Research map / Label-efficient and robust learning / Recognition and long-tail
Fast Multi-Instance Partial-Label Learning
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 FASTMIPL, a mixed-effects (GLMM) linear model with variational inference and posterior predictive disambiguation for multi-instance partial-label learning that models bag dependencies and is much faster than prior MIPL methods.
Problem
Multi-instance partial-label learning (MIPL) methods assume instances are independent, ignoring bag-level dependencies and heterogeneity, and are very slow on complex datasets.
Approach
Model bags with a generalized linear mixed model (fixed effects from bag covariates, random effects from attention-pooled instance embeddings with Gaussian priors), fit via reparameterized variational inference with mini-batch SGD, and disambiguate candidate labels using the posterior predictive probability.
Key results
- On MNIST-MIPL with r=1, FASTMIPL reaches 0.999±0.002 accuracy versus 0.991 for ELIMIPL (best baseline).
- On CRC-MIPL (C-Row) FASTMIPL achieves 0.487±0.038 accuracy versus 0.434 for ELIMIPL, and it wins 9 of 11 MIPL comparisons on real-world data.
- FASTMIPL reduces training time by up to nearly 20x versus ELIMIPL and up to about 30x versus DEMIPL on CRC-MIPL while maintaining accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FASTMIPL instead of ELIMIPL | MIPL classification · MNIST-MIPL, FMNIST-MIPL, Birdsong-MIPL, and SIVAL-MIPL (r=1, 2, 3) | accuracy: MNIST: .999±.002 vs .991±.005 (r=1); .998±.004 vs .989±.013 (r=2); .975±.074 vs .749±.148 (r=3). FMNIST: .911±.022 vs .9 | Helped |
| Use FASTMIPL instead of the strongest reported MIPL baseline | MIPL classification · CRC-MIPL (C-Row, C-SBN, C-KMeans, C-SIFT) | accuracy: C-Row: .487±.038 vs .435±.006; C-SBN: .573±.031 vs .510±.008; C-KMeans: .573±.013 vs .545±.013; C-SIFT: .526±.029 vs .53 | Mixed |
| Add random-effect coefficient β (ablation) | FASTMIPL · Benchmark and real-world MIPL datasets | accuracy | Helped |
| Add instance-weighting parameter γ for embedding and instance-level aggregation (ablation) | FASTMIPL · Benchmark and real-world MIPL datasets | accuracy | Helped |
| Use FASTMIPL instead of ELIMIPL or DEMIPL | MIPL classification · CRC-MIPL | training time: up to nearly 20 times compared to ELIMIPL; up to about 30 times compared to DEMIPL | Helped |
Surprising findings
- MIPLGP did not improve with more sophisticated image bag generators on CRC-MIPL, which the authors attribute to its inability to use such features.
Limitations
- Some baselines (MIPLGP) could not be run on the single-GPU setup, so efficiency comparison excludes them.
- FASTMIPL-V1 and V2 ablations show the model depends on both random effects and learned embeddings; the framework is a linear model, limiting expressiveness.
- The C-SIFT dataset could not be evaluated for MIPLGP due to memory overflow.
Experimental setup
Datasets: MNIST-MIPL FMNIST-MIPL Birdsong-MIPL SIVAL-MIPL CRC-MIPL
Benchmarks: MNIST-MIPL FMNIST-MIPL Birdsong-MIPL SIVAL-MIPL CRC-MIPL
Metrics: classification accuracy training time
Baselines: MIPLGP DEMIPL ELIMIPL PRODEN RC LWS CAVL PL-AGGD MIVAE ATTEN ATTEN-GATE LOSS-ATTEN
Builds on: GLMM attention-based MIL ELBO variational inference
Closest papers by meaning
- Multi-Instance Partial-Label Learning with Margin Adjustment (NeurIPS 2024 · similarity 0.92)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.86)
- Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning (NeurIPS 2025 · similarity 0.86)
- Multi-Instance Multi-Label Classification from Crowdsourced Labels (AAAI 2025 · similarity 0.85)
- Federated Partial Label Learning with Local-Adaptive Augmentation and Regularization (AAAI 2024 · similarity 0.84)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · 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.