Search papers, labs, and topics across Lattice.
This paper introduces local population-risk certificates that provide a two-sided confidence band for the risk increment of a model when updated with a local candidate set. The significance lies in the ability to control risk during model updates, ensuring that updates are only accepted if they do not increase risk beyond a certified threshold. The key result is that this method allows for a systematic approach to model updates, enhancing safety and reliability in machine learning applications.
Risk-controlled model updates can now be certified, ensuring that only safe improvements are accepted, which could transform how we approach model training.
This paper develops local certificates for population-risk increments around a current model. For a local candidate set \(\mathcal D\), the certificate is a two-sided confidence band for \(P({\ell_{胃+v}-\ell_胃})\) over \(v\in\mathcal D\). As an application, the upper endpoint of this band yields a risk-controlled update rule: an update is accepted only when its certified upper endpoint is nonpositive; otherwise the current model is retained.
