What You See is What You Get: Principled Deep Learning via Distributional Generalization
This work addresses the challenge of unreliable model behavior in deep learning for practitioners, offering a principled approach to generalization, though it builds incrementally on existing DP and stability concepts.
The paper tackles the problem of ensuring machine learning models behave similarly at training and test time, termed the WYSIWYG property, by showing that Differentially-Private (DP) training provably achieves this through distributional generalization. It introduces algorithms that are competitive with state-of-the-art in distributional robustness, improve privacy-disparate impact trade-offs, and mitigate robust overfitting, with significant gains in specific applications.
Having similar behavior at training time and test time $-$ what we call a "What You See Is What You Get" (WYSIWYG) property $-$ is desirable in machine learning. Models trained with standard stochastic gradient descent (SGD), however, do not necessarily have this property, as their complex behaviors such as robustness or subgroup performance can differ drastically between training and test time. In contrast, we show that Differentially-Private (DP) training provably ensures the high-level WYSIWYG property, which we quantify using a notion of distributional generalization. Applying this connection, we introduce new conceptual tools for designing deep-learning methods by reducing generalization concerns to optimization ones: to mitigate unwanted behavior at test time, it is provably sufficient to mitigate this behavior on the training data. By applying this novel design principle, which bypasses "pathologies" of SGD, we construct simple algorithms that are competitive with SOTA in several distributional-robustness applications, significantly improve the privacy vs. disparate impact trade-off of DP-SGD, and mitigate robust overfitting in adversarial training. Finally, we also improve on theoretical bounds relating DP, stability, and distributional generalization.