Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Instance Partial-Label Learning with Margin Adjustment
Read the paperCheck your idea against 108,967 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 MIPLMA, an end-to-end multi-instance partial-label learning method that adds a margin-aware attention mechanism (temperature-annealed, normalized attention) and a margin distribution loss on candidate vs non-candidate predicted probabilities, outperforming MIPL, PLL, and MIL baselines.
Problem
Multi-instance partial-label learning (dual inexact supervision: bag-level instance ambiguity and candidate label ambiguity) suffers from margin violations, where negative instances get higher attention than positives and non-candidate labels get higher probability than candidate labels, hurting gen
Approach
Introduce a margin-aware attention with an annealed temperature and per-bag normalization of attention scores to widen the attention margin between positive and negative instances, plus a margin distribution loss that maximizes the mean and minimizes the variance of the margin between the highest candidate and highest non-candidate predicted probabilities.
Key results
- MIPLMA achieves the best accuracy in 33 of 36 comparisons against the three MIPL baselines on benchmark datasets, e.g., .703±.026 on SIVAL with r=1 versus .675 for ELIMIPL.
- On the CRC dataset, MIPLMA is best in all 11 comparisons, e.g., .557±.010 on C-KMeans versus .546 for ELIMIPL.
- Overall win/tie/loss tally shows MIPLMA statistically superior in 96.4% of cases (477 wins of 495 counted comparisons).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MIPLMA instead of ELIMIPL | MIPL classification · Birdsong, r=3 | accuracy: .717$\pm$.017 -> .746$\pm$.013 | Helped |
| Use MIPLMA instead of ELIMIPL | MIPL classification · FMNIST, r=3 | accuracy: .702$\pm$.055 -> .654$\pm$.055 | Hurt |
| Use margin adjustment in both instance and label spaces rather than either alone or neither (ablation) | MIPLMA · SIVAL, r ∈ {1,2,3} | accuracy | Helped |
| Add margin adjustment in the label space only (ablation) | MIPLMA · SIVAL, r ∈ {1,2,3} | accuracy | Helped |
| Add margin-aware attention (instance-space margin adjustment) (ablation) | MIPLMA · SIVAL, r ∈ {1,2,3} | accuracy | Helped |
| Add margin distribution loss (ablation) | ProDEN · Kuzushiji-MNIST, q ∈ {0.1, 0.3, 0.5, 0.7, 0.9} | accuracy: .772$\pm$.017 -> .816$\pm$.008 at q=0.9 | Helped |
| Use margin-aware attention | MIL attention models · MNIST (MIL) and FMNIST (MIL) | accuracy: .962$\pm$.010 -> .991$\pm$.006 on MNIST; .859$\pm$.032 -> .930$\pm$.021 on FMNIST | Helped |
| Use 25 image patches rather than 16 ResNet-34 feature instances | CRC ResNet-34 MIPL · CRC, C-R34-16 vs C-R34-25 | accuracy: .631$\pm$.008 -> .685$\pm$.011 for MIPLMA | Helped |
Surprising findings
- Adding the margin distribution loss to PRODEN improves PLL accuracy, most strongly under high disambiguation difficulty (q=0.9: .816 vs .772 on Kuzushiji-MNIST).
- Several PLL algorithms with MLP do not consistently outperform their linear-model versions on benchmark datasets, suggesting overfitting.
Limitations
- Like other attention-based MIL/MIPL methods, it cannot process multiple bags simultaneously.
- Shows slight overfitting on the relatively simple MNIST-MIPL dataset.
- Not suitable for instance-level classification tasks.
Experimental setup
Datasets: MNIST-MIPL FMNIST-MIPL Birdsong-MIPL SIVAL-MIPL CRC (colorectal cancer crowdsourced) MNIST (MIL) FMNIST (MIL) Kuzushiji-MNIST
Benchmarks: MNIST-MIPL FMNIST-MIPL Birdsong-MIPL SIVAL-MIPL CRC C-Row CRC C-SBN CRC C-KMeans CRC C-SIFT C-R34-16 C-R34-25 MNIST (MIL) FMNIST (MIL)
Metrics: classification accuracy win/tie/loss counts from pairwise t-tests
Baselines: ELIMIPL DEMIPL MIPLGP PRODEN RC LWS CAVL POP AGGD PLSVM MMMPL VWSGP
Builds on: DEMIPL ELIMIPL attention-based MIL PLL disambiguation margin distribution learning
Closest papers by meaning
- Fast Multi-Instance Partial-Label Learning (AAAI 2025 · similarity 0.92)
- Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning (NeurIPS 2025 · similarity 0.86)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.86)
- Federated Partial Label Learning with Local-Adaptive Augmentation and Regularization (AAAI 2024 · similarity 0.85)
- What Makes Partial-Label Learning Algorithms Effective? (NeurIPS 2024 · similarity 0.85)
- Multi-Instance Multi-Label Classification from Crowdsourced Labels (AAAI 2025 · similarity 0.85)
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.