MIT AGI: Cognitive Architecture (Nate Derbinsky)

Core Motivation and AGI Definition

• Goal: The speaker defines AGI as systems exhibiting human-level intelligence that persist over long periods, operate robustly across varying conditions, and learn continuously without pre-definition of tasks.
• Inspiration: Current AI (e.g., Alexa) lacks the ability to engage in teach-back dialogues (e.g., "teach me how to write a script"), a capability the speaker aims to develop via systems like "Rosie."
• Research Trajectory: The field moves from observing human behavior (cognitive modeling) to predicting human responses (applications in medicine, HCI, defense) and finally to building functional systems that perform tasks beyond current machine learning capabilities.
• Distinction from Standard AI: Unlike "rational agents" (logic-based) or standard "black-box" ML (optimal action focus), cognitive architecture researchers aim to replicate the process of human thought, allowing for predictions regarding reaction times, error patterns, and emotional states.

Historical Context and Theoretical Foundations

• Alan Newell's Unified Theories: Cited as the foundational shift from disparate psychological theories to "Unified Theories of Cognition" (cognitive architecture), which posits that fixed mechanisms (memory, learning, representation) can be integrated into a single theory to constrain design space.
• Lactoseum Science: A proposed methodology where a "core" of beliefs grows as assumptions survive testing across multiple tasks, moving from ephemeral hypotheses to established constraints.
• Time Scales of Action: Proposes that cognitive phenomena occur at distinct scales:
  • Neuronal/Substrate: Neuroscience, cognitive neuroscience.
  • Deliberate Acts: Psychology and cognitive science (seconds to minutes).
  • Social Interaction: Sociology and economics (hours to years).
• Bounded Rationality: Based on Herbert Simon's Nobel-winning concept, arguing humans cannot solve complex problems optimally due to constraints in time, memory, and computation; systems must instead "satisfice" (find satisfactory solutions).
• Symbol System Hypothesis: The view that intelligent behavior can be generated via the manipulation of symbols; while necessary, modern approaches acknowledge the need for non-symbolic (statistical/ML) processing for perception.

Key Cognitive Architectures and Approaches

• Biological Level (Substrate):
  • Spaun (Spawn): A large-scale model simulating 2.5 million neurons; integrates perception (vision) and motor control (arm movement) to categorize images and mimic handwriting styles.
  • Libra: Another architecture modeling low-level biological processes and neuronal firing rates.
• Psychological Level (Cognitive Modeling):
  • ACT-R: Developed by John Anderson; focuses on predicting human errors, reaction times, and brain activity (validated via MRI). Widely used in Human-Computer Interaction (HCI) for interface design.
  • Sigma: A newer architecture by Paul Rosenblum (formerly SOAR); re-implements SOAR logic using factor graphs and message passing to better integrate modern machine learning and optimization techniques.
• Functional Level (Task Execution):
  • SOAR: The primary focus of the talk; emphasizes efficiency (cycle times under 50ms), symbolic representation, and modularity.
  • Applications: Included R1 (computer configuration), TAC Air (defense simulation), mobile robotics (Splinter Bot), and a mobile app for playing Liars' Dice.

SOAR Architecture: Mechanisms and Implementation

• Processing Cycle:
  • Input: Perception updates a directed graph (working memory).
  • Matching: Production rules (if-then) fire in parallel to identify potential actions.
  • Selection: A decision procedure ranks preferences to select a single operator.
  • Output: The system executes actions (internal or external) and updates memory.
  • Constraint: The entire cycle must complete within 50ms to maintain human-like reactivity.
• Memory Systems:
  • Procedural: Production rules for task execution.
  • Declarative: Facts and episodic memories (stored as experiences).
  • Chunking: A learning mechanism that compiles sub-goal reasoning into new rules (memoization).
• Efficiency Heuristics:
  • Change Detection: The system only processes rules when relevant data in the working memory changes, allowing billions of rules to run in under 1ms.
  • Approximation: Used to calculate memory activation decay without re-computing all history at every step, ensuring performance remains high.

Case Studies on "Forgetting" and Memory Management

• Rational Analysis of Memory: Humans exhibit recency and frequency effects in memory retrieval; SOAR implemented these as "base level activation" decay to prioritize relevant memories.
• Word Sense Disambiguation: Implementing decay mechanisms improved performance in NLP tasks by naturally prioritizing contextually relevant meanings.
• Mobile Robotics (Mapping):
  • Problem: Large working memory (maps) slowed the robot as it aged.
  • Solution: Forgetting old, low-activation map segments; reconstruction was possible from semantic memory (room connectivity).
  • Result: Outperformed hand-tuned cleanup rules and maintained operation under 50ms.
• Reinforcement Learning (Liars' Dice):
  • Problem: Massive state space caused memory overflow (2GB) on mobile devices.
  • Solution: Forgetting low-activation value estimates unless they had been recently updated.
  • Result: Reduced memory usage to <1GB (fitting on an iPhone) while maintaining game-winning performance comparable to non-forgetting systems.

Integration with Machine Learning and Future Challenges

• Hybrid Systems: SOAR does not replace deep learning but interfaces with it (e.g., using CNNs for object recognition while SOAR handles symbolic reasoning).
• Symbol Grounding: A key open challenge is bridging the gap between low-level non-symbolic (pixel) data and high-level symbolic reasoning.
  • Spaun addresses this via distributed representations.
  • Future Work: Exploring vector representations (similar to Word2Vec) to allow symbolic architectures to operate on continuous sensory inputs.
• Metacognition: Current research includes agents that can self-assess when knowledge is harmful and override it, though active "learning to forget" is less developed.
• Open Issues:
  • Transfer Learning: Difficulty combining distinct theories or modules across different tasks.
  • Scalability: Cognitive architectures often require complex software engineering compared to the scalability of pure deep learning models.
  • Multi-Agent Systems: Lack of built-in theories for how multiple cognitive agents should share and fuse knowledge.

Q&A Highlights

• Action Selection: Uses a fixed, simple decision procedure based on "acceptable operator preferences" pushed up from high-level knowledge to keep cycle times low.
• Learning to Forget: No current mechanism for the system to learn when to forget; it relies on static decay thresholds, though metacognition allows overriding bad habits.
• AGI Viability: Cognitive architectures are seen as "brave" but face the challenge of rare "home run" commercial applications compared to narrow AI successes; they prioritize understanding general intelligence over solving specific narrow tasks.