This paper presents a deep learning approach for face recognition by jointly using face identification and verification signals as supervision. The proposed method, DeepID2, learns effective face representations through deep convolutional networks that enhance inter-personal variations and reduce intra-personal variations. The DeepID2 features are extracted from different face regions and resolutions, and are concatenated after PCA dimension reduction to form a comprehensive feature representation. The method achieves 99.15% face verification accuracy on the LFW dataset, outperforming previous deep learning results by significantly reducing the error rate. The key idea is to use both identification and verification signals during training. Identification helps separate different identities, while verification ensures that faces of the same identity are close in the feature space. This dual supervision leads to more discriminative features. The deep convolutional networks are designed to extract hierarchical nonlinear features, with the DeepID2 layer fully connected to both the third and fourth convolutional layers. The features are learned under two supervisory signals: identification, which classifies faces into different identities, and verification, which ensures that faces of the same identity are similar. The DeepID2 features are then used in a Joint Bayesian model for face verification, which models the joint probability of two faces being the same or different. The model uses log-likelihood ratio tests to determine face verification accuracy. The method is evaluated on the LFW dataset, achieving high accuracy and outperforming previous methods. The results show that combining identification and verification signals leads to better feature learning and improved face verification performance. The system is efficient, with fast feature extraction and high accuracy. The paper concludes that the joint use of identification and verification signals is essential for learning effective face representations.This paper presents a deep learning approach for face recognition by jointly using face identification and verification signals as supervision. The proposed method, DeepID2, learns effective face representations through deep convolutional networks that enhance inter-personal variations and reduce intra-personal variations. The DeepID2 features are extracted from different face regions and resolutions, and are concatenated after PCA dimension reduction to form a comprehensive feature representation. The method achieves 99.15% face verification accuracy on the LFW dataset, outperforming previous deep learning results by significantly reducing the error rate. The key idea is to use both identification and verification signals during training. Identification helps separate different identities, while verification ensures that faces of the same identity are close in the feature space. This dual supervision leads to more discriminative features. The deep convolutional networks are designed to extract hierarchical nonlinear features, with the DeepID2 layer fully connected to both the third and fourth convolutional layers. The features are learned under two supervisory signals: identification, which classifies faces into different identities, and verification, which ensures that faces of the same identity are similar. The DeepID2 features are then used in a Joint Bayesian model for face verification, which models the joint probability of two faces being the same or different. The model uses log-likelihood ratio tests to determine face verification accuracy. The method is evaluated on the LFW dataset, achieving high accuracy and outperforming previous methods. The results show that combining identification and verification signals leads to better feature learning and improved face verification performance. The system is efficient, with fast feature extraction and high accuracy. The paper concludes that the joint use of identification and verification signals is essential for learning effective face representations.