Multi-Subspace FAISS with Entropy Gating and Whitening A Mode-Aware Exemplar Search Approach Benjamin J Gilbert College of the Mainland Robotic Process AutomationDownload
We present a mode-aware exemplar index that
learns K subspaces via Gaussian mixture models, K-means, or
Bayesian Gaussian mixture models, and routes queries using
soft cluster responsibilities with entropy gating and optional
per-subspace whitening. The approach demonstrates improved
retrieval accuracy through adaptive subspace selection while
maintaining computational efficiency. Goal-aware sparsity masks
are applied before scaling and whitening for both indexing and
search operations, enabling fine-grained feature control.
Index Terms—Vector search, FAISS, subspace learning, mixture models, entropy gating, whitening, exemplar retrieval
