Analyzing the Performance of Multilayer Neural Networks for Object\n Recognition

In the last two years, convolutional neural networks (CNNs) have achieved an\nimpressive suite of results on standard recognition datasets and tasks.\nCNN-based features seem poised to quickly replace engineered representations,\nsuch as SIFT and HOG. However, compared to SIFT and HOG, we understand much\nless about the nature of the features learned by large CNNs. In this paper, we\nexperimentally probe several aspects of CNN feature learning in an attempt to\nhelp practitioners gain useful, evidence-backed intuitions about how to apply\nCNNs to computer vision problems.\n