Research map / Label-efficient and robust learning / Noisy labels
Learning Causal Transition Matrix for Instance-dependent 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 a causal graph and "causal transition matrix" P(Ŷ|do(Y),X) for instance-dependent label noise, splitting inputs into noise-resistant and noise-sensitive components and training end-to-end with co-teaching, a transition model, and a policy-gradient component.
Problem
Learning with noisy labels where the transition matrix is instance-dependent and generally considered unidentifiable, so standard instance-independent assumptions fail in real-world label noise.
Approach
Model label noise with a causal graph where latent Z can affect the instance and/or the annotation; split X into a noise-resistant component X1 (used to predict clean Y, decorrelated from noisy labels via a uniform-output regularizer) and a noise-sensitive component X2 (learned via policy gradient to explain the noisy-label transition), so the causal transition matrix P(Ŷ|do(Y),X) is identifiable
Key results
- On CIFAR10 with 50% instance-dependent noise, the method reaches 78.37% vs. 77.39% for CausalNL; on CIFAR100 with 50% IDN it reaches 34.44% vs. 32.13% for CausalNL.
- On Clothing1M the method reaches 72.25% test accuracy vs. 68.90% for CausalNL and 69.91% for Forward.
- On CIFAR-N with ProMix as backbone, Promix+Ours reaches 96.16% on CIFAR10-N Worst and 73.66% on CIFAR100-N Noisy Fine, slightly above ProMix.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the policy model (X2) (ablation) | Ours · FashionMNIST SYM 20% and 80%, ASYM 20% and 40%, IDN 20% and 40% | test accuracy: 92.1 -> 92.1; 75.8 -> 71.5; 89.6 -> 91.4; 82.4 -> 88.7; 91.3 -> 91.0; 90.3 -> 90.4 | Mixed |
| add the policy model (X2) (ablation) | Ours · SVHN IDN 20%, 30%, 40%, 45%, 50% | test accuracy: 93.86±0.17 -> 94.13±0.08; 93.82±0.18 -> 93.97±0.11; 93.50±0.25 -> 93.94±0.16; 93.19±1.25 -> 93.33±1.12; 92.91±1.54 -> 92 | Mixed |
| add the policy model (X2) (ablation) | Ours · CIFAR10 IDN 20%, 30%, 40%, 45%, 50% | test accuracy: 82.57±0.33 -> 82.94±0.29; 81.24±0.36 -> 82.15±0.25; 79.36±0.81 -> 81.04±0.23; 78.43±0.51 -> 80.24±0.39; 75.59±2.07 -> 78 | Helped |
| add the policy model (X2) (ablation) | Ours · CIFAR100 IDN 20%, 30%, 40% | test accuracy: 45.50±0.99 -> 46.37±0.46; 44.67±0.60 -> 43.34±0.39; 38.44±1.40 -> 39.61±1.04 | Mixed |
| add the policy model (X2) (ablation) | Ours · CIFAR100 IDN 45% and 50% | test accuracy: 34.88±2.53 -> 37.04±1.83; 33.05±1.47 -> 34.44±1.86 | Mixed |
Surprising findings
- The ablation without the policy model (w.o/pg) performs best on several settings, e.g. FashionMNIST SYM 80% and CIFAR10-IDN 30-40%, which the authors attribute to X1 decorrelation alone.
Limitations
- Relies on the assumption that the clean label Y is partially observable via confidence sampling, which may fail under high noise rates.
- Training requires a twin network, roughly doubling training compute.
- Some ablation results and baseline comparisons are mixed; the method does not win everywhere (e.g., BLTM-V on SVHN at low noise, and CIFAR100 results where w.o/pg is better at 30%).
Experimental setup
Datasets: FashionMNIST SVHN CIFAR10 CIFAR100 Food101 Clothing1M CIFAR-N
Benchmarks: FashionMNIST SVHN CIFAR10 CIFAR100 Food101 Clothing1M CIFAR-N
Metrics: test accuracy
Baselines: Co-teaching Joint JoCoR CORES2 SCE LS REL Forward DualT TVR MentorNet Mixup
Builds on: co-teaching CausalNL T-Revision Gumbel-Softmax
Closest papers by meaning
- Learning the Latent Causal Structure for Modeling Label Noise (NeurIPS 2024 · similarity 0.95)
- Identifiability of Label Noise Transition Matrix (ICML 2023 · similarity 0.91)
- Instance-dependent Label-noise Learning under a Structural Causal Model (NeurIPS 2021 · similarity 0.90)
- Learning Mixtures of Unknown Causal Interventions (NeurIPS 2024 · similarity 0.87)
- Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.87)
- Noisy Label Learning with Instance-Dependent Outliers: Identifiability via Crowd Wisdom (NeurIPS 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.