Model Transplantation: Absorbing Pretrained Neural Knowledge into Binary Hypervectors

Large pretrained neural networks encode rich semantic knowledge in dense embeddingmatrices, but this knowledge is locked in a format incompatible with symbolic reasoning. Wepresent Model Transplantation, a pipeline that converts dense float embeddings into 10,240-bit binary hypervectors (HVs) while preserving similarity structure, enabling the full suite ofVector Symbolic Architecture (VSA) operations—binding, bundling, and analogical query—over transplanted knowledge. The core projection step applies a seeded random Gaussianmatrix P ∈ Rd×D followed by sign binarisation, an instance of the Johnson–Lindenstrauss(JL) random projection family.We implement and benchmark the pipeline in the NSCK (Neuro-Symbolic CognitiveKernel) codebase and evaluate on four embedding sources: (i) Gaussian synthetic embeddings(d = 32, n ≤ 1,000); (ii) bert-base-uncased contextual sentence embeddings (d = 768,n = 4,096); (iii) all-MiniLM-L6-v2 curated sentence embeddings (d = 384, n = 280); and(iv) all-MiniLM-L6-v2 scaled phrase embeddings (d = 384, n = 5,000). Across all sources weobserve Spearman rank correlation ϱ > 0.98 between source cosine similarities and projectedHV Hamming similarities. Contextual BERT achieves ϱ = 0.9906, Recall@10 = 0.889, andARI = 0.723; MiniLM-curated achieves Recall@10 = 0.898; and the scaled MiniLM vocabulary(n = 5,000) retains ϱ = 0.9839 with Recall@10 = 0.954. Round-trip encoding fidelity is 1.000(deterministic). We further demonstrate VSA analogy on a balanced 50-equation benchmark(5 categories × 10 analogies): king ⊕ man ⊕ woman ≈ queen. Aggregate accuracy is Top-1= 39/50 = 78% and Top-5 = 48/50 = 96%. Finally, on a downstream synonym-substitutedrelation-retrieval task, transplanted HVs achieve Top-1 accuracy 0.650 versus a random-HVbaseline of 0.250 (absolute gain +0.400), demonstrating functional reasoning gains beyondgeometric preservation. The JL analysis confirms D = 10,240 exceeds the lower bound by≥1.40× for the tested vocabulary sizes.Keywords: Vector Symbolic Architecture, model transplantation, random projection,Johnson–Lindenstrauss lemma, hypervector computing, neuro-symbolic integration, em-bedding compression