核心概念
DyCE introduces a dynamic configurable early-exit framework for deep learning compression and scaling, allowing real-time adaptation to varying performance-complexity requirements.
要約
DyCE is a novel approach that decouples design considerations from base models, enabling dynamic switching of exits in real-time. It offers optimized configurations for performance-complexity trade-offs, enhancing model scalability and compression efficiency. By utilizing multiple exits collaboratively, DyCE significantly improves overall system performance compared to individual exits. The proposed search algorithms efficiently generate configurations tailored to specific targets, ensuring fine-grained performance tuning. DyCE's flexibility extends beyond image classification tasks, making it adaptable to various applications and hierarchical models.
統計
DyCE significantly reduces the computational complexity by 23.5% of ResNet152 and 25.9% of ConvNextv2-tiny on ImageNet.
引用
"Dynamic compression methods allocate computational resources based on the complexity of each input sample."
"Designing an early exit system necessitates answering three critical questions: How to construct efficient exit networks? Where should the exits be positioned? When should the system exit?"
"DyCE simplifies the design process of early-exit-based dynamic compression systems by partitioning the considerations during the design of such systems."