Research map / Probabilistic and scientific modeling / Causal discovery
Learning the Latent Causal Structure for Modeling Label Noise
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 CSGN, a VAE-based generative label-noise method that learns a latent causal graph (linear SCM over latent causal variables with masks) to estimate instance-dependent noise transition matrices without similarity assumptions.
Problem
Estimating instance-dependent noise transition matrices from only noisy data requires relating transition matrices of a few identifiable "special" instances to others, which existing methods do via similarity assumptions that are hard to verify and may fail.
Approach
Model the noisy-data generating process with latent causal variables Z that are causally related via a learnable linear DAG, with Y-dependent weights and two learnable masks selecting which latent variables generate the instance versus the noisy label; learn this with a VAE/ELBO using clean examples selected by warmup, so transition similarities emerge implicitly from shared latent causes.
Key results
- On CIFAR-10 with IDN-50% noise, CSGN reaches 95.88% accuracy versus 86.30% for the best baseline DivideMix.
- On Fashion-MNIST IDN-50%, CSGN achieves 94.59% versus 83.35% for DivideMix.
- CSGN achieves lowest transition matrix estimation error across MNIST, Fashion-MNIST, CIFAR-10 and CIFAR-100 at noise rates 0.1 to 0.5 compared with BLTM and MEIDTM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn latent causal structure for noisy-label generation | CSGN · CIFAR-10 IDN-50% versus DivideMix | accuracy: 95.88 ± 0.08 vs 86.30 ± 0.14 | Helped |
| learn latent causal structure for noisy-label generation | CSGN · CIFAR-100 IDN-50% versus CCR | accuracy: 74.60 ± 0.17 vs 64.66 ± 0.11 | Helped |
| replace CausalNL's generative model with the proposed generative model | CausalNL on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 57.93 ± 7.53 -> 88.83 ± 0.09 | Helped |
| replace InstanceGM's generative model with the proposed generative model | InstanceGM on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 95.81 ± 0.07 -> 96.32 ± 0.07 | Helped |
| replace CausalNL's generative model with the proposed generative model | CausalNL on CIFAR-100 IDN-50% · CIFAR-100 IDN-50% | accuracy: 38.69 ± 1.72 -> 59.64 ± 0.99 | Helped |
| replace InstanceGM's generative model with the proposed generative model | InstanceGM on CIFAR-100 IDN-50% · CIFAR-100 IDN-50% | accuracy: 76.64 ± 0.38 -> 79.23 ± 0.12 | Helped |
| add semi-supervised learning warmup rather than regular early stopping (ablation) | CSGN on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 94.43 ± 0.07 -> 95.88 ± 0.08 | Helped |
| add ELBO and LM losses (ablation) | CSGN on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 88.10 ± 0.26 -> 95.88 ± 0.08 | Helped |
| add ELBO loss (ablation) | CSGN on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 88.47 ± 0.13 -> 95.88 ± 0.08 | Helped |
| add LM loss (ablation) | CSGN on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 95.74 ± 0.05 -> 95.88 ± 0.08 | Helped |
| use semi-supervised learning and clean-sample selection instead of PES (ablation) | CSGN-PES on CIFAR-10 IDN-50% · CIFAR-10 IDN-50% | accuracy: 85.45 ± 0.05 -> 95.88 ± 0.08 | Helped |
| replace CSGN with the proposed model | CIFAR-10N and CIFAR-100N classification · CIFAR-10N and CIFAR-100N | accuracy | Helped |
Surprising findings
- Plugging the proposed generative model into CausalNL and InstanceGM improved their accuracy, e.g., CausalNL on CIFAR-100 IDN-50% rose from 38.69% to 59.64%.
Limitations
- Causal relations among latent variables are assumed linear for identifiability; nonlinear relations unresolved.
- Requires a set of clean examples selected from noisy data, and identifiability needs clean examples across distinct classes for each style combination.
- Generative VAE-based design adds computational cost.
Experimental setup
Datasets: Fashion-MNIST CIFAR-10 CIFAR-100 MNIST CIFAR-10N CIFAR-100N WebVision ImageNet ILSVRC12
Benchmarks: Fashion-MNIST CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision ImageNet ILSVRC12 MNIST
Metrics: test accuracy relative estimation error of noise transition matrix
Baselines: CE MentorNet Co-Teaching Reweight Forward PTD CausalNL CCR MEIDTM BLTM NPC RENT
Builds on: CausalNL InstanceGM Co-Teaching DivideMix PTD
Closest papers by meaning
- Learning Causal Transition Matrix for Instance-dependent Label Noise (AAAI 2025 · similarity 0.95)
- Estimating Noise Transition Matrix with Label Correlations for Noisy Multi-Label Learning (NeurIPS 2022 · similarity 0.91)
- Identifiability of Label Noise Transition Matrix (ICML 2023 · similarity 0.91)
- Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.88)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.88)
- Label-Noise Robust Diffusion Models (ICLR 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.