Do Neurons Dream of Primitive Operators? Wake-Sleep Compression Rediscovers Schank’s Event Semantics
Summary: arXiv:2603.25975v1 Announce Type: cross
In a groundbreaking study, researchers have explored the potential for discovering event primitives through an innovative approach known as wake-sleep compression. This research revisits Roger Schank’s conceptual dependency theory, which posits that all events can be decomposed into fundamental operations, such as ATRANS, PTRANS, and MTRANS, which were originally hand-coded based on linguistic intuition. The primary question posed is whether these primitives can be automatically identified through the lens of compression pressure.
The Wake-Sleep Approach
The study adapts DreamCoder’s wake-sleep learning methodology to analyze event state transformations. The system operates on the premise of using events represented as pairs of before and after world states. During the “wake” phase, the system identifies operator compositions that adequately explain each event. In the subsequent “sleep” phase, recurring patterns are extracted and optimized under the Minimum Description Length (MDL) principle, leading to the discovery of new operators.
Findings and Discoveries
Starting with four generic primitives, the research successfully identified operators that directly correspond to Schank’s original concepts. These include:
- MOVE_PROP_has = ATRANS
- CHANGE_location = PTRANS
- SET_knows = MTRANS
- SET_consumed = INGEST
Additionally, the study revealed the existence of compound operators, such as “mail,” which combines ATRANS and PTRANS, as well as novel emotional state operators that were absent from Schank’s taxonomy.
Validation and Results
The findings were validated against both synthetic events and real-world commonsense data derived from the ATOMIC knowledge graph. The results showed that:
- On synthetic data, the newly discovered operators achieved Bayesian MDL within 4% of Schank’s hand-coded primitives while successfully explaining 100% of the events, compared to Schank’s 81%.
- In the ATOMIC dataset, the contrast was even more pronounced, with Schank’s primitives explaining only 10% of naturalistic events, while the discovered library accounted for 100%.
Shifting Perspectives on Event Primitives
Interestingly, the dominant operators identified were not the physical-action primitives one might expect, but rather mental and emotional state changes, including:
- CHANGE_wants (20%)
- CHANGE_feels (18%)
- CHANGE_is (18%)
None of these mental and emotional state operators were included in Schank’s original taxonomy, suggesting that the landscape of event primitives is not only richer than previously understood but also places greater emphasis on cognitive processes in naturalistic data.
Conclusion
This research provides the first empirical evidence supporting the idea that event primitives can indeed be derived from compression pressure. It solidifies the notion that Schank’s core primitives are not only information-theoretically justified but that a more comprehensive inventory exists, fundamentally altering our understanding of event semantics in artificial intelligence.