Skip to content

mossrake/learning-system

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Learning Through Narratives: Episodic Compression to Latent Variables for Sample-Efficient Intelligence

A three-part research paper proposing an alternative approach to sample-efficient learning that leverages narrative-based experience rather than embodied interaction.

Overview

Current machine learning systems require orders of magnitude more training examples than humans to achieve comparable competence. While LeCun (2022) argues that embodied physical interaction is necessary for sample-efficient learning, we observe that humans learn extensively from narratives—therapists understand conditions they haven't experienced, children learn dangers they haven't encountered, and scientists build on literature they haven't personally validated.

This work proposes episodic compression: a mechanism where experiences compress into narrative-formatted episodes marked by salience (novelty/surprise), which generate generalizations through pattern detection. These episodes and generalizations serve as learned latent variables, enabling sample-efficient learning from narratives using existing pre-trained language models.

Getting Started

New to this work? Start with the Reader's Guide which provides:

  • Discipline-specific reading recommendations
  • What we are and aren't claiming in each domain
  • Common misunderstandings to avoid
  • How the three parts fit together

Want a quick overview? See the Summary for a condensed version covering:

  • The sample efficiency challenge
  • Episodic compression mechanism
  • Experimental validation results
  • Scaling architecture
  • Multi-timescale learning properties
  • Practical implications and limitations

Paper Structure

Part 1: Core Mechanism and Validation

File: episodic-compression-part1.pdf

  • Introduces the episodic compression mechanism (experiences → episodes → generalizations)
  • Validates the approach using the Shimmer Valleys artificial world
  • Demonstrates a system with 9 generalizations and 4 episodes successfully predicting outcomes in 5 novel scenarios
  • Shows appropriate confidence calibration and multi-step causal reasoning
  • Requires no neural network training—leverages pre-trained LLMs

Key contribution: Demonstrates that narrative learning with episodic compression can achieve sample efficiency without embodied interaction.

Part 2: Scaling Architecture

File: episodic-compression-part2.pdf

  • Addresses scaling beyond "toy scale" demonstrations
  • Proposes a two-layer architecture:
    • Intuition layer: Small model (1-3B params) with large context (100K-200K tokens) for fast pattern-matching
    • Executive layer: Large capable model with standard context for deliberate reasoning and consolidation
  • Introduces "micro-sleep" consolidation cycles for continuous learning
  • Provides implementation specifications and guidance
  • Compares with RAG, JEPA, and other approaches

Key contribution: Practical architecture for deploying episodic compression at scale with continuous learning during operation.

Part 3: Multi-Timescale Learning

File: episodic-compression-part3.pdf

  • Explores emergent temporal properties of the architecture
  • Examines learning across timescales: real-time (seconds), episodic formation (hours), consolidation (days), long-term accumulation (months-years)
  • Addresses catastrophic forgetting (shows why traditional problem doesn't apply)
  • Discusses deployment learning, personalization, and continuous adaptation
  • Identifies open research questions

Key contribution: Positions episodic compression as a framework for continuous learning during deployment, addressing gaps in current ML approaches.

Supporting Materials

The Shimmer Valleys World

File: shimmer-valleys.pdf

An artificial world with novel causal rules used for experimental validation:

  • Entities: Globs, Whisps, Resonators, Flutter Seeds, Shade Pools, Living Glass
  • Internally consistent but unprecedented causal rules
  • No correspondence to physics or pre-training knowledge
  • Enables testing without pre-training contamination

Learning System Components

File: shimmer-valleys-learning-components.pdf

Detailed breakdown showing:

  • Raw experiences (stream of interactions)
  • Episodes (salient compressed summaries)
  • Generalizations (enduring patterns)
  • Example of how experiences compress through the system

Key Claims

  1. Humans learn extensively from narratives rather than requiring direct experience
  2. LLMs provide suitable cognitive substrates due to pre-training and attention mechanisms
  3. Episodic compression achieves sample-efficient learning through progressive compression
  4. Two-layer architecture enables scaling beyond single-context capacity
  5. Architecture naturally operates across multiple timescales for continuous learning
  6. Approach requires no new neural network training

Key Distinctions

vs. LeCun's JEPA:

  • Learning source: Narratives vs. embodied interaction
  • Training: Uses pre-trained models vs. trains from scratch
  • Latent variables: Symbolic (via consolidation) vs. continuous (via gradient descent)
  • Timeline: Immediate deployment vs. weeks of training

vs. RAG:

  • Retrieval: Pattern-matching on causal structure vs. semantic similarity
  • Learning: Dynamic (forms new generalizations) vs. static
  • Novelty: Explicitly detects when patterns don't match vs. always returns best match

vs. Traditional Continual Learning:

  • Catastrophic forgetting: Doesn't apply (additive learning, no weight updates)
  • Mechanism: Consolidation through reasoning vs. gradient descent
  • Transparency: Human-readable generalizations vs. opaque weight changes

Limitations

  • Empirical validation limited to toy scale (Part 1 only)
  • Long-term deployment (months-years) unproven
  • Optimal parameters require empirical tuning
  • Scaling thresholds remain estimates
  • Real-world performance across diverse domains untested

Future Research Directions

  • Extended real-world deployment studies
  • Consolidation optimization and scheduling
  • Hierarchical knowledge organization
  • Memory management strategies
  • Multi-agent and transfer learning
  • Safety and adversarial robustness
  • Theoretical foundations

Citation

Part 1 - Mossrake Group, LLC. (2025a). Learning Through Narratives: Episodic Compression to Latent Variables for Sample-Efficient Intelligence.
Part 2 - Mossrake Group, LLC. (2025b). Scaling episodic compression: A two-layer architecture for continuous learning.
Part 3 - Mossrake Group, LLC. (2025c). Multi-Timescale Continuous Learning: Emergent Properties of Episodic Compression.

License

© 2025 Mossrake Group, LLC


Note: This is conceptual and theoretical work. Part 1 provides experimental validation at toy scale. Parts 2 and 3 are primarily architectural proposals and discussion requiring comprehensive empirical validation.

About

Episodic Compression to Latent Variables for Sample-Efficient Intelligence

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published