Algorithmic Stability
Overview
Algorithmic stability studies how much a learned model or prediction changes when training examples, initialization, or computation are perturbed, and how that sensitivity affects generalization and reproducibility. Common tools include uniform, on-average, and pointwise stability, analyzed through coupled runs on neighboring datasets and combined with concentration inequalities to bound generalization gaps. These methods are applied to stochastic gradient descent and its variants, neural networks, decentralized and bilevel optimization, and ensemble or subsampling procedures, with newer work addressing heavy-tailed data, data reuse, and finite-precision updates. A central tension is that stronger stability can improve generalization or reliability but may constrain accuracy, convergence, or regret, while available guarantees can depend sharply on assumptions and may not capture every form of real-world robustness.
Latest papers
48 papers2 papers in the last four weeks, down 33% on the four weeks before. 0.0% of all new papers.