Pattern Mining

help edit space_dashboard
Draft — This content has not been approved for publication.

Responsible: Ben Goertzel

Papers: Hyperon for AGI⇒ASI Whitepaper (2025), §6.4; O'Neill & Goertzel (2012), FISHGRAM: Pattern Mining for General Intelligence; Goertzel, Sanders, O'Neill, Yu (2013), DeSTIN-FISHGRAM-PLN

Status: Multi-paradigm. Three paradigms exist as parallel research tracks: symbolic (opencog/miner + hyperon-miner trio), neural / GNN (rejuve-bio/neural-subgraph-matcher-miner), and perception / embodiment (Vepstas's Hyperon-era AtomSpace-Scheme portfolio). The three paradigms have zero cross-grep references in current Hyperon code per the Perception/Neural-Symbolic cluster pilot (closed 2026-05-01). MORK-native pattern mining, TransWeave integration, and the PLN/WILLIAM/Symbolic-Heads/SubRep downstream paths described below are whitepaper / roadmap claims, not currently-wired implementation.

Concept

Pattern mining discovers frequent and surprising subgraph structures in the AtomSpace metagraph — templates that recur across contexts. The whitepaper frames mined patterns as reusable building blocks for reasoning (PLN proof heuristics), learning (MOSES program templates), neural integration (Symbolic Heads for transformers), and knowledge compression (WILLIAM). Roadmap framing — see "What This Card Is Not" below.

Algorithm (symbolic paradigm)

The classical (opencog/miner + hyperon-miner) approach is incremental and stream-based: generate abstract templates, specialize them, filter by support, expand conjunctions, and rank by I-surprisingness — an information-theoretic metric measuring how much a pattern's frequency deviates from what independence assumptions would predict. Results are maintained as nondeterministic streams until final tally, making it natural to run continuously as knowledge drifts. Per the Perception cluster pilot Source 2, opencog/miner is paper-faithful to the Chi/Xia/Yang/Muntz 2005 pattern-mining methodology family (where the 2013 FISHGRAM was paper-leaning-divergent).

The neural / GNN paradigm (rejuve-bio/neural-subgraph-matcher-miner) follows a separate [SPMiner-LINEAGE] tradition with strictly empirical patterns. The two paradigms are [PARALLEL-NON-INTEGRATED]; see Neural Pattern Mining for the implementation surface.

Trilateral Tradition Map (Perception cluster pilot, 2026-05-01)

ParadigmAnchor reposVerdict tagsWiki home
Symbolic mining opencog/miner (legacy C++/Scheme baseline) + hyperon-miner trio (trueagi-io / iCog-Labs-Dev / zariuq) [METTA-PORT-OF-MINER] for the hyperon-miner trio; paper-faithful to Chi/Xia/Yang/Muntz 2005 family on the legacy side This card
Neural / GNN mining rejuve-bio/neural-subgraph-matcher-miner [SPMiner-LINEAGE] / [GNN-NEURAL-MINING] / [PARADIGM-DISTINCT-NEURAL] / [STRICTLY-EMPIRICAL] / [PARALLEL-NON-INTEGRATED] Implementation Families+Neural Pattern Mining
Perception / embodiment Vepstas's Hyperon-era AtomSpace-Scheme portfolio: opencog/sensory, agents, motor, evidence, atomese-simd [ATOMSPACE-SCHEME-STACK]; portfolio not wired to miners Knowledge Representations+Sensory

What This Card Is Not (disambiguation)

  • NOT FISHGRAM continuation by name. The 2013 paper title's "FISHGRAM" is a wiki/editorial term-of-art for the algorithm in O'Neill & Goertzel 2012; current Hyperon code does not use the FISHGRAM name. See Publications+FISHGRAM Pattern Mining for the editorial-terminology disambiguation.
  • NOT current DeSTIN-FISHGRAM-PLN wiring. The 2013 Goertzel/Sanders/O'Neill/Yu paper described a coupling of DeSTIN (perception), FISHGRAM (pattern mining), and PLN (reasoning) as a fragmented architectural experiment. [PARTIAL-FRAGMENTED-REVIVAL] per Perception cluster pilot Source 1 (2026-05-01): the pieces exist as parallel tracks today, but the 2013 coupling is NOT wired in current Hyperon code; do not read the whitepaper roadmap as evidence of operational coupling.
  • NOT ECAN attention. Pattern mining surfaces structural patterns; ECAN allocates economic attention. Adjacent but distinct concerns; see ECAN.
  • NOT a single integrated Hyperon runtime path. The three paradigms (symbolic / neural / perception) are research tracks with no current cross-grep references in code.
  • NOT pattern indexing. See "Pattern Index — adjacent legacy" below.

