Standard machine learning models often fail when they encounter distribution shifts because they latch onto "spurious" shortcuts—features that work in the training set but fail in the real world. Diversify and Disambiguate (DivDis) is a two-stage framework that first forces a model to discover multiple competing explanations for the data, and then uses a tiny amount of extra information to pick the one that actually generalizes.
The Challenge of Underspecification
Machine learning datasets are often underspecified, meaning there are multiple plausible hypotheses that can explain the training data equally well. For example, if every photo of a cow in your training set is on a grassy field and every camel is on sand, a model might learn that "grass = cow." This works perfectly until the model sees a cow on a beach.
The core issue is simplicity bias: models (and even ensembles of models) tend to gravitate toward the easiest feature to learn, even if that feature is a spurious shortcut. Existing robust learning methods often fail when these shortcuts are perfectly correlated with the labels in the training data. Instead of trying to find a single "robust" solution from the start, DivDis proposes finding all plausible solutions and then figuring out which one is correct.
The DivDis Framework
DivDis operates in two distinct stages to overcome simplicity bias:
Stage 1: Diversify. The model is trained using a neural network with a shared backbone but multiple "heads." Each head is trained to predict the correct labels on the source (training) data. Crucially, a Mutual Information (MI) minimization loss is added, which forces the different heads to make disagreeing predictions on a separate set of unlabeled target data. This pushes the heads to look for different predictive features (e.g., one head might look at the core object, while another looks at the background).
Stage 2: Disambiguate. Once a diverse "menu" of hypotheses is generated, the framework must select the best one. DivDis uses Active Querying: it identifies specific data points in the target set where the heads disagree the most and requests the true labels for just those points. Because the heads are already diverse, it often takes very few labels—sometimes just one—to identify which head has captured the true underlying logic.
Key Insights
- Unlabeled Data as a Compass: Labels aren't needed to find different hypotheses; they are only needed to choose between them. Unlabeled target data provides the necessary "directions" in which the functions can vary.
- Ensembles Aren't Enough: Standard ensembles fail because every member usually falls into the same simplicity bias trap. DivDis explicitly penalizes heads for agreeing on the target data, forcing them apart.
- Extreme Label Efficiency: By focusing the disambiguation stage on the most informative (most disagreed-upon) points, DivDis can outperform traditional fine-tuning methods while using up to 32x fewer labels.
- Hyperparameter Robustness: Unlike many Out-of-Distribution (OOD) methods that require a perfectly tuned validation set with group labels, DivDis can be tuned using only the average accuracy on the source data.
Experimental Results
DivDis was evaluated across a wide range of benchmarks, including toy 2D tasks, MNIST-CIFAR collages, standard subpopulation shift datasets (Waterbirds, CelebA, MultiNLI), and real-world medical imaging (CXR-14) and tumor classification (Camelyon17-WILDS). To truly stress-test the method, the authors also introduced "Complete Correlation" (-CC) variants of Waterbirds and CelebA, where the spurious feature is 100% correlated with the label in the training set.
- Breaking the 50% Barrier: On the MNIST-CIFAR task with complete correlation, standard methods like Group DRO or JTT couldn't do better than random guessing (50%). DivDis achieved over 75% accuracy.
- Waterbirds-CC: In the most difficult "complete correlation" setting, DivDis reached 82.4% worst-group accuracy, while a standard ensemble remained stuck at 34.8%.
- Medical Imaging: On the CXR-14 dataset, models often mistakenly use "chest drains" as a shortcut for diagnosing pneumothorax. DivDis significantly improved the Area Under the Curve (AUC) for patients without chest drains, proving it successfully ignored the spurious shortcut.
- Scalability: On the large-scale Camelyon17-WILDS benchmark, DivDis achieved 90.4% accuracy, outperforming established domain adaptation methods like DANN and CORAL.
Limitations and Future Work
While highly effective, DivDis does require the user to decide how many heads ($N$) to train. If the data contains more than $N$ plausible explanations, the model might miss the "correct" one. Furthermore, it assumes that the correct robust function actually exists within the model's hypothesis class and can be found by the optimizer.
Training multiple heads and calculating the MI loss adds computational overhead, though it is only about 2x the cost of standard Empirical Risk Minimization (ERM). Future directions include exploring how this framework could be applied to disentangled feature learning and more complex, non-classification tasks.
Summary updated