| name | discovery-by-dreaming-cross-domain-recombination |
| description | A skill for implementing cross-domain recombination inspired by dreaming, based on the paper "Discovery by Dreaming: Cross-Domain Recombination in Artificial Memory" (arXiv:2607.16256). This skill outlines how to implement a LoRA fine-tuning pipeline (DREAMS) and a symbolic engine (SAPIENCE) to recombine knowledge across domains, enhancing AI discovery and insight generation. |
| activation | discovery by dreaming, cross-domain recombination, dream-inspired AI, memory consolidation AI |
Discovery by Dreaming: Cross-Domain Recombination in Artificial Memory
Overview
This skill implements the methodology from arXiv:2607.16256, which proposes that dreaming (offline recombination of experiences) is not merely for memory consolidation but for discovering novel cross-domain connections. The approach consists of two complementary systems:
- DREAMS: A LoRA fine-tuning pipeline that recombines neural network weights.
- SAPIENCE: A symbolic engine that replays structured knowledge objects.
Both systems generate cross-domain associations that improve performance on tasks requiring transfer learning, such as unseen math reasoning.
Core Concepts
- Recombination vs. Rehearsal: Within-domain rehearsal does not yield discovery; cross-domain recombination does.
- Substrate-General: The principle applies across neural networks and symbolic systems.
- Falsifiable Prediction: Hippocampal recordings can distinguish recombination from rehearsal.
Implementation Steps
1. DREAMS (LoRA Fine-Tuning Pipeline)
- Prepare Dataset: Collect a diverse corpus spanning multiple domains (e.g., math, language, science).
- LoRA Configuration: Apply Low-Rank Adaptation to the target model (e.g., LLaMA, Mistral).
- Recombination Objective: During training, sample batches that mix examples from different domains to encourage cross-domain weight updates.
- Training Loop: Train for a limited number of epochs, ensuring the model does not overfit to any single domain.
- Evaluation: Test on tasks requiring cross-domain transfer (e.g., GSM8K for math reasoning) and compare to baseline.
2. SAPIENCE (Symbolic Engine)
- Knowledge Representation: Encode knowledge as structured objects (e.g., semantic triples, frames).
- Replay Mechanism: Periodically activate a recombination process that randomly pairs knowledge objects from different domains.
- Composition Rule: Define how to combine objects (e.g., concatenation, property transfer, analogy mapping).
- Output Generation: Generate novel hybrid concepts and evaluate their usefulness via downstream tasks or human judgment.
Validation
- Quantitative: Measure improvement on benchmark tasks (e.g., +21 pp for symbolic arm, +5.64 pp overall for neural arm, +14.5 pp on cross-domain subtasks).
- Qualitative: Assess novelty and usefulness of generated combinations.
- Neuroscience Alignment: Predict hippocampal replay patterns that favor cross-domain co-activation.
Usage Notes
- Ensure computational resources for LoRA training (GPU memory).
- For symbolic engine, define a clear knowledge schema and recombination operators.
- Combine both approaches for synergistic effects.
References
- arXiv:2607.16256 - Discovery by Dreaming: Cross-Domain Recombination in Artificial Memory
Activation Keywords
dreaming, recombination, cross-domain, memory consolidation, LoRA, symbolic AI, neuroscience-inspired AI