MIT AGI: Cognitive Architecture (Nate Derbinsky)

In this MIT AGI lecture, Nate Derbinsky introduces cognitive architecture as one research approach toward AGI, situating it across neuroscience, psychology, cognitive science, and AI. He surveys the field from low-level biological models like Spaun up through psychological models like ACT-R and functional systems like Soar and Sigma. He tells a detailed research story about how the team gave Soar a human-inspired memory-activation mechanism and discovered that deliberate forgetting improved performance on robot mapping and a reinforcement-learning dice game. He closes with open problems including symbol grounding, transfer learning, and integrating deep learning, followed by an extended audience Q&A.

• Cognitive architecture means fixing your core assumptions about the mechanisms intelligent agents use across tasks, so interconnections add constraints that shrink the design space over time.
  00:11:32
• Soar is engineered for extreme efficiency, holding itself to a 50-millisecond per-cycle ceiling so agents stay reactive, with most cycles under a millisecond.
  00:34:01
• The team deliberately added forgetting to Soar, an architecture that does not aim to mimic humans, after realizing the human habit of forgetting might actually be useful.
  00:47:36
• On the mobile robot, the general forgetting mechanism outperformed hand-tuned task-specific cleanup rules at keeping the system under its 50ms threshold.
  00:55:54
• In the liar's dice game, the forgetting system cut memory needs by more than half while still playing as well as a system that forgot nothing.
  00:58:31
• Soar can be loaded with literally billions of rules and still turn over in under a millisecond because it only processes what changes.
  01:05:25
• In the huge dice state space, under 1% of value estimates ever got updated, so almost all forgotten values could be exactly reconstructed by recomputation.
  01:13:17
• The chunking mechanism generates new rules by memorizing the solution to a subgoal so the same situation never needs to be solved from scratch again.
  01:06:29

• AGI here means systems that persist over long periods, stay robust, learn over time, and handle tasks they were never told about in advance.
  00:01:32
• Allen Newell delivered the 1987 Harvard lectures, published in 1990 as Unified Theories of Cognition, founding the cognitive architecture idea.
  00:10:29
• Herb Simon won the Nobel Prize in Economics partly for bounded rationality, the idea that humans satisfice under constraints rather than optimize.
  00:15:11
• ACT-R encodes knowledge as if-then production rules and has been used to predict brain-area activation later confirmed with MRI scans.
  00:27:10
• ACT-R has over 1,100 papers, a Lisp main distribution, and ports to Java for robotics and Erlang for parallel processing.
  00:30:19
• Sigma reimplements Soar's core functionality using factor graphs and message passing, enabling modern machine learning, vision, and optimization.
  00:31:22
• Word sense disambiguation asks what a task-independent memory should return for an ambiguous word like 'run' given recency and frequency effects.
  00:48:38
• The Rosie project teaches a robot arm new tasks and concepts like color and 'move' through restricted natural language and pointing.
  00:40:50
• Spaun is a brain model with about two and a half million simulated neurons that can recognize images of numbers.
  00:21:27
• TacAir-Soar was one of the largest rule-based systems, with tens of thousands of rules running a 48-hour large-scale defense simulation.
  00:36:35