Abstract
Face recognition models have achieved remarkable advances through deep learning, with many techniques matching or surpassing human-level recognition performance under diverse environmental conditions. However, while prior research has predominantly focused on improving recognition accuracy, little attention has been paid to improving computational efficiency and reducing memory usage. These aspects are critical for deploying face recognition systems, efficiently processing large-scale data, and enabling fast inference. To address this gap, we propose an approach that employs 1-bit activations and weights in widely used face recognition models, such as AdaFace, while preserving high recognition accuracy. Additionally, we significantly accelerate inference by using a custom CUDA kernel tailored to our specific convolutional requirements. Finally, we demonstrate the generalizability of our method, achieving promising results across five standard face recognition datasets. This work paves the way for more efficient and scalable face recognition solutions without compromising performance.