This paper introduces a novel unsupervised feature learning approach called instance-level discrimination, which learns a feature representation that captures apparent similarity among instances rather than classes. The method is based on a non-parametric softmax formulation and uses noise-contrastive estimation (NCE) to handle the computational challenges of large-scale instance classification. The approach is evaluated on ImageNet and Places datasets, demonstrating superior performance compared to state-of-the-art methods. The method is also effective for semi-supervised learning and object detection tasks. The non-parametric model is highly compact, requiring only 600MB storage for a million images, enabling fast nearest neighbor retrieval. The paper also compares the method with other unsupervised learning approaches, including self-supervised learning and exemplar CNN, and shows that the proposed method outperforms them in terms of accuracy and efficiency. The results indicate that the learned features generalize well to new tasks and datasets, demonstrating the effectiveness of the unsupervised learning approach. The method is also shown to be scalable with larger training sets and deeper networks, making it a promising approach for unsupervised feature learning.This paper introduces a novel unsupervised feature learning approach called instance-level discrimination, which learns a feature representation that captures apparent similarity among instances rather than classes. The method is based on a non-parametric softmax formulation and uses noise-contrastive estimation (NCE) to handle the computational challenges of large-scale instance classification. The approach is evaluated on ImageNet and Places datasets, demonstrating superior performance compared to state-of-the-art methods. The method is also effective for semi-supervised learning and object detection tasks. The non-parametric model is highly compact, requiring only 600MB storage for a million images, enabling fast nearest neighbor retrieval. The paper also compares the method with other unsupervised learning approaches, including self-supervised learning and exemplar CNN, and shows that the proposed method outperforms them in terms of accuracy and efficiency. The results indicate that the learned features generalize well to new tasks and datasets, demonstrating the effectiveness of the unsupervised learning approach. The method is also shown to be scalable with larger training sets and deeper networks, making it a promising approach for unsupervised feature learning.