Learning a nonlinear feature extractor with K-means via classifier-constrained posteriors
caglar_ee · x · 2026-11-11
The author sketches how to learn a neural-network-like nonlinear feature extractor using K-means: introduce hidden continuous feature vectors yₙ for inputs xₙ and couple the posteriors q(yₙ|xₙ) with linear classifier margin constraints, wᵀ Eq(yₙ|xₙ)[yₙ]+b ≥ 1 and ≤ −1, jointly optimizing clustering and classification.
More from Research
- Sable proves bit-exact LLM inference: Qwen2.5-32B verified on one GPU in 92.5s — MoonMath-ai · 2026-11-11
- AACL 2026 tutorial maps diffusion language models, slides now public — Bollegala · 2026-11-11
- K-means-based output classifier still trainable via alternating minimization, with a KL-constrained E-step — caglar_ee · 2026-11-11
- Molence Bio open-sources Mol-1-constraint for constraint-guided biomolecular prediction — AllThingsApx · 2026-11-11
- Compass 1 navigation foundation model hits 79.4% success in 320 real-world trials, zero-shot transfers to 4 robot types — abursuc · 2026-11-11
- Hollow-core fibre cuts data link lag 30% over 1,000km in Chinese trial — AIFlow_ML · 2026-11-11