Enregistré dans:
| Auteurs principaux: | , |
|---|---|
| Format: | Recurso digital |
| Langue: | |
| Publié: |
Zenodo
2025
|
| Accès en ligne: | https://doi.org/10.5281/zenodo.17819180 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
Table des matières:
- Deep learning models have achieved remarkable success across various domains, but their computational demands often limit their deployment on resource-constrained devices. This paper introduces a novel approach called "Adaptive Early Exits" that aims to bridge the gap between accuracy and efficiency by dynamically adjusting the inference process based on available resources and input complexity. We propose a framework that integrates multiple exit points within a deep neural network, each associated with a lightweight classifier. An adaptive control mechanism learns to select the earliest exit point that satisfies a pre-defined accuracy threshold, effectively reducing computational cost for simpler inputs while maintaining high accuracy for more complex ones. Our resource-aware optimization strategy considers factors such as latency, energy consumption, and memory footprint during training, allowing the model to adapt to different deployment scenarios. We evaluate our approach on several benchmark datasets and demonstrate significant improvements in inference speed and energy efficiency compared to traditional deep learning models, with minimal impact on accuracy. The results highlight the potential of Adaptive Early Exits to enable the deployment of deep learning models on edge devices and in real-time applications where resource constraints are paramount.