ReCIPH: Relational Coefficients for Input Partitioning Heuristic - Inria - Institut national de recherche en sciences et technologies du numérique
Poster Communications Year : 2022

ReCIPH: Relational Coefficients for Input Partitioning Heuristic

Abstract

With the rapidly advancing improvements to the already successful Deep Learning artifacts, Neural Networks (NN) are poised to permeate a growing number of everyday applications, including ones where safety is paramount and, therefore, formal guarantees are a precious commodity. To this end, Formal Methods, a long-standing, mathematically-inspired field of research saw an effervescent outgrowth targeting NN and advancing almost as rapidly as AI itself. Without a doubt, the most challenging problem facing this new research direction is the scalability to the evergrowing NN models. This paper stems from this need and introduces Relational Coefficients for Input partitioning Heuristic (ReCIPH), accelerating NN analysis. Extensive experimentation is supplied to assert the added value to two different solvers handling several models and properties (coming, in part, from two industrial use-cases).
Fichier principal
Vignette du fichier
26_CameraReady_camera_ready_3.pdf (306.59 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-03926281 , version 1 (06-01-2023)

Identifiers

  • HAL Id : hal-03926281 , version 1

Cite

Serge Durand, Augustin Lemesle, Zakaria Chihani, Caterina Urban, François Terrier. ReCIPH: Relational Coefficients for Input Partitioning Heuristic. 1st Workshop on Formal Verification of Machine Learning (WFVML 2022), Jul 2022, Baltimore, United States. ⟨hal-03926281⟩
155 View
185 Download

Share

More