Abstract
This paper explores the formal relationship between connectionist models of reflexive reasoning and classical symbolic logic. By introducing Pointwise Isolated Connections (PICs), the authors demonstrate that the efficiency of neural-inspired systems like SAM can be replicated within the Connection Method of automated theorem proving. This equivalence suggests that traditional logic engines, when properly optimized for parallelism, can achieve the rapid "common sense" reasoning characteristic of human cognition.
Introduction: The AI Paradox
A long-standing challenge in artificial intelligence is the "AI Paradox": humans perform complex "reflexive" reasoning—such as understanding natural language or navigating social contexts—almost instantaneously, while classical automated theorem provers often struggle with even simple tasks. Connectionist models, such as Shastri and Ajjanagadde's SAM (System for Assembling Models), were developed to address this by using temporal synchrony for variable binding, allowing for fast, parallel inference. However, these models were often viewed as distinct from formal logic. This paper bridges that gap by asking if a formal logical property, termed adequateness, allows a theorem prover to solve simple problems with the same speed as connectionist systems.
Bridging Connectionism and the Connection Method
The authors focus on two primary frameworks:
- SAM (System for Assembling Models): A connectionist system performing backward reasoning over Horn clauses, restricted to ensure computational efficiency.
- The Connection Method: A classical proof procedure that identifies "spanning sets" of connections between logical literals to establish validity.
To unify these, the authors introduce Pointwise Isolated Connections (PICs). While standard reduction techniques in logic simplify formulas based on unique matches for literals, PICs operate at the level of individual predicate arguments (points). If an argument is a constant or uniquely bound, it can be unified and propagated across the knowledge base in parallel.
Key Insights: Reflexive Reasoning as Logic Reduction
The central thesis is that the "reflexive reasoning" observed in connectionist systems is essentially the parallel application of specific logical reductions.
- Adequateness: A proof method is considered "adequate" if its time complexity scales with the problem's inherent difficulty rather than the total size of the knowledge base. The inclusion of PICs makes the Connection Method adequate for reflexive reasoning tasks.
- Massive Parallelism: The Connection Method is naturally suited for parallel architectures. In such environments, PIC reductions can occur simultaneously across millions of facts, potentially outperforming connectionist models that propagate information step-by-step through rule chains.
Results and Formal Equivalence
The authors prove that the operational steps of SAM—specifically constant propagation and the propagation of success/failure labels—correspond exactly to reductions within the Connection Method.
Key findings include:
- Soundness: SAM is formally sound because its operations are a subset of valid logical reductions.
- Complexity: With PICs, the Connection Method can solve reflexive reasoning tasks in parallel time proportional to the shortest proof length, with space requirements bounded by the knowledge base size.
Conclusion and Limitations
While PICs provide a path toward high-speed logical reasoning, the approach remains "incomplete" to maintain efficiency. For instance, assuming a rule is used only once prevents infinite loops but may miss certain proofs. Furthermore, dynamic constraints on variable usage must be monitored during runtime. Despite these limitations, the work demonstrates that symbolic logic remains a powerful and efficient tool for modeling human-like reflexive reasoning when implemented with appropriate parallel reductions.
Summary updated