Balancing Accuracy and Efficiency in Budget-Aware Early-Exiting Neural Networks - GREYC image
Pré-Publication, Document De Travail Année : 2024

Balancing Accuracy and Efficiency in Budget-Aware Early-Exiting Neural Networks

Résumé

This paper presents an Early Exit Neural Network (EENN) architecture, which enables budgeted classification by dynamically selecting the most relevant exit point for each input sample of a dataset to achieve the best performance while adhering to a pre-defined computational budget. The key contribution of this work is a novel method that jointly learns the classifier model and the sample exiting policy, in contrast to prior approaches that treated these components separately. Specifically, the paper introduces a bi-level optimization framework that simultaneously optimizes the cross-entropy loss of the classifier and the probabilities of each sample exiting at different stages of the network. This joint learning approach allows the classifier parameters and the sample-dependent exiting policy to be mutually optimized, leading to improved classification accuracy under computational constraints. The proposed EENN method is evaluated on three computer vision benchmarks -CIFAR-10, CIFAR-100, and ImageNet -and demonstrates stateof-the-art results in budgeted classification compared to existing early exit strategies. The code for this work will be made publicly available upon acceptance of the paper.

Fichier principal
Vignette du fichier
01748_paper.pdf (466.98 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04683503 , version 1 (02-09-2024)

Identifiants

  • HAL Id : hal-04683503 , version 1

Citer

Youva Addad, Alexis Lechervy, Frédéric Jurie. Balancing Accuracy and Efficiency in Budget-Aware Early-Exiting Neural Networks. 2024. ⟨hal-04683503⟩
57 Consultations
27 Téléchargements

Partager

More