Pattern Index — adjacent legacy (Source 3 close, 2026-05-06)

singnet/pattern-index (forked from opencog/pattern-index; HEAD a72459e, 2020-07-24; 1 ahead / 3 behind upstream) is a separate paradigm: disk-based AtomSpace pattern indexing for huge SCM-file lookup, NOT a wrapper around OpenCog's PatternMatcher or PatternMiner. It does ship its own pattern-mining algorithms (PatternIndexAPI.h:182-208 minePatterns; TypeFrameIndex.cc:581-745 mining routine; tests at tests/pattern-index/PatternIndexAPIUTest.cxxtest:170-179) following the OpenCog wiki Pattern_miner methodology, but the architectural distinction is RAM-only AtomSpace mining (the symbolic paradigm above) vs disk-based subgraph indexing (this repo). Verdict: [PATTERN-INDEX-SEPARATE-PARADIGM-ADJACENT] — adjacent legacy infrastructure, not a fourth Hyperon-era tradition.

Roadmap claims (qualify before citing as implementation)

The following are whitepaper / roadmap descriptions, not currently-wired implementation:

  • MORK-native pattern mining: the whitepaper describes patterns as subtrees with PathMap traversal matches, capsule-summary trie-node counts, and weakness-guided search. Implementation pending; depends on the AtomSpace Integration Phase 4+/5+ work.
  • TransWeave framing: mining as two-ended geodesic discovery between data-driven priors and goal-driven posteriors. Conceptual framing only.
  • PLN/WILLIAM/Symbolic-Heads/SubRep downstream paths: mined patterns feeding inference templates, compression-worthy templates, structured transformer-attention priors, or candidate subgoals — paper-described, not currently exercised by Hyperon code paths.

Implementations

For full implementation detail, defer to the cluster-specific cards:

  • Symbolic (this card's primary scope): trueagi-io/hyperon-miner — MeTTa/SWI-Prolog implementation with three approaches: pattern matching, dependent type system, Prolog-based. iCog-Labs-Dev fork is the active HEAD with team-overlap (Basliel/Amsalu/Tezena); zariuq/hyperon-miner is Zar Goertzel's own fork. Active.
  • opencog/miner — legacy C++/Guile Scheme implementation using OpenCog's Unified Rule Engine (Miner, HandleTree, Valuations, Surprisingness; requires cogutil + atomspace + unify + URE). Maintained baseline; ngeiswei/miner is the upstream maintainer's pre-merge staging fork (currently 0 ahead / 8 behind opencog/master at 2026-05-06). leungmanhin/miner and singnet/miner are [STRICT-FORK-STALE] historical forks.
  • Neural / GNN: see Implementation Families+Neural Pattern Mining — rejuve-bio/neural-subgraph-matcher-miner.
  • Perception / embodiment portfolio: see Knowledge Representations+Sensory for Vepstas's 5-repo Hyperon-era AtomSpace-Scheme stack. Not wired to either miner family at HEAD.

Key References

  • Goertzel, B. (2025). Hyperon for AGI⇒ASI Whitepaper, §6.4: Pattern Mining on General Graphs
  • O'Neill, J. and Goertzel, B. (2012). Pattern Mining for General Intelligence: The FISHGRAM Algorithm, AGI-12. See Publications+FISHGRAM Pattern Mining.
  • Goertzel, B., Sanders, T., O'Neill, J., Yu, G. (2013). Integrating Deep Learning Based Perception with Probabilistic Logic via Frequent Pattern Mining, AGI-13. Local: publication_texts/2013_Goertzel_Sanders_ONeill_Yu_DeSTIN_PLN.pdf. [PARTIAL-FRAGMENTED-REVIVAL].
  • Chi, Y., Xia, Y., Yang, Y., Muntz, R. (2005). Methodology family that opencog/miner is paper-faithful to.

Cluster-pilot extraction archives:

  • Perception/Neural-Symbolic cluster pilot (closed 2026-05-01) — primary source for trilateral framing: scripts/archive/perception_pilot/ (sources 1-5 + reconciliations).
  • Non-clustered HAA cluster pilot Source 3 (closed 2026-05-06) — residual scope; verified miner-fork stalenesses, locked pattern-index adjacency: scripts/archive/non_clustered_haa_pilot/source3_pattern_mining_residual/.



Discussion