AToM: An Analogical Theory of Mind

Irina RabkinaIJCAI 2017 · DOI

TL;DR

This paper introduces ATOM, a novel computational model for Theory of Mind (ToM) that leverages analogical reasoning. By grounding ToM in structure-mapping principles, ATOM aims to enable AI systems to understand and predict human mental states, a crucial step towards more natural human-AI interaction.

What Problem Does This Paper Address?

Humans possess a remarkable ability called Theory of Mind (ToM), which allows us to infer the beliefs, desires, and intentions of others. This "mind-reading" capability is fundamental to social interaction, enabling us to predict behavior and navigate complex social dynamics. For artificial intelligence systems to truly integrate into human society, they must also acquire this social intelligence. While ToM has been extensively studied in psychology, developing computational models that can replicate this ability remains a significant challenge. Existing approaches often struggle to capture the nuanced and analogical nature of human reasoning about mental states. This paper proposes ATOM as a computational framework to address this gap, building upon established theories of analogy to model ToM development and function.

Methods

ATOM is built upon the Structure-Mapping Engine (SME), a computational model of Structure-Mapping Theory (SMT), which is a foundational theory of analogy. It also incorporates elements from the SAGE model of analogical generalization and the MAC/FAC model of analogical retrieval.

The core idea is to represent knowledge as predicate calculus cases in a long-term memory (LTM). When a situation requiring ToM reasoning arises, ATOM uses MAC/FAC to retrieve relevant past experiences from LTM.

  • Retrieval: Relevant cases are retrieved from LTM.
  • Interim Generalization: If a retrieved case is a specific event, an interim generalization is created in working memory. This involves projecting candidate inferences from the retrieved case onto the current situation and re-representing parts of the current situation based on these inferences.
  • Feedback Loop: ATOM seeks feedback on its reasoning in natural language, mimicking human interaction.
  • Generalization and Storage: Based on feedback, ATOM uses SAGE to generalize the original probe with the retrieved case, creating more abstract and robust schemas that are stored back in LTM. This iterative process allows ATOM's ToM abilities to improve over time.

ATOM's approach to interim generalization is a novel adaptation, focusing on projecting inferences and re-representing the target case.

Datasets

Not discussed in the paper. The paper describes modeling studies that draw upon existing research on children's ToM development and pretend play, rather than using specific datasets for training or evaluation.

Key Insights and Ideas

  • Analogy as the Foundation of ToM: The paper posits that analogical reasoning, specifically through structure-mapping, is a core mechanism underlying Theory of Mind. This is a significant departure from purely symbolic or statistical approaches to ToM.
  • Iterative Learning and Generalization: ATOM's learning process is iterative. By retrieving, reasoning, receiving feedback, and generalizing, the model continuously refines its understanding of mental states, mirroring human cognitive development.
  • Pretense as a Testbed for ToM: The paper highlights the utility of modeling pretend play as a simplified yet effective domain for testing and validating ToM mechanisms, particularly candidate inference verification and re-representation.
  • Bridging Cognitive Science and AI: ATOM aims to bridge theoretical insights from cognitive science (e.g., Bach's work on ToM and SMT) with practical AI implementation, potentially leading to more human-like AI.

Key Results

The paper presents two completed studies demonstrating the plausibility of ATOM:

  • Pretense Modeling: ATOM's mechanisms were tested by modeling two studies of pretend play. The model's results were consistent with both successful and failed instances of pretend play observed in children, providing support for ATOM's underlying mechanisms.
  • ToM Training Study: ATOM was able to successfully perform a series of false belief tasks (a standard measure of ToM development) using only three input vignettes. This suggests that ATOM's mechanisms are cognitively plausible and that the resulting ToM reasoning is comparable to human reasoning.

Limitations and Future Work

  • Computational Appropriateness and Cognitive Plausibility: A major challenge identified is testing ATOM in a way that is both computationally efficient and cognitively realistic. The current tests use simplified tasks, and the authors acknowledge that completing these tasks is not sufficient for a complete ToM.
  • Identifying Additional Cognitive Tasks: The authors are actively working on identifying more complex cognitive tasks to further test ATOM.
  • Full Implementation and Testing: The current work is an extended abstract. A full implementation and comprehensive testing of ATOM are planned as future steps.
  • Integration with Companion Architecture: ATOM is intended to be integrated into the Companion cognitive architecture to enhance its interactive abilities, which is a significant future direction.

Summary updated

BibTeX
@inproceedings{rabkina2017atom,
  author = {Irina Rabkina},
  title = {{AToM: An Analogical Theory of Mind}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/761},
}
Figure