OpenCog NS: A Deeply-Connected, Hybrid Neural-Symbolic Architecture

Author: Ben Goertzel
Year: 2010
Venue: Proceedings of BICA-2010
Links: PDF

Summary

Proposes a deeply-connected hybrid neural-symbolic architecture where neural and symbolic processing are tightly integrated rather than loosely coupled.

Relevance to Hyperon

Informs Hyperon's neural-symbolic integration strategy, particularly how MeTTa can interface with neural network components.

---

Curated Excerpts (cluster-pilot extraction, 2026-04-25)

Curated notes and excerpts from publication_texts/json/OpenCog_NS_Hybrid_Neural-Symbolic.json. See RawData+Publications+OpenCog NS Hybrid Neural-Symbolic for the complete raw extraction.

Architectural Definition

The paper defines a deeply-interactive hybrid neural-symbolic architecture: neural and symbolic components frequently intervene in each other's internal operations, forming one coupled dynamical system. This is the contrast with loose-coupling neural-symbolic systems where translation happens only at boundaries.

Hierarchical Attractor Neural Net (HANN) Decomposition

OCNS combines OpenCog Prime with HANNs, not generic neural embeddings. Proposed HANN layout: primary, high-level procedural, low-level procedural, and episodic HANNs, each linked to OpenCog memory and process types. Architectural objective: global/local memory synergy — each memory type has globalist and localist substores that interact synergetically.

The Grounding Seam

HANN attractors connect to AtomSpace via fuzzy MemberLink weights between attractors and ConceptNodes. MemberLink weight represents fuzzy attractor membership.

NN Activation ↔ STI Currency Conversion

Activation may spread from NN nodes to ConceptNodes; STI may spread back. A conversion rate between NN activation and STI currency must be maintained by the "OCP central bank." This couples ECAN's resource-allocation layer (see Publications+Economic Attention Networks) directly into the neural-symbolic interface.

Author's Caveat

The paper explicitly states OCNS has not yet been implemented as of writing — properties are speculative even as an architecture specification.

Historical Context

This is a 2009/2010 architectural proposal predating MeTTa entirely. The vocabulary is OpenCog/AtomSpace/HANN. The Relevance summary above ("how MeTTa can interface with neural network components") should be read as analogy: the paper specifies AtomSpace ↔ NN coupling, not MeTTa ↔ NN coupling. Modern Hyperon work would need to translate the HANN/MemberLink coupling pattern into MeTTa terms.

Implementation Status

Per the cluster-pilot review, this paper is paper-only / architectural proposal, never implemented in any cluster repo verified through this pass. Orthogonal to the post-2024 World-Model PLN line documented in RawData+Publications+xiPLN and RawData+Publications+World-Model Calculus: those address PLN's evidence semantics; this paper addresses neural-symbolic perception grounding. Different problem, different solution paradigm.