In classical statistics, importance weighting is a fundamental tool for handling biased data. If a certain group is underrepresented in your dataset, you simply assign those samples a higher weight in your loss function. This forces the model to "pay more attention" to them, shifting the optimal solution to favor performance on that group.
However, as we move into the regime of over-parameterized deep learning—where models are large enough to perfectly interpolate (achieve zero error on) the training data—this classic intuition breaks down. Recent research reveals that for modern neural networks, importance weights often become "invisible" to the model by the end of training.
The Vanishing Gradient Problem of Interpolation
The core reason importance weighting fails in deep learning is the phenomenon of interpolation. When a high-capacity model reaches zero training loss, the gradient of the loss function with respect to the parameters also goes to zero.
Mathematically, the gradient of a weighted loss is $w_i \cdot \nabla \mathcal{L}_i$. If the model perfectly classifies sample $i$, then $\nabla \mathcal{L}_i$ becomes zero. At this point, it doesn't matter if the weight $w_i$ is 1 or 1,000,000; the product is still zero. As Natalia Martinez, Martin Andres Bertran, and their colleagues point out in Achieving Group Distributional Robustness and Minimax Group Fairness with Interpolating Classifiers, this makes models "blind" to standard fairness adjustments. Once the model has learned to separate the data, the weights lose their power to influence the decision boundary further.
Destination versus Path
If weights don't change the final "destination" of an interpolating model, what do they do? Da Xu, Yuting Ye, and Chuanwei Ruan explore this in Understanding the role of importance weighting for deep learning. They find that while the weights don't change the eventual max-margin solution—the specific decision boundary the model settles on due to the implicit bias of gradient descent—they do significantly impact the speed of convergence.
Think of it like two hikers climbing to a mountain pass. One hiker is pulled by a strong bungee cord (high weight), and the other is not. They might both end up at the exact same narrow pass (the max-margin solution), but the hiker with the bungee cord gets there much faster. In practical deep learning, we rarely train to absolute infinity; we stop after a finite number of steps. Therefore, the weights influence the model's generalization error by changing which part of the "valley" the model occupies when training is halted.
Restoring Influence via Temperature and Margins
Because standard importance weighting vanishes at interpolation, researchers have had to develop new mechanisms to maintain group robustness and fairness.
The framework introduced by Natalia Martinez and Guillermo Sapiro suggests that we need to decouple the learner from the weights. Their approach, TempLoss, uses temperature-scaled margins. By adjusting the "sharpness" of the loss function and enforcing a margin even after samples are correctly classified, they ensure that the gradients do not vanish prematurely. This keeps the importance weights "active," allowing the model to remain sensitive to group-level performance even as it approaches zero training error.
Why This Matters
This shift in understanding changes how we approach model reliability. We can no longer assume that simply re-weighting a dataset will fix a bias or handle a distribution shift in a large-scale Transformer or ResNet. Understanding that the implicit bias of the optimizer often overrides explicit weights is the first step toward building models that are truly robust, not just on the training set, but across diverse real-world groups.