Introduction
In the development of connectionist models during the mid-1980s, lateral inhibition was a cornerstone for achieving selective behavior in neural networks. To resolve competition between hypotheses—such as disambiguating the multiple meanings of a word—standard architectures employed explicit inhibitory links. These negative weights allowed a "winner" node to suppress its rivals. However, this approach suffers from two significant drawbacks: a quadratic explosion in connection complexity as the number of competing nodes increases, and a lack of psychological evidence, as human associative memory rarely exhibits such explicit negative relationships. James Reggia proposes that we can achieve the behavior of inhibition without the baggage of inhibitory links through a competition-based theory of spreading activation.
The Competition-Based Spreading Activation Model
Reggia’s model shifts the paradigm from nodes "pushing" activation to neighbors "pulling" activation from a finite resource pool at the source node. This mechanism is governed by a "strong get richer" rule, where the amount of activation a receiving node $n_j$ obtains from a source $n_i$ is proportional to $n_j$'s current activation level.
Crucially, this competition remains a local computation. A node only requires information about the activation levels and weights of its immediate neighbors to determine the flow of energy. This architecture eliminates the need for explicit negative weights, creating "virtual lateral inhibition" as an emergent property of the network's resource constraints. By removing negative weights, the model becomes much easier to initialize using frequency-based empirical data, which is almost always positive.
Experimental Framework
To validate this theory, a word-sense disambiguation network was implemented using the Parallel Activation Network (PAN) simulator. The network is organized into functional layers representing words (W), senses (S), and context (C). This small, hand-crafted network tests how the model handles ambiguity, such as one word mapping to multiple potential meanings.
Results and Discussion
Simulations demonstrate that the model effectively performs complex logical gating and switching without inhibitory connections. Two primary behaviors were observed:
Automatic Disambiguation
When a word node is activated, it distributes activation to its associated senses. Due to inherent weight differences or initial activation levels, one sense naturally begins to dominate. As it "pulls" more activation, its competitor receives less, eventually being suppressed to near-zero. This is illustrated in the simulation data where one sense (S2) absorbs the majority of activation while the rival (S3) remains near baseline, effectively being "virtually inhibited."
The Switch Effect
The model also exhibits high sensitivity to contextual priming. By introducing a minimal "contextual" input to a theoretically weaker sense, the entire flow of activation can be redirected. This allows the network to settle into a stable state representing the contextually appropriate hypothesis, mimicking a "winner-take-all" circuit. This confirms that the competition-based model can perform complex logical "gating" and "switching" without a single negative connection.
Conclusion
Virtual lateral inhibition provides a more parsimonious and biologically plausible account of selective activation in associative memory. By replacing explicit negative weights with resource competition, the model simplifies network initialization and aligns more closely with empirical findings on human word association. While the current simulations are on a "toy" scale, the author points toward scaling this approach to massive networks, such as those required for full-scale print-to-sound mapping, to further test the robustness of virtual inhibition.
Summary updated