Core Concepts
Auto-Train-Once (ATO) introduces an innovative network pruning algorithm designed to automatically reduce the computational and storage costs of DNNs.
Abstract
The content discusses the Auto-Train-Once (ATO) algorithm for automatic network pruning. It addresses limitations in existing methods by introducing a controller network to guide the pruning process. The algorithm aims to simplify the training process and improve performance across various model architectures on benchmark datasets like CIFAR-10, CIFAR-100, and ImageNet.
Directory:
- Abstract
- Proposes ATO for automatic network pruning.
- Introduction
- Discusses challenges in deep neural network pruning.
- Data Extraction
- "Our approach achieves state-of-the-art performance across various model architectures."
- Related Works
- Mentions structural pruning methods and their effectiveness.
- Proposed Method
- Introduces ATO for automatic network pruning with a controller network.
- Convergence and Complexity Analysis
- Provides theoretical analysis ensuring convergence of ATO.
- Experiments
- Evaluates ATO on image classification tasks with ResNet models and MobileNetV2.
Stats
"Our approach achieves state-of-the-art performance across various model architectures."