The Living Thing / Notebooks :

Stability (in learning)

Your estimate is robust to a deleted data point? it is a stable estimate. This implies generalisability, apparently. The above statements can be made precise, I am told. Making them precise might give us new ideas for risk bounds, model selection, or connections to optimization.

Supposedly there is also connection to differential privacy, but since I don't yet know anything about differential privacy I can't take that statement any further except to note I would like to work it out one day. This is also presumably a connection to robust inference, since the setup sounds very similar.



Reproducibility is imperative for any scientific discovery. More often than not, modern scientific findings rely on statistical analysis of high-dimensional data. At a minimum, reproducibility manifests itself in stability of statistical results relative to “reasonable” perturbations to data and to the model used. Jacknife, bootstrap, and cross-validation are based on perturbations to data, while robust statistics methods deal with perturbations to models.

Moritz Hardt, Stability as a foundation of machine learning:

Central to machine learning is our ability to relate how a learning algorithm fares on a sample to its performance on unseen instances. This is called generalization.

In this post, I will describe a purely algorithmic approach to generalization. The property that makes this possible is stability. An algorithm is stable, intuitively speaking, if its output doesn’t change much if we perturb the input sample in a single point. We will see that this property by itself is necessary and sufficient for generalization.