Indexado en

Licencia y uso

Citaciones

Cited 4 times in

Altmetrics

Grant support

E. Goubault-Supported in part by the academic Chair "Engineering of Complex Systems", Thales-Dassault Aviation-Naval Group-DGA-Ecole Polytechnique-ENSTA Paris-Telecom Paris, and AID project "Drone validation and swarms of drones". S. Sankaranarayanan-Supported by US National Science Foundation (NSF) award #1932189. All opinions expressed are those of the authors and not necessarily of the sponsors.

Impacto en los Objetivos de Desarrollo Sostenible (ODS)

Análisis de autorías institucional

Compartir

Static Analysis of ReLU Neural Networks with Tropical Polyhedra

Publicado en:Lecture Notes In Computer Science. 12913 166-190 - 2021-01-01 12913(), DOI: 10.1007/978-3-030-88806-0_8

Autores: Goubault, Eric; Palumby, Sebastien; Putot, Sylvie; Rustenholz, Louis; Sankaranarayanan, Sriram

Afiliaciones

Resumen

This paper studies the problem of range analysis for feedforward neural networks, which is a basic primitive for applications such as robustness of neural networks, compliance to specifications and reachability analysis of neural-network feedback systems. Our approach focuses on ReLU (rectified linear unit) feedforward neural nets that present specific difficulties: approaches that exploit derivatives do not apply in general, the number of patterns of neuron activations can be quite large even for small networks, and convex approximations are generally too coarse. In this paper, we employ set-based methods and abstract interpretation that have been very successful in coping with similar difficulties in classical program verification. We present an approach that abstracts ReLU feedforward neural networks using tropical polyhedra. We show that tropical polyhedra can efficiently abstract ReLU activation function, while being able to control the loss of precision due to linear computations. We show how the connection between ReLU networks and tropical rational functions can provide approaches for range analysis of ReLU neural networks. We report on a preliminary evaluation of our approach using a prototype implementation.

Palabras clave