Source-linked AI summary
EnD: Entangling and Disentangling deep representations for bias correction
Enzo Tartaglione, Carlo Alberto Barbano, Marco Grangetto
TL;DR
Training-data biases can undermine deep models' generalization, while the specific cues that encode a bias may be unknown. EnD regularizes intermediate representations to disentangle bias information while retaining task-relevant features, and the paper reports improved generalization on unbiased test sets and favorable comparisons with existing debiasing methods.
Problem
Biases in training data can undermine deep models' generalization, and their specific manifestations may be unknown.
Source: §1 · p. 1Method
EnD regularizes an intermediate layer by entangling features with the same target class and disentangling features with the same bias label.
Source: §1 · p. 1Results
Takeaways & Limitations
Abstract
from arXiv · showhide
Artificial neural networks perform state-of-the-art in an ever-growing number of tasks, and nowadays they are used to solve an incredibly large variety of tasks. There are problems, like the presence of biases in the training data, which question the generalization capability of these models. In this work we propose EnD, a regularization strategy whose aim is to prevent deep models from learning unwanted biases. In particular, we insert an "information bottleneck" at a certain point of the deep neural network, where we disentangle the information about the bias, still letting the useful information for the training task forward-propagating in the rest of the model. One big advantage of EnD is that we do not require additional training complexity (like decoders or extra layers in the model), since it is a regularizer directly applied on the trained model. Our experiments show that EnD effectively improves the generalization on unbiased test sets, and it can be effectively applied on real-case scenarios, like removing hidden biases in the COVID-19 detection from radiographic images.
1. Introduction
Training-data biases can let deep models make confident predictions using unintended cues, raising concerns about robustness and generalization. EnD regularizes intermediate features to separate bias-related information from task-relevant representations without adding training overhead.
1. Introduction
- Biased datasets can lead models to make high-confidence predictions from unintended cues, including medical devices, age, or data origin in COVID-19 radiographic-image studies.The work assumes a bias may be present even when its manifestation is unknown.Source: §1 · p. 1
- EnD entangles features from inputs with the same target class and disentangles features from inputs with the same bias label.The regularizer is applied at a layer within the deep model.Source: §1 · p. 1
- EnD is minimized with the task loss by standard back-propagation, without training extra models, adding learned parameters, or modifying input data.The authors report competitive generalization compared with other debiasing techniques.Source: §1 · pp. 1–2
2. Related works
Related work groups debiasing methods into data-level correction, adversarial or ensemble approaches, and techniques applied within the trained model. EnD is positioned in the last group as a direct regularizer without additional learned parameters.
2. Related works
- Data-level debiasing includes strategies that balance datasets and approaches that reduce bias by borrowing data from different sources, with reported improvements in model generalization.Torralba and Efros examined dataset biases; Tommasi et al. tested strategies for balancing data; Gupta et al. [11] used data from different sources.Source: §2 · p. 2
- Adversarial and ensemble approaches include gradient inversion [16], HEX [32], BlindEye [1], ReBias [4], LfF [21], and LearnedMixin [6].These methods use bias classifiers, feature extraction, or ensembles to reduce bias influence.Source: §2 · p. 2
- Within-model methods include corrective losses, logit re-weighting, and Group-DRO; EnD likewise regularizes the trained model without extra learned parameters.The section contrasts these approaches with dataset-level methods, which may lack direct control over removed information, and GAN-based methods, which can be computationally complex.Source: §2 · p. 2
3. Entangling and Disentangling deep representations
EnD regularizes deep features to discourage learning biases while preserving task-relevant information. It combines feature decorrelation within bias classes with feature correlation across bias classes for examples sharing a target class.
