Searcharxiv⌕ Search

arXiv subjects

Gabriele Sanguin

Publications and source records attributed to Gabriele Sanguin.

2 recordsLinked to original sources

Direct Search meets Frank-Wolfe: Inexact Derivative-Free Bilevel Optimization over Polytopes

In this work we propose BLDS-SIMPLEX, a derivative-free direct search method for the solution of bilevel optimization with polytope constraints at the upper level. The proposed method reformulates the feasible set using barycentric coordinates and explores it using Frank-Wolfe-type directions, relying on a fixed-accuracy black-box oracle for the upper-level objective function. We analyze its convergence properties and complexity in the case of smooth reduced objective; additionally, we analyze a variant for the case of nonsmooth objectives and show its convergence to a specific kind of stationary point that we introduce as a natural generalization to the classic stationarity notions in this setting. Numerical tests on modified instances from BOLIB support the viability of our new approach.

math.OC↗

Exploring the Potential of Bilevel Optimization for Calibrating Neural Networks

Handling uncertainty is critical for ensuring reliable decision-making in intelligent systems. Modern neural networks are known to be poorly calibrated, resulting in predicted confidence scores that are difficult to use. This article explores improving confidence estimation and calibration through the application of bilevel optimization, a framework designed to solve hierarchical problems with interdependent optimization levels. A self-calibrating bilevel neural-network training approach is introduced to improve a model's predicted confidence scores. The effectiveness of the proposed framework is analyzed using toy datasets, such as Blobs and Spirals, as well as more practical simulated datasets, such as Blood Alcohol Concentration (BAC). It is compared with a well-known and widely used calibration strategy, isotonic regression. The reported experimental results reveal that the proposed bilevel optimization approach reduces the calibration error while preserving accuracy.

cs.LG↗