This paper introduces DeepID2+, a high-performance deep convolutional network designed for face recognition. DeepID2+ is trained using identification-verification supervisory signals, achieving state-of-the-art performance on LFW and YouTube Faces benchmarks. The authors discover three critical properties of its deep neural activations: sparsity, selectiveness, and robustness. Sparsity is observed in the moderate activation of neurons, which maximizes the discriminative power of the network. Selectiveness is evident in the high selectivity of neurons to identities and identity-related attributes, allowing for accurate classification with a single neuron. Robustness is demonstrated in the network's ability to handle occlusions, maintaining high accuracy even when faces are partially occluded. These properties are naturally achieved through large-scale training without explicit regularization, providing valuable insights into the intrinsic properties of deep networks.This paper introduces DeepID2+, a high-performance deep convolutional network designed for face recognition. DeepID2+ is trained using identification-verification supervisory signals, achieving state-of-the-art performance on LFW and YouTube Faces benchmarks. The authors discover three critical properties of its deep neural activations: sparsity, selectiveness, and robustness. Sparsity is observed in the moderate activation of neurons, which maximizes the discriminative power of the network. Selectiveness is evident in the high selectivity of neurons to identities and identity-related attributes, allowing for accurate classification with a single neuron. Robustness is demonstrated in the network's ability to handle occlusions, maintaining high accuracy even when faces are partially occluded. These properties are naturally achieved through large-scale training without explicit regularization, providing valuable insights into the intrinsic properties of deep networks.