This paper introduces Exemplar-CNN, a discriminative unsupervised feature learning method that trains a convolutional neural network (CNN) using only unlabeled data. The key idea is to train the network to discriminate between surrogate classes, where each surrogate class is formed by applying various transformations to a randomly sampled 'seed' image patch. The resulting feature representation is robust to these transformations and outperforms state-of-the-art unsupervised methods on several datasets (STL-10, CIFAR-10, Caltech-101, Caltech-256). While these features are not as effective as class-specific features from supervised training for classification tasks, they perform well on geometric matching problems, outperforming SIFT descriptors. The method generates surrogate training data by applying transformations such as translation, scaling, rotation, contrast, and color variations to image patches. These transformations are used to create surrogate classes, which the CNN is trained to distinguish. The learned features are invariant to the transformations applied during training, making them robust to variations in the data. The Exemplar-CNN is evaluated on classification and descriptor matching tasks, showing superior performance on both. On descriptor matching, the features outperform SIFT and AlexNet, especially when blur is included as a transformation during training. The paper also discusses the invariance properties of the learned features and their performance on different datasets. The results show that the Exemplar-CNN can generalize well to new datasets and that the inclusion of blur as a transformation improves performance on matching tasks. The method is flexible and can be adapted to various applications by extending the list of transformations. Overall, the Exemplar-CNN demonstrates the effectiveness of unsupervised feature learning for tasks beyond classification, particularly in geometric matching.This paper introduces Exemplar-CNN, a discriminative unsupervised feature learning method that trains a convolutional neural network (CNN) using only unlabeled data. The key idea is to train the network to discriminate between surrogate classes, where each surrogate class is formed by applying various transformations to a randomly sampled 'seed' image patch. The resulting feature representation is robust to these transformations and outperforms state-of-the-art unsupervised methods on several datasets (STL-10, CIFAR-10, Caltech-101, Caltech-256). While these features are not as effective as class-specific features from supervised training for classification tasks, they perform well on geometric matching problems, outperforming SIFT descriptors. The method generates surrogate training data by applying transformations such as translation, scaling, rotation, contrast, and color variations to image patches. These transformations are used to create surrogate classes, which the CNN is trained to distinguish. The learned features are invariant to the transformations applied during training, making them robust to variations in the data. The Exemplar-CNN is evaluated on classification and descriptor matching tasks, showing superior performance on both. On descriptor matching, the features outperform SIFT and AlexNet, especially when blur is included as a transformation during training. The paper also discusses the invariance properties of the learned features and their performance on different datasets. The results show that the Exemplar-CNN can generalize well to new datasets and that the inclusion of blur as a transformation improves performance on matching tasks. The method is flexible and can be adapted to various applications by extending the list of transformations. Overall, the Exemplar-CNN demonstrates the effectiveness of unsupervised feature learning for tasks beyond classification, particularly in geometric matching.