Integrating a Compositional Spatiotemporal Deep Learning Network with Symbolic Representation/Reasoning

Author: Ben Goertzel
Year: 2011
Venue: Proceedings of AAAI Symposium on Cognitive Systems
Links: PDF

Summary

Proposes integrating deep learning spatiotemporal networks with symbolic reasoning through an intermediary semantic network layer.

Relevance to Hyperon

Informs the neural-symbolic integration strategy in Hyperon, particularly how to bridge subsymbolic perception with symbolic metagraph representations.

---

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

Curated notes and excerpts from local source extraction at publication_texts/json/Compositional_Spatiotemporal_Deep_Learning.json.

Identity Note: 2011 CSDLN Paper, NOT 2013 FISHGRAM

This paper is the 2011 "A Novel Strategy for Hybridizing Subsymbolic and Symbolic Learning and Representation" (Goertzel, AAAI Cognitive Systems Symposium 2011). It is not the 2013 paper "Integrating Deep Learning Based Perception with Probabilistic Logic via Frequent Pattern Mining" (Goertzel et al., 2013), often referenced as the FISHGRAM paper. Those are separate works. The 2013 FISHGRAM paper is currently a wiki gap — see Publications+Deep Learning Perception with PLN stub which still needs full-text retrieval from a separate source.

CSDLN Framework

The paper introduces Compositional Spatiotemporal Deep Learning Networks (CSDLNs) as a category covering HTM, DeSTIN, and related systems whose hierarchies mirror spacetime structure. The integration claim: meaningful symbolic/subsymbolic integration requires dynamic adaptive linkage between CSDLN attractors and analogous symbolic representations — not just translation at boundaries.

Tripartite Semantic CSDLN

The proposed bridge structure is a tripartite semantic CSDLN: semantic-perceptual, semantic-motoric, and semantic-goal hierarchies bridging CSDLN attractors to OpenCog AtomSpace. Semantic CSDLN nodes are NOT raw vectors — they contain abstract patterns and local semantic networks, optionally with spatial/temporal relationship labels.

Feedback Loop

Learning includes a feedback loop: mine frequent patterns from a standard perceptual CSDLN → create semantic CSDLN subnetworks/parent nodes → update semantic clusters via CSDLN dynamics → use semantic clusters as seeds for more mining.

Author's Status Note

The paper explicitly says most ideas were not implemented as of July 2011. It is a conceptual integration strategy, not a deliverable system.

Implementation Status

Per the cluster-pilot review, this paper is paper-only. No CSDLN/DeSTIN/OpenCog bridge in any cluster repo verified through this pass. Agrees with the broader neural-symbolic stream represented by Publications+OpenCog NS Hybrid Neural-Symbolic at the high level (attractors should link to symbols through intermediate structures, not as raw vectors), but is orthogonal to the post-2024 World-Model PLN line — those papers concern PLN's evidence semantics, not perceptual grounding.