A Core Method for the Weak Completion Semantics with Skeptical Abduction (Extended Abstract)

Emmanuelle-Anna Dietz Saldanha, Steffen Hölldobler, Carroline Dewi Puspa Kencana Ramli, Luis Palacios MedinacelliIJCAI 2019 · DOI

TL;DR

This paper presents a novel connectionist (neural network) approach to implement "Weak Completion Semantics," a cognitive theory of human reasoning. It tackles the challenge of modeling human skepticism in logical deduction by integrating a mechanism for "skeptical abduction" into a neural network architecture, aiming to bridge the gap between symbolic logic and neural computation for cognitive modeling.

What Problem Does This Paper Address?

Human reasoning is complex and often deviates from purely logical deduction. While formal logic provides powerful tools for reasoning, it doesn't always capture the nuances of human thought, particularly when dealing with incomplete information or the need to explain observations. The "Weak Completion Semantics" is a cognitive theory that attempts to model human reasoning by using logic programs and a specific type of logical inference called "skeptical abduction." However, implementing this theory computationally, especially in a way that aligns with how neural networks process information, has been a significant challenge. This paper addresses the need for a connectionist realization of Weak Completion Semantics, specifically focusing on how to represent and compute logical inferences, handle three-valued logic (which allows for "unknown" truth values), and crucially, how to implement "skeptical abduction" within a neural network framework. Skeptical abduction is important because humans often don't consider all possible explanations for an observation; they tend to be skeptical and focus on a subset, a behavior that pure logical systems often miss.

Methods

The authors propose a connectionist architecture built upon the "CORE method," which uses recurrent neural networks to compute semantic operators for logic programs. This architecture is extended to handle three-valued Łukasiewicz logic. The core of their approach involves several interconnected subnetworks:

  • NA (Explanation Generator): A network designed to sequentially generate all possible "abducibles" (potential explanations) for a given observation. This is based on finite automata principles and implemented using binary threshold units.
  • NP (Least Model Computer): A recurrent neural network that computes the least model of a logic program combined with a specific explanation. It uses a feed-forward core that is made recurrent by feeding its output back to its input, allowing it to reach a stable state representing the least fixed point of the semantic operator.
  • Ns (Skeptical Conclusion Computer): A two-layer feed-forward network that aggregates the results from NP for all valid explanations. It determines if an atom is true, false, or unknown across all explanations.
  • Nc (Controller): A network that orchestrates the entire process, prompting NA, passing explanations to NP, checking for stable states and the validity of explanations against observations and integrity constraints, and managing the flow of information between subnetworks.

The key novelty lies in the integration of skeptical abduction. While NP computes models for individual explanations, Ns is designed to derive skeptical conclusions by only asserting a conclusion if it holds true across all valid explanations.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and computational realization rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Bridging Symbolic Logic and Neural Networks: The paper demonstrates a concrete pathway to implement complex symbolic reasoning (Weak Completion Semantics) within a connectionist framework, moving beyond simple pattern recognition.
  • Three-Valued Logic in Neural Networks: The extension of the CORE method to handle three-valued Łukasiewicz logic is crucial for modeling uncertainty and unknown states, which are common in human reasoning.
  • Skeptical Abduction as a Cognitive Mechanism: The core contribution is the implementation of skeptical abduction, reflecting the human tendency to not exhaustively test all possible explanations, which is vital for modeling human reasoning more accurately.
  • Modular Network Architecture: The decomposition of the reasoning process into specialized subnetworks (explanation generation, model computation, conclusion derivation, control) provides a structured and understandable approach to complex cognitive tasks.

Key Results

The paper outlines a theoretical framework and a computational architecture for Weak Completion Semantics with skeptical abduction. It describes how the proposed network can:

  • Represent logic programs and their weak completions.
  • Compute least models under three-valued Łukasiewicz logic.
  • Generate possible explanations for observations.
  • Derive skeptical conclusions by considering all valid explanations.

While specific quantitative results are not presented in this extended abstract, the paper establishes the plausibility of a connectionist realization for these complex reasoning tasks.

Limitations and Future Work

  • Computational Complexity: The paper acknowledges that skeptical conclusions can be exponentially complex in the number of abducibles. This suggests that the current approach might be computationally expensive for very large sets of abducibles.
  • Human-like Explanation Selection: The authors hypothesize that humans don't test all explanations systematically and suggest that a future direction is to develop an "attention formalism" to identify which explanations are actually tested, leading to "bounded skeptical abduction."
  • Experimental Validation: The current work is a theoretical specification. Future work will involve setting up experiments to test the hypothesis of bounded skeptical abduction and building networks that incorporate this mechanism.
  • Alternative Network Architectures: The paper mentions that Elman networks have been used to generate explanations in an arbitrary order, suggesting potential improvements in the explanation generation subnetwork.

Summary updated

BibTeX
@inproceedings{saldanha2019core,
  author = {Emmanuelle-Anna Dietz Saldanha and Steffen Hölldobler and Carroline Dewi Puspa Kencana Ramli and Luis Palacios Medinacelli},
  title = {{A Core Method for the Weak Completion Semantics with Skeptical Abduction (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/887},
}
Figure