Beyond MACs: Hardware Efficient Architecture Design for Vision Backbones
Researchers have developed a new vision backbone architecture called LowFormer, designed for improved hardware efficiency, particularly on edge devices. Unlike previous methods that relied on MACs (Multiply Accumulate operations) as a primary efficiency metric, this paper demonstrates the limitations of MACs and identifies key factors for optimizing backbone design. LowFormer incorporates 'Lowtention,' a more efficient alternative to Multi-Head Self-Attention, and has shown superior performance on ImageNet and various downstream tasks, including object detection and segmentation, across different hardware platforms. AI
IMPACT Introduces a more hardware-efficient vision backbone, potentially accelerating AI applications on edge devices.