One-vs-one N-pair loss
Mathematical Methods AIMech:OneVsOneNPairLoss DRAFT
Aliases: N-pair-ovo loss
A supervised metric-learning objective that sums separate positive-versus-negative softplus comparisons for each anchor in a batch of positive pairs from distinct classes.
Identity Notes & Concept Scope
This is the named N-pair-ovo objective in Sohn’s Equation (8). It shares the N-pair batch construction with N-pair-mc but uses a sum of binary softplus comparisons, not one joint log-sum-exp over negatives. They coincide for one negative per anchor (N=2), but generally differ for N>2. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 218 preceding native IDs/names/aliases. Word-delimited fixed-string searches of both full checkouts, including ignored and hidden files, returned 15 lines each in existing discovery/review context. Five neighboring native records and five exact AIO nodes with scoped synonyms and outgoing edges were inspected.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Given N>=2 positive pairs (x_i,x_i+) from N distinct classes and embeddings f_i,f_i+, define L_ovo=(1/N) sum_i sum_(j!=i) log[1+exp(f_i^T f_j+−f_i^T f_i+)]. Other positive-side examples supply N−1 different-class negatives to each anchor. The printed divisor is N, not N(N−1), so it averages across anchors while summing each anchor’s pairwise comparisons. In contrast, N-pair-mc places the sum over negatives inside a single logarithm. The expression uses dot products without an explicit temperature or cosine normalization.
⚠️ Curated Limitations & Negative Results
- Use N positive pairs from N distinct classes; other positive-side examples must represent different classes for the stated objective. Pair sampling and class selection are part of this configuration, not interchangeable with arbitrary augmented-view negatives.
- The main paper uses embedding-norm regularization instead of a hard unit-length constraint. This record specifies the loss term; it does not prescribe a regularization coefficient, encoder, optimizer or full training schedule.
- No unconditional equivalence to a margin triplet hinge or an information lower bound is asserted. Objective shape and label-based batching do not establish downstream accuracy. No training or benchmark was reproduced.
- The proceedings Supplemental.zip inspected during retrieval contains a different paper by Steinhardt and Liang. It is not supporting evidence; this record relies on the inspected nine-page N-pair main paper.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ContrastiveLearning
Contrastive Learning
|
related_match | The objective contrasts same-class positives with different-class negatives. The inspected AIO target explicitly describes a self-supervised deep-network approach, whereas this objective is supervised and mathematical. Their shared positive-versus-negative representation-learning mechanism supports a related association only, not an equivalence or subclass assertion. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Kihyuk Sohn
|
primary paper
full_text
|
ID: https://papers.nips.cc/paper_files/paper/2016/hash/6b180037abbebea991d8b1232f8a8ca9-Abstract.html
sha256:83bee5d1a7ed…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T09:32:03Z
Updated: 2026-10-11T09:32:03Z
Source File:
data/curated/OneVsOneNPairLoss.yaml