Research map / Label-efficient and robust learning / Recognition and long-tail

Representation Rigidity in Face Embeddings: Orthogonal Identifiability and Backfill-Free Compatibility

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Margin-based face recognition losses

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that independently trained angular-margin face embedding models are related by a single global orthogonal transform (representation rigidity), enabling backfill-free compatibility upgrades for face recognition.

Problem

Upgrading a deployed face-recognition model makes new query embeddings incompatible with the legacy gallery, requiring costly backfilling of stored embeddings.

Approach

Independently trained angular-margin face models have embedding spaces well approximated by one global orthogonal transformation, so queries from a new model can be mapped into the legacy gallery space without re-extracting gallery embeddings.

Surprising findings

• Independently trained angular-margin face models are well approximated by a single global orthogonal transformation of their embedding spaces.

Closest papers by meaning

• Attributes Shape the Embedding Space of Face Recognition Models (ICML 2025 · similarity 0.87)
• Towards Learning Structure via Consensus for Face Segmentation and Parsing (CVPR 2020 · similarity 0.86)
• Generalizable Face Landmarking Guided by Conditional Face Warping (CVPR 2024 · similarity 0.85)
• HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.84)
• LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · similarity 0.84)
• Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.84)