expected to remain prevalent, with SOAR requiring processing cycles under 50 milliseconds for real-time human-interaction tasks.
- SOAR is anticipated to deploy across diverse platforms including mobile devices and real-time processors via C++, Java, and Python interfaces, while Sigma serves as the basis for the Virtual Human project.
- Applications for these systems span medicine, Human-Computer Interaction (HCI), robotics, defense simulations, natural language processing, and domains like Liars' Dice and mobile robotics.
- The Spawn model is planned to simulate aging processes and coordinate information flow, while the ROSI system aims to learn via multimodal natural language and gesture commands.
- A "forgetting" mechanism based on base-level activation is expected to reduce memory footprints for mobile robotics and game reinforcement learning, enabling operation on devices like the iPhone.
- The "Lumini" installation aims to demonstrate co-creative AI through a virtual dancer that learns human movements and improvises based on viewpoint theory.
- Integration of deep learning is projected to handle low-level visual tasks via APIs, while cognitive architectures manage structured decision-making, potentially using distributed representations like Word2Vec to bridge sensory and symbolic gaps.
- Future challenges include solving open issues in time-scale integration, transfer learning, and building multimodal representations, with progress potentially relying on efficient scaling configurations similar to Sigma.
- Achieving human-level AI is predicted to raise complex ethical and social issues comparable to servant-like synthetic beings, while the "chunking" mechanism in SOAR will generate new production rules through memoization.
- While specific algorithms like A* can ensure optimality within constrained search spaces, the field acknowledges no guarantees for complex decision-making across all tasks without further optimization.