Seminar Details

The safe and responsible deployment of machine learning in real-world applications requires learned models to satisfy critical safety, reliability, and interpretability constraints without compromising predictive accuracy or computational efficiency. This work addresses the challenge of enforcing three such critical properties: properness, adversarial robustness, and discrete Lipschitz continuity. Expressing these properties as global constraints to be enforced throughout the entirety of the training makes the problem computationally expensive.

To overcome this, we propose a framework based on “local correctors”—efficient query-time procedures that transform a base hypothesis satisfying a relaxed constraint into one that fully and strictly satisfies the target property. This localized approach reduces the complex globally-constrained optimization task during training to a relaxed version, which can then be solved efficiently using traditional polynomial approximation methods and standard optimization techniques. We design and analyze several such local correction algorithms and build learners and testers around them, producing provably safe and interpretable learned models with efficiency comparable to that of traditional learning.