Paper proposes simplex features for backward-compatible retrieval
Original title: Hyperspherical Simplex Representations for Backward-Compatible Retrieval
The method needs no extra compatibility loss, learned map, or architecture change, including when classes are added.
A paper proposes hyperspherical simplex representations from softmax outputs and logits so retrieval embeddings from independently trained models stay comparable without full gallery re-indexing. A deterministic linear transform realigns class prototypes when new classes expand the simplex. The authors report competitive results on standard compatibility benchmarks, including larger class sets and public independently trained models, under stated von Mises–Fisher assumptions.
- Existing methods often need compatibility training, mappings, or the previous model.
- Alignment uses the simplex from softmax cross-entropy with one-hot labels.
- Theory gives two geometric conditions under which compatibility inequalities hold in expectation.
From the original article
1 min read at arxiv.orgTraining modern AI models has become increasingly expensive, and updating deployed models can alter their internal feature representations, affecting applications built upon them. In retrieval systems, such updates often require complete gallery re-indexing, a computationally expensive process at scale.
By Simone Ricci · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.