Shadows and headless shadows: a worlds-based, autobiographical approach to reasoning

Ladislau BölöniarXiv 2012 · OpenReview

TL;DR

This paper introduces Xapagy, a novel cognitive architecture that reasons about the world exclusively through an autobiographical memory, eschewing traditional rule-based systems. It proposes a "shadowing" mechanism where current events are mentally projected onto past experiences, and "headless shadows" to predict future events or infer hidden information, offering a unique, fragmented view of reality.

What Problem Does This Paper Address?

Many AI systems rely on structured knowledge bases, explicit rules, or multiple, distinct models to represent and reason about the world. This approach can lead to brittle systems that struggle with ambiguity, novel situations, and the nuanced, often incomplete, nature of human memory and reasoning. The authors aim to build a cognitive system that mimics human-like story understanding and reasoning by leveraging a more fluid, memory-centric approach. They question how a system can reason and predict without explicit rules, relying solely on a raw, chronological record of experiences.

Methods

Xapagy's core is its autobiographical memory, which stores events as "scenes" without any post-hoc generalization into rules. Reasoning occurs through two key mechanisms:

  • Shadowing: When the system encounters a new event (in its "focus"), it searches its autobiographical memory for similar past events. These past events are "shadowed" onto the current event, creating a mental projection of how this situation might have unfolded or been experienced before.
  • Headless Shadows (HLSs): These are extrapolated shadows that don't directly correspond to a current event. They represent predictions of future events, inferences about hidden causes or relationships, or summaries of past events. HLSs are built from "shadow VI relatives" (SVRs) and "shadow VI relative interpretations" (SVRIs), which are essentially templates for potential future events or inferred relationships, weighted by their likelihood based on past experiences.

The system also employs "spike activities" (instantaneous operations) and "diffusion activities" (gradual changes) to maintain and update these shadows. A key design choice is the fragmented representation of entities, including the self, where a single real-world entity might be represented by multiple instances across different scenes.

Datasets

Not discussed in the paper. The authors mention manually translating a 500-word version of "Little Red Riding Hood" into their Xapi language, creating a synthetic autobiography for the agent to learn from. This serves as a demonstration of their approach rather than a formal dataset for evaluation.

Key Insights and Ideas

  • Autobiography as the sole foundation: The radical idea that a rich, nuanced understanding of the world can emerge solely from a raw, chronological memory, without explicit rule extraction or generalization.
  • Fragmented Self and Entities: The notion that representing entities (including the agent's "self") as multiple, loosely connected instances across time and context is a more natural way to model how humans might experience and recall information, leading to a less monolithic but potentially more flexible representation.
  • Reasoning through "What If" and "What Then": Xapagy's reasoning is framed around predicting potential outcomes (headless shadows) and inferring missing information, mirroring human intuitive reasoning about stories and events.
  • "Surprise" as a measure of novelty: The system defines surprise not just by unexpected events, but by the degree of reorganization required in its shadow structures, suggesting a more sophisticated understanding of novelty than simple deviation.

Key Results

The paper is primarily a conceptual and architectural proposal, not a quantitative evaluation. The key "results" are qualitative demonstrations:

  • The system can represent complex narratives like "Little Red Riding Hood" using its scene-based autobiographical memory.
  • It can generate predictions about story continuations (e.g., the wolf eating LRRH) and infer hidden events based on its shadowing and headless shadow mechanisms.
  • The authors illustrate how the system might handle unexpected events (like a sneeze) and major plot points (like the wolf eating LRRH), showing how these impact the system's internal state and predictions.

Limitations and Future Work

  • Lack of quantitative evaluation: The paper presents a theoretical framework and a qualitative demonstration. There are no empirical results or benchmarks to assess Xapagy's performance against other systems.
  • Manual translation effort: The manual translation of the LRRH story into Xapi highlights the current labor-intensive nature of populating the system with narrative data.
  • Scalability: While the authors suggest the autobiographical memory is computationally manageable, the practical scalability of the shadowing and HLS mechanisms for very large autobiographies remains an open question.
  • Future Work: The authors suggest exploring how to automate the creation of synthetic autobiographies, further refining the HLS mechanism for different types of reasoning (e.g., inference of missing actions), and investigating the implications of their fragmented self-representation.

Summary updated

Abstract

Many cognitive systems deploy multiple, closed, individually consistent models which can represent interpretations of the present state of the world, moments in the past, possible futures or alternate versions of reality. While they appear under different names, these structures can be grouped under the general term of worlds. The Xapagy architecture is a story-oriented cognitive system which relies exclusively on the autobiographical memory implemented as a raw collection of events organized into world-type structures called {\em scenes}. The system performs reasoning by shadowing current events with events from the autobiography. The shadows are then extrapolated into headless shadows corresponding to predictions, hidden events or inferred relations.
BibTeX
@misc{boloni2012shadows,
  author = {Ladislau Bölöni},
  title = {{Shadows and headless shadows: a worlds-based, autobiographical approach to reasoning}},
  year = {2012},
  eprint = {1211.5643},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1211.5643},
}
Figure