Research map / Label-efficient and robust learning / Recognition and long-tail
Partial Label Causal Representation Learning for Instance-Dependent Supervision and Domain Generalization
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 CausalPLL+, a VAE-based instance-dependent partial label learning method that learns identifiable content-style disentangled causal representations, with contrastive learning and candidate label refinement.
Problem
Instance-dependent partial label learning (IDPLL) where candidate labels depend on instance features, and existing PLL methods ignore this dependency for causal representation learning and are vulnerable to style variation and domain shift.
Approach
Model instance and candidate label generation with a conditional VAE whose content latent has a class-conditioned prior (via a prior network over a normalized candidate-label vector), split from a standard-normal style latent; add a prior-based contrastive loss and momentum-based candidate label refinement.
Key results
- On KMNIST at tau=16, CausalPLL+ reaches 98.49% accuracy vs 91.77% for the best baseline CC.
- On CIFAR10 at tau=16, PLCR (96.28%) slightly exceeds CausalPLL+ (95.91%), so CausalPLL+ is not uniformly best.
- On USPS→SVHN at tau=16, CausalPLL+ achieves 86.72% vs 80.23% for CC and 83.70% for PLCR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CausalPLL+ instead of the strongest reported baseline | IDPLL classification · FashionMNIST, τ = 16/32/64 | accuracy: 93.28 ± 0.24% -> 94.49 ± 0.37%; 92.46 ± 0.13% -> 93.60 ± 0.10%; 90.72 ± 0.15% -> 92.75 ± 0.18% | Helped |
| Use CausalPLL+ instead of the strongest reported baseline | IDPLL classification · KMNIST, τ = 16/32/64 | accuracy: 97.84 ± 0.04% -> 98.49 ± 0.08%; 96.03 ± 0.60% -> 97.89 ± 0.14%; 91.43 ± 0.58% -> 96.96 ± 0.10% | Helped |
| Use CausalPLL+ instead of the strongest reported baseline | IDPLL classification · SVHN, τ = 16/32/64 | accuracy: 97.15 ± 0.09% -> 97.50 ± 0.21%; 96.59 ± 0.15% -> 97.05 ± 0.37%; 95.97 ± 0.18% -> 96.56 ± 0.26% | Helped |
| Use CausalPLL+ instead of the strongest reported baseline | IDPLL classification · CIFAR10, τ = 16/32/64 | accuracy: 96.28 ± 0.09% -> 95.91 ± 0.28%; 93.97 ± 0.07% -> 94.04 ± 0.26%; 88.82 ± 0.11% -> 89.66 ± 0.32% | Mixed |
| Use CausalPLL+ instead of the strongest reported baseline | Domain generalization · MNIST→MNIST-M, τ = 16/32/64 | accuracy: 98.64 ± 0.07% -> 97.85 ± 0.11%; 95.34 ± 0.09% -> 96.58 ± 0.10%; 94.59 ± 0.11% -> 94.67 ± 0.18% | Mixed |
| Use CausalPLL+ instead of the strongest reported baseline | Domain generalization · MNIST→SVHN, τ = 16/32/64 | accuracy: 95.57 ± 0.08% -> 94.13 ± 0.15%; 92.50 ± 0.09% -> 93.11 ± 0.12%; 87.39 ± 0.18% -> 91.04 ± 0.19% | Mixed |
| Use CausalPLL+ instead of PLCR | Domain generalization · USPS→SVHN, τ = 16 | accuracy: 83.70 ± 0.17% -> 86.72 ± 0.16% | Helped |
Surprising findings
- On MNIST→MNIST-M at tau=16, PiCO (98.64%) outperforms CausalPLL+ (97.85%), though PiCO collapses at higher ambiguity levels.
Limitations
- Candidate label refinement and the prior network rely on a learned, partly heuristic procedure; no theoretical guarantee for the refinement step is given.
- Experiments are limited to small image benchmarks and simple domain shifts.
Experimental setup
Datasets: FashionMNIST Kuzushiji-MNIST SVHN CIFAR10 MNIST MNIST-M USPS
Benchmarks: FashionMNIST Kuzushiji-MNIST SVHN CIFAR10 MNIST→MNIST-M MNIST→SVHN USPS→SVHN
Metrics: classification accuracy
Baselines: PLCR VALEN LWS PRODEN RC CC PiCO
Builds on: iVAE Khemakhem et al. 2020 identifiable VAE PRODEN RC CC
Closest papers by meaning
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.90)
- Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning (NeurIPS 2025 · similarity 0.89)
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.88)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.88)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.87)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · similarity 0.87)
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.