Kernel Affine Hull Machines Reduce Query Encoding Costs
Kernel Affine Hull Machines (KAHMs) introduce a cost-effective analytical estimator to substitute costly neural inference in semantic retrieval. This innovative technique transforms lexical characteristics into a fixed semantic embedding space by calculating prototype-mixture weights within a precisely defined RKHS, while also enhancing prototypes through normalized least-mean-squares. In an evaluation based on Austrian law involving 5,000 queries, 84 laws, and 10,762 units, KAHM demonstrated impressive teacher-space reconstruction, achieving a mean squared error of 0.000091 and an R² value of 0.90.
Key facts
- KAHMs replace repeated neural inference with an analytical estimator.
- The method uses prototype-mixture weights in an RKHS.
- Prototypes are refined via normalized least-mean-squares.
- Tested on an Austrian-law benchmark with 5,000 queries.
- Achieved MSE 0.000091 and R² 0.90.
- The study addresses the fixed-teacher query-adaptation problem.
- Encoding error is decomposed into posterior-approximation, generalization, and teacher-noise components.
- The approach is designed for compute-efficient query-side semantic encoding.
Entities
—