essay · 2026-10-11

How do I calculate intra-class variance using LLMs

Research on person re-identification reveals that standard batch-hard triplet losses often suffer from significant variance between different positive samples, a problem known as the Distance Variance among Different Positives (DVDP). This variance creates redundancy in learned features, which can hinder the model's ability to generalize effectively across different contexts or domains. A proposed solution is the Attribute-aware Identity-hard Triplet Loss (AITL), a method specifically designed to reduce this intra-class variation by incorporating attribute distance calculations into the loss function. The underlying theory suggests that controlling intra-class variance is critical for managing generalization bounds and eliminating redundant information in the feature space. By minimizing DVDP, models can achieve more stable representations where features of the same identity cluster tightly together regardless of other attributes present. This matters because systems relying on such feature embeddings need high consistency to function correctly when processing varied inputs from different sources or environments.

Sources

• Support Neighbor Loss for Person Re-Identification — 2018-08-18T01:40:56Z
• An Enhanced Federated Prototype Learning Method under Domain Shift — 2024-09-27T09:28:27Z
• Attribute-aware Identity-hard Triplet Loss for Video-based Person Re-identification — 2020-06-13T09:15:38Z
• [Paper Note] Explaining Grokking and Information Bottleneck through...
• Too Categorical to be Human: Emotion Concepts in LLMs and Humans
• Proceedings of the 2023 Conference on Empirical... - ACL Anthology
• How Text Quality Interventions Reshape Neural Scaling... | OpenReview
• How Andrej Karpathy uses LLMs: complete example-driven, practical...