Introduction
This paper, originally presented at AAAI-94, challenges the traditional "symbolic" baggage often carried into neural network design. It argues that by forcing rigid distinctions between rules and facts, or types and predicates, researchers actually cripple the reasoning power of connectionist systems. The author proposes a uniform connectionist representation that collapses these traditional AI silos into a single architecture of "feature nodes" and "link bundles."
The Problem: Rigid Symbolic Distinctions
In the early 1990s, researchers sought to bridge the gap between connectionism (neural networks) and symbolic reasoning (logic). A leading system at the time, developed by Ajjanagadde and Shastri (A&S), utilized "activation synchrony"—where nodes firing in the same time-slice are considered "bound" to each other—to perform rapid logical inferences.
However, the A&S system inherited rigid distinctions from classical logic that created architectural asymmetries:
- Rules vs. Facts: Rules (general) and facts (specific) were encoded using entirely different structures.
- Roles vs. Fillers: The "slots" in a relation (like giver) were treated differently than the objects filling them (like John).
- Types vs. Predicates: Categories (like Cat) were handled in a separate "type hierarchy" module, distinct from relational knowledge (like Prey-on).
By splitting these into different "bins," systems became asymmetrical. For instance, a system might prove a statement was true but struggle to "extract" an answer (e.g., proving someone bought a book, but failing to identify John as the buyer).
A Uniform Connectionist Architecture
The author proposes an Alternative Scheme of Representation that treats all knowledge uniformly. Instead of separate modules, the system utilizes a unified network of feature nodes and link bundles.
In this scheme, an object is represented as a collection of "basis features" (microfeatures). Crucially, everything is treated as a rule. A "fact" like John loves Mary is simply a highly specific rule: if (features of John) AND (features of Mary), then (Love relationship). To query the network, one clamps "assertion" nodes at high activation and "target" nodes at low levels; if the target's activation rises, the proposition is proved.
This uniform approach allows for bidirectional "answer extraction" and class-based inference without specialized circuitry. Inhibitory logic within "link bundles" ensures binding consistency, preventing the network from incorrectly associating unrelated entities.
Key Insights and Results
- The Generality Continuum: Facts are just rules at the extreme end of a spectrum of specificity. Treating them as such enables bidirectional binding, allowing information to flow from the rule to the fact and back again.
- Types as Unary Predicates: There is no need for a separate type hierarchy. Saying "a cat is an animal" is functionally identical to the rule
cat(x) → animal(x). This allows the system to reason about classes using the same mechanism it uses for relations. - Functional Superiority: The proposed system handles queries that previous systems could not, such as n-ary relations within a type hierarchy and general class-level queries (e.g., "Are birds scared of cats?").
Conclusion
By eliminating the need for specialized "collector" and "enabler" nodes, the proposed architecture achieves significant simplicity. It naturally supports finding specific fillers for variables during inference—a task that previously required complex, multi-stage processes. While questions remain regarding the selection of "basis features" and scaling to massive knowledge bases, this research demonstrates that collapsing symbolic distinctions can lead to more robust and flexible connectionist reasoning.
Summary updated