3. Entangling and Disentangling deep representations
- EnD is a regularization term applied to deep features to discourage the model from learning biases.Source: §3 · p. 2
3.1. Preliminaries
- The objective adds EnD term R to task loss L, applying R to normalized features at the output of layer Γ.Source: §3.1 · p. 3, Figure 1 · p. 3
3.2. Data correlations
- The similarity matrix G records pairwise feature correlations, with entries indicating directional differences between feature vectors.Source: §3.2 · pp. 3–4
3.3. The EnD regularizer
- EnD combines disentangling R⊥ and entangling R∥ terms using multipliers α and β.Source: §3.3 · p. 4
3.3.1 Disentangling term
- The disentangling term minimizes off-diagonal entries of Gramian matrices for examples sharing a bias class, encouraging their feature vectors toward orthogonality.Source: §3.3.1 · p. 4
3.3.2 Entangling term
- The entangling term maximizes correlations between examples sharing a target class but having different bias classes, avoiding a rule that ignores bias labels.Source: §3.3.2 · p. 4
4. Experiments
The experiments test EnD on controlled color bias and real-world age, gender, and medical-image biases. Across these settings, the reported results show improved performance on unbiased or bias-conflicting evaluations, including COVID chest X-rays.
4. Experiments
- Experiments span synthetic color bias and real-world age, gender, and medical-image biases; EnD results are averaged over three runs.The controlled setup is intended to isolate EnD’s behavior from real-world data noise and uncertainty.Source: §4 · pp. 4–5, §4.1 · p. 5
4.1. Controlled experiments
- Controlled Biased MNIST varies color–digit correlation from ρ = 0.990 to 0.999, with biased and unbiased tests at ρ = 1.0 and 0.1, respectively.The EnD term is applied at the average-pooling layer before the fully connected classifier.Source: §4.1.1 · p. 5
4.1.1 Biased MNIST
- At ρ = 0.999, EnD reaches 52.30% accuracy on the unbiased Biased MNIST test set versus 10.4% for vanilla, and leads the compared debiasing algorithms.The paper reports the highest results among compared methods, particularly at ρ = 0.999 and 0.997.Source: §4.1.1 · p. 5, Table 1 · p. 5
- At ρ = 0.997, both EnD terms improve generalization and their combination performs best; at ρ = 0.995, the kick-in region begins when loss falls below a threshold and R becomes predominant.During this phase, unbiased accuracy rises while biased accuracy briefly drops as the model shifts from color to digit shape.Source: §4.1.1 · pp. 5–6, Table 2 · p. 6
4.2. Real world datasets
- Real-world experiments target age and gender bias in face datasets and hidden bias in COVID-19 computer-aided diagnosis.CelebA and IMDB Face use a ResNet-18, while CORDA uses a DenseNet-121 encoder followed by a two-layer classifier.Source: §4.2 · p. 6
4.2.1 CelebA
- On CelebA, EnD outperforms vanilla on both tasks and exceeds Group DRO and LfF on most evaluation sets, with comparable results on HeavyMakeup bias-conflicting data.Accuracy is computed as the average over target–bias pairs.Source: Table 3 · p. 7, §4.2.1 · p. 7
4.2.2 IMDB Face
- EnD improves results across IMDB Face evaluations and ranks best among compared debiasing methods; after EB2 training, age-prediction accuracy on EB1 rises from 48.91% to 74.25%.The EB1 and EB2 splits reverse the age–gender bias patterns.Source: §4.2.2 · p. 7, Table 4 · p. 7
4.2.3 COVID CXR dataset
- On CORDA-CDSS, EnD raises TNR from 59.26% to 76.30%; on CORDA-SLG, it raises TPR from 52.14% to 68.37%, increasing balanced accuracy on both sets.The vanilla model’s TPR and TNR differ across the institution-specific test sets; EnD maintains similar TPR on CDSS and TNR on SLG.Source: §4.2.3 · p. 8, Table 5 · p. 8
5. Conclusion
EnD regularizes deep models trained on biased datasets to discourage biased features without adding learnable parameters or modifying input data. The reported experiments show benefits on heavily biased datasets and COVID-19 chest X-ray diagnosis.
5. Conclusion
- EnD outperforms other state-of-the-art techniques on heavily biased data, including Biased MNIST at ρ = 0.999 and IMDB.The conclusion highlights these as cases where EnD excels.Source: §5 · p. 8
- On COVID-19 diagnosis from chest X-rays, EnD improves overall test-set performance when bias comes from data source.The source bias is described as not straightforward to detect.Source: §5 · p. 8
- EnD discourages biased deep-feature selection without adding learnable parameters, modifying input data, or introducing additional data priors.The regularizer is applied to the model itself.Source: §5 · p. 8