Research map / Label-efficient and robust learning / Recognition and long-tail
Boundary Mass in Feature Space as a Label-Budget Diagnostic for Representations
Siming Zhang, Zhehui Shen, Shijie Chen, Yansen Yu, Hang Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a boundary-mass diagnostic in feature space that estimates how much target data lies near the fitted decision boundary, as a label-budget indicator for representations.
Problem
Predicting how many labels are needed to learn a downstream decision boundary in low-label transfer; existing transferability and probe-quality scores (global alignment, likelihood, margins, calibration) do not measure target mass near the boundary where labels are most costly.
Approach
Introduces an effective boundary-mass score measuring how much target data remains close to the fitted decision boundary in feature space, used as a diagnostic of label budget rather than global alignment or calibration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| expand features along the boundary normal (s=1 to s=2) (ablation) | frozen CIFAR-100 features with linear probe · CIFAR-100 label budget | budget (spc): -1.54 spc; AULC +2.26 | Improved |
| scale features along a tangent direction (s=1 to s=2) (ablation) | frozen CIFAR-100 features · CIFAR-100 label budget | budget (spc): +0.02 spc; AULC -0.22 | No clear effect |
| add smooth boundary-mass penalty to linear head training | linear head on frozen features · CIFAR-100 / Tiny-ImageNet | boundary complexity and top-1: psi 201.68 -> 169.22; top-1 53.97 -> 53.84 | Improved |
| add top-2 boundary-pressure regularizer to ResNet-18 training | ResNet-18 · CIFAR-100 | Top-1: 53.68 -> 57.09 | Improved |
| use weight decay, margin or entropy regularization instead of boundary pressure (ablation) | ResNet-18 · CIFAR-100 | boundary estimate: 23.57 -> 23.61 (weight decay) | No clear effect |
| add boundary mass to conventional and transferability diagnostics (ablation) | label-budget regression · CIFAR-100 binary pairs | R^2: 0.734 -> 0.749 | Improved |
| second-stage tie-break by boundary mass among two shortlisted backbones | backbone selection · CIFAR-100 binary pairs | regret (spc): 0.058 -> 0.030 | Improved |
| normalize sample size by psi_eff in sample-complexity fit (ablation) | plug-in boundary study · controlled synthetic | R^2: 0.698 -> 0.895 | Improved |
| use uncertainty/hybrid acquisition instead of random (ablation) | label acquisition · CIFAR-100 binary | mean realized budget (spc): 7.05 -> 5.57 / 5.47 | Improved |
Closest papers by meaning
- Devil Is in the Edges: Learning Semantic Boundaries From Noisy Annotations (CVPR 2019 · similarity 0.87)
- MaxSup: Overcoming Representation Collapse in Label Smoothing (NeurIPS 2025 · similarity 0.85)
- Boundary Matters: A Bi-Level Active Finetuning Method (NeurIPS 2024 · similarity 0.85)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.84)
- Discovering Fine-Grained Visual-Concept Relations by Disentangled Optimal Transport Concept Bottleneck Models (CVPR 2025 · similarity 0.84)
- Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
