Research map / Label-efficient and robust learning / Recognition and long-tail
NeuronTune: Towards Self-Guided Spurious Bias Mitigation
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 NeuronTune, a post hoc, annotation-free method that identifies "biased" embedding dimensions via activation distributions of correct vs. incorrect predictions and zeroes them while retraining the last layer to reduce spurious bias.
Problem
Deep networks trained with ERM rely on spurious correlations, hurting worst-group accuracy; existing mitigation methods need group labels or only indirectly reweight samples, so mitigation without spurious-correlation annotations is needed.
Approach
For each class and embedding dimension, compute a spuriousness score (median activation of misclassified minus correctly classified samples); dimensions with positive scores are treated as biased and suppressed while the last layer is retrained with class-balanced data, with a model-selection metric (SFit) that needs no group labels.
Key results
- On Waterbirds, NeuronTune^† reaches 92.5% WGA with a 2.0% accuracy gap in the unsupervised setting, versus 90.4% for AFR (which uses group labels).
- On CelebA, NeuronTune^† reaches 87.3% WGA, vs. 83.9% for SELF (which uses validation group labels) and 81.1% for JTT.
- On ImageNet-A (ResNet-18 trained on ImageNet-9), NeuronTune attains 37.3% accuracy, the best reported, with the smallest accuracy gap (56.4) among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply NeuronTune instead of ERM | ResNet-50 · Waterbirds | worst-group accuracy (%): 72.6 -> 92.2 | Helped |
| apply NeuronTune instead of ERM | ResNet-50 · CelebA | worst-group accuracy (%): 47.2 -> 83.1 | Helped |
| apply NeuronTune instead of ERM | BERT · MultiNLI | worst-group accuracy (%): 67.9 -> 72.1 | Helped |
| apply NeuronTune instead of ERM | BERT · CivilComments | worst-group accuracy (%): 57.4 -> 82.4 | Helped |
| apply NeuronTune instead of ERM | ResNet-18 · ImageNet-9 | average accuracy (%): $90.8_{\pm0.6}$ -> $93.7_{\pm0.1}$ | Hurt |
| apply NeuronTune instead of ERM | ResNet-18 · ImageNet-A | average accuracy (%): $24.9_{\pm1.1}$ -> $\textbf{37.3}_{\pm 0.5}$ | Helped |
| use D_val rather than D_train for identification (ablation) | NeuronTune · Waterbirds | worst-group accuracy (%): 78.0$_{\pm 2.3}$ -> 92.2$_{\pm 0.3}$ | Helped |
| use D_val rather than D_train for identification (ablation) | NeuronTune · CelebA | worst-group accuracy (%): 58.5$_{\pm 1.2}$ -> 83.1$_{\pm 1.1}$ | Helped |
| use D_val rather than D_train for identification (ablation) | NeuronTune · MultiNLI | worst-group accuracy (%): 42.0$_{\pm 10.5}$ -> 72.1$_{\pm 0.1}$ | Helped |
| use D_val rather than D_train for identification (ablation) | NeuronTune · CivilComments | worst-group accuracy (%): 80.0$_{\pm 10.5}$ -> 82.4$_{\pm 0.2}$ | Helped |
| add neuron-based tuning to class-balanced tuning (ablation) | NeuronTune · Waterbirds | worst-group accuracy (%): 82.7$_{\pm 0.4}$ -> 92.2$_{\pm 0.3}$ | Helped |
| add neuron-based tuning to class-balanced tuning (ablation) | NeuronTune · CelebA | worst-group accuracy (%): 53.9$_{\pm 0.0}$ -> 83.1$_{\pm 1.1}$ | Helped |
Surprising findings
- Using the training data to both identify biased dimensions and tune the model gives low WGA (e.g., 42.0% on MultiNLI), so a held-out split is needed for identification.
Limitations
- Trade-off between average accuracy and worst-group accuracy, especially when same-attribute samples of different classes are hard to separate in latent space.
- Theory relies on a linear data model and Gaussian assumptions; the threshold lambda=0 is a heuristic.
- Identification and tuning data choice (held-out split) affects results and is left open; using validation data for identification acts like model selection.
Experimental setup
Datasets: Waterbirds CelebA ImageNet-9 ImageNet-A MultiNLI CivilComments
Benchmarks: Waterbirds CelebA ImageNet-9 ImageNet-A MultiNLI CivilComments
Metrics: worst-group accuracy average accuracy accuracy gap
Baselines: ERM JTT SELF CNC BAM AFR DFR BPA GEORGE StylisedIN RUBi ReBias
Builds on: Just Train Twice (JTT) Deep Feature Reweighting (DFR) Approximate Feature Reweighting (AFR) BAM
Closest papers by meaning
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.86)
- From Hope to Safety: Unlearning Biases of Deep Models via Gradient Penalization in Latent Space (AAAI 2024 · similarity 0.85)
- Severing Spurious Correlations with Data Pruning (ICLR 2025 · similarity 0.85)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.84)
- Neuron Activation Coverage: Rethinking Out-of-distribution Detection and Generalization (ICLR 2024 · similarity 0.84)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 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.