Learning Modules

Learning in an AGI system is not a one-time "training phase." It is a continuous, multi-strategy process that happens as the agent interacts with the world.

The Three Pillars of AGI Learning

1. Perceptual Learning

Adapting how the system perceives the world.

• Mechanism: Fine-tuning neural weights for better vision, speech, or pattern recognition.
• Goal: Turning raw data into clean "Atoms" for the memory.

2. Structural Learning

Modifying the knowledge graph (AtomSpace) itself.

• Mechanism: Adding new concept nodes, strengthening links between related items.
• Goal: Building a more accurate mental model of reality.

3. Program/Skill Learning

Discovering new algorithms or behavioral sequences.

• Mechanism: Program evolution (MOSES) or Reinforcement Learning (RL).
• Goal: Learning how to do things more efficiently.

Learning Architecture

graph TD
    Experience[World Experience] -->|Perception| PL[Perceptual Learning]
    PL -->|Structured Facts| SL[Structural Learning]
    
    Experience -->|Reward Signal| RL[Policy/Goal Learning]
    
    subgraph Core Knowledge
    AtomSpace[AtomSpace]
    end
    
    SL --> AtomSpace
    RL --> AtomSpace
    
    AtomSpace -->|Expectations| PL

Sample-Efficient Learning

One of the biggest challenges for AGI is Sample Efficiency. Unlike big neural nets that need trillions of tokens, an AGI must be able to learn from a single demonstration (one-shot learning) by using its existing knowledge to fill in the gaps.

Metacognitive Learning

Advanced AGI systems learn about their own learning process (Meta-Learning).

• "Which reasoning strategy worked best for this type of problem?"
• "Do I need more data before I can make this decision?"

Next: Agent Loops