ActNAS: Generating Efficient YOLO Models Using Activation NAS
Sudhakar Sah, Ravish Kumar, Darshan C. Ganji, Ehsan Saboori · 2025
Activation functions introduce non-linearity into neural networks, allowing them to learn complex patterns. Different activation functions impact differently on speed and accuracy; for instance, ReLU is fast but often less precise, while SiLU offers higher accuracy at the expense of speed. Traditionally, a single activation function is used throughout a model. In this work, we conducted a comprehensive study on the effects of using mixed activation functions in YOLO-based models, examining their impact on latency, memory usage, and accuracy across CPU, NPU, and GPU edge devices. We propose Activation NAS (Act-NAS)-a Hardware-Aware Neural Architecture Search (HANAS) method that optimizes activation functions per layer for specific hardware. ActNAS-generated models maintain comparable mean Average Precision (mAP) to baselines, while achieving up to 1.67 ⋉ faster inference and/or 64.15 % lower memory usage. Additionally, we demonstrate that hardware-aware models learn to leverage architectural and compiler-level optimizations, resulting in highly efficient performance tailored to each hardware platform.