TurboQuant/TurboVec Retrieval Notes#
TurboQuant/TurboVec is optional retrieval, compression, and audit infrastructure around NFC. It is not the core NeuroTwin model contribution and must not become a required dependency.
Math Summary#
Normalize:
Random rotation:
High-dimensional coordinate approximation:
Quantization map:
MSE objective:
Inner-product distortion objective:
Score error bound:
Retrieval-kNN Baseline#
This baseline may use train labels and test query features. It must never use test targets.
Semantic Near-Duplicate Audit#
This can flag train/test semantic leakage in stimulus features, especially for Algonauts/CNeuroMod.
Why It Could Help#
Compress large stimulus-feature stores.
Support exact or approximate nearest-neighbor retrieval baselines.
Audit semantic near-duplicates between train and test stimuli.
Store latent field summaries for memory/debug workflows.
Risks#
Low-dimensional vectors can violate high-dimensional approximations.
Approximate retrieval can miss duplicates.
Quantization distortion can change rankings.
Optional dependency friction can break reproducibility.
Quantization is not automatically differentiable or claim-relevant.
Implementation Priority#
Numpy exact vector store.
Optional lazy TurboVec adapter.
Retrieval baseline.
Semantic duplicate audit.
No implementation is added in this pass.