Deep Neural Networks Abstraction using An Interval Weights Based Approach - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Deep Neural Networks Abstraction using An Interval Weights Based Approach

Résumé

In this work, we present a Neural Network (NN) abstraction approach to reduce the state-space (number of nodes) of NN towards solving the non-scalability of NN formal verification approaches. The main idea consists in merging neurons on the NN layers in order to build an abstract model that over-approximates the original one. Concretely, the outgoing weights of the abstract network are computed as the sum of the absolute value of the weights on the original one, while the incoming weights are intervals determined based on the signs of the outgoing and the incoming weights of the original model.
Fichier principal
Vignette du fichier
2022_Boudardara_ConfianceAIdays22.pdf (882.79 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04426163 , version 1 (30-01-2024)

Identifiants

  • HAL Id : hal-04426163 , version 1

Citer

Fateh Boudardara, Abderraouf Boussif, Pierre-Jean Meyer, Mohamed Ghazel. Deep Neural Networks Abstraction using An Interval Weights Based Approach. Confiance.ai Days, Oct 2022, Saclay (92), France. ⟨hal-04426163⟩
7 Consultations
4 Téléchargements

Partager

Gmail Facebook X LinkedIn More