Cognitive Architectures

Cognitive architectures provide unified frameworks for building complete intelligent systems. Unlike narrow AI approaches, they aim to integrate perception, learning, reasoning, and action into coherent agents.

What is a Cognitive Architecture?

A cognitive architecture specifies:

• Fixed Structure: Core mechanisms that don't change
• Learning Components: What improves with experience
• Knowledge Representation: How information is stored
• Control Flow: How processing is orchestrated
• Integration: How components interact

Classical Architectures

SOAR (State, Operator, And Result)

Overview: Production system architecture developed since 1983.

Key Features:

• Production Rules: IF-THEN patterns
• Working Memory: Current state
• Long-Term Memory: Procedural, semantic, episodic
• Impasse-Driven Learning: Learning when stuck

Strengths:

• Well-studied and mature
• Strong problem-solving capabilities
• Chunking for speedup learning

Applications: Game playing, robotics, tutoring systems

ACT-R (Adaptive Control of Thought-Rational)

Overview: Cognitive architecture modeling human cognition.

Key Features:

• Modules: Declarative, procedural, perceptual, motor
• Buffers: Limited-capacity working memory
• Subsymbolic Level: Activation, utilities, learning
• Production Compilation: Learning by doing

Strengths:

• Cognitive plausibility
• Predictive modeling of human performance
• Strong psychological grounding

Applications: Human-computer interaction, cognitive modeling, education

CLARION (Connectionist Learning with Adaptive Rule Induction ON-line)

Overview: Hybrid architecture with explicit and implicit knowledge.

Key Features:

• Dual Representation: Symbolic and subsymbolic (neural)
• Bottom-Up Learning: Extracting rules from neural nets
• Top-Down Learning: Refining explicit knowledge
• Metacognition: Monitoring and control

Strengths:

• Integrates symbolic and connectionist approaches
• Models implicit learning

Symbolic vs. Subsymbolic

Symbolic Architectures

• Representation: Discrete symbols and rules
• Reasoning: Logical inference, search
• Learning: Rule induction, chunking
• Strengths: Explainability, compositionality
• Weaknesses: Brittleness, grounding problem

Subsymbolic (Connectionist)

• Representation: Distributed patterns in neural networks
• Reasoning: Pattern matching, constraint satisfaction
• Learning: Gradient descent, backpropagation
• Strengths: Robustness, learning from data
• Weaknesses: Interpretability, catastrophic forgetting

Hybrid Architectures

• Combining Strengths: Neural perception + symbolic reasoning
• Examples: CLARION, NARS, OpenCog, Sigma
• Challenge: Seamless integration

Modern Cognitive Architectures

NARS (Non-Axiomatic Reasoning System)

Overview: Reasoning system designed for AGI with limited resources.

Key Features:

• Non-Axiomatic Logic: Uncertain, inconsistent knowledge
• Experience-Grounded Semantics: Pragmatic truth
• Resource-Conscious: Time and space constraints
• Real-Time Adaptation: Continuous learning

Strengths:

• Handles uncertainty naturally
• Flexible knowledge representation
• Designed for open-world environments

OpenCog

Overview: Integrative AGI framework (predecessor to Hyperon).

Key Components:

• AtomSpace: Hypergraph knowledge store
• PLN (Probabilistic Logic Networks): Uncertain inference
• MOSES: Program evolution
• Neural-symbolic integration

Strengths:

• Multiple reasoning paradigms
• Extensible architecture
• Active research community

Sigma

Overview: Integrative architecture based on graphical models.

Key Features:

• Unified Representation: Factor graphs
• Multiple Levels: Reactive, routine, deliberative, reflective
• Perception to Cognition: End-to-end integration

Strengths:

• Probabilistic reasoning throughout
• Cognitive hierarchy

Lida (Learning Intelligent Distribution Agent)

Overview: Biologically inspired cognitive cycle.

Key Features:

• Global Workspace Theory: Consciousness model
• Cognitive Cycle: ~10 Hz processing
• Action Selection: Behavior networks

Strengths:

• Models consciousness mechanisms
• Continuous operation

Functional Components

Perception

• Sensory Processing: Vision, audition, touch
• Feature Extraction: Pattern recognition
• Object Recognition: Identification and categorization
• Integration: Multimodal fusion

Memory Systems

• Working Memory: Current focus, limited capacity
• Episodic Memory: Personal experiences, events
• Semantic Memory: Facts and concepts
• Procedural Memory: Skills and habits

Reasoning

• Deductive: Logical inference
• Inductive: Generalization from examples
• Abductive: Explanation generation
• Analogical: Transfer from similar situations

Learning

• Supervised: Learning from labeled data
• Unsupervised: Pattern discovery
• Reinforcement: Learning from rewards
• Transfer: Applying knowledge to new domains

Action Selection

• Planning: Goal-directed behavior
• Reactive: Immediate responses
• Habits: Automatic behaviors
• Deliberation: Weighing alternatives

AGI-Oriented Architectures

Requirements for AGI

• Generality: Operate across diverse domains
• Autonomy: Self-directed learning and adaptation
• Efficiency: Practical resource usage
• Robustness: Handle uncertainty and novelty
• Scalability: Grow with experience

Hyperon (Next-generation OpenCog)

Modern AGI architecture covered in detail in Hyperon section.

Key innovations:

• MeTTa language for knowledge and code
• Distributed hypergraph
• Pattern matching and rewriting
• Integration of neural and symbolic

Evaluation and Benchmarks

Criteria

• Task Performance: Accuracy on benchmarks
• Learning Speed: Sample efficiency
• Transfer: Generalization to new tasks
• Explainability: Interpretability of decisions
• Resource Efficiency: Computational requirements

Benchmarks

• General Intelligence Tests: Raven's matrices, ARC
• Multi-Task Learning: Meta-learning benchmarks
• Commonsense Reasoning: Winograd schemas, COPA
• Embodied AI: Robotics, simulation environments

Design Principles

Unified Cognition

Single framework for all cognitive phenomena (Newell's hypothesis).

Bounded Rationality

Satisficing under resource constraints (Simon).

Embodiment

Cognition arises from sensorimotor interaction (embodied cognition).

Developmental Learning

Learning trajectories from simple to complex (constructivism).

Recommended Resources

• Unified Theories of Cognition - Allen Newell
• The Cambridge Handbook of Cognitive Architectures - Müller (Ed.)
• Artificial General Intelligence - Goertzel and Pennachin
• How to Build a Brain - Chris Eliasmith

---

Next: Information Theory