This paper presents a Light CNN framework for learning a compact face representation from large-scale data with massive noisy labels. The framework introduces a Max-Feature-Map (MFM) operation, which replaces ReLU to suppress low-activation neurons and enhances generalization. MFM uses a competitive relationship between feature maps to separate informative and noisy signals. Three Light CNN models are designed, based on AlexNet, VGG, and ResNet, to reduce parameters and computational costs. A semantic bootstrapping method is proposed to re-label noisy data using pre-trained networks, improving consistency with true labels. Experimental results show that the proposed framework achieves state-of-the-art performance on various face benchmarks without fine-tuning. The Light CNN-29 model with a 256-D representation outperforms other models in terms of accuracy, speed, and storage. The framework is efficient, robust, and suitable for real-time applications. The code is available on GitHub. The paper also discusses the importance of handling noisy labels in face recognition and evaluates the effectiveness of the proposed methods on multiple datasets, including LFW, YTF, MegaFace, IJB, and cross-domain databases. The results demonstrate that the Light CNN framework is effective in learning discriminative face representations and generalizes well across different tasks and domains.This paper presents a Light CNN framework for learning a compact face representation from large-scale data with massive noisy labels. The framework introduces a Max-Feature-Map (MFM) operation, which replaces ReLU to suppress low-activation neurons and enhances generalization. MFM uses a competitive relationship between feature maps to separate informative and noisy signals. Three Light CNN models are designed, based on AlexNet, VGG, and ResNet, to reduce parameters and computational costs. A semantic bootstrapping method is proposed to re-label noisy data using pre-trained networks, improving consistency with true labels. Experimental results show that the proposed framework achieves state-of-the-art performance on various face benchmarks without fine-tuning. The Light CNN-29 model with a 256-D representation outperforms other models in terms of accuracy, speed, and storage. The framework is efficient, robust, and suitable for real-time applications. The code is available on GitHub. The paper also discusses the importance of handling noisy labels in face recognition and evaluates the effectiveness of the proposed methods on multiple datasets, including LFW, YTF, MegaFace, IJB, and cross-domain databases. The results demonstrate that the Light CNN framework is effective in learning discriminative face representations and generalizes well across different tasks and domains.