This paper introduces Additive Margin Softmax (AM-Softmax), a novel and interpretable objective function for deep face verification. The authors aim to address the limitations of traditional softmax loss, which excels at separating different classes but struggles with reducing intra-class variation. Unlike existing methods that incorporate angular margins in a multiplicative manner, AM-Softmax uses an additive margin by subtracting a scalar value \( m \) from the cosine similarity \( \cos \theta \). This approach is simpler and more intuitive, making it easier to implement and tune. The paper highlights the importance of feature normalization, which helps in stabilizing the optimization process and improving performance on low-quality images. Experiments on the LFW and MegaFace datasets demonstrate that AM-Softmax consistently outperforms state-of-the-art methods using the same network architecture and training dataset. The authors also provide a geometric interpretation of the loss function, showing how it affects the decision boundaries and intra-class variance. The paper concludes by discussing future directions, including the potential for more creative margin specifications and the need to automatically determine margins and incorporate class-specific or sample-specific margins.This paper introduces Additive Margin Softmax (AM-Softmax), a novel and interpretable objective function for deep face verification. The authors aim to address the limitations of traditional softmax loss, which excels at separating different classes but struggles with reducing intra-class variation. Unlike existing methods that incorporate angular margins in a multiplicative manner, AM-Softmax uses an additive margin by subtracting a scalar value \( m \) from the cosine similarity \( \cos \theta \). This approach is simpler and more intuitive, making it easier to implement and tune. The paper highlights the importance of feature normalization, which helps in stabilizing the optimization process and improving performance on low-quality images. Experiments on the LFW and MegaFace datasets demonstrate that AM-Softmax consistently outperforms state-of-the-art methods using the same network architecture and training dataset. The authors also provide a geometric interpretation of the loss function, showing how it affects the decision boundaries and intra-class variance. The paper concludes by discussing future directions, including the potential for more creative margin specifications and the need to automatically determine margins and incorporate class-specific or sample-specific margins.