{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/spectral-representations-for-convolutional","title":"Spectral Representations for Convolutional Neural Networks","arxiv_id":"1506.03767","date":"2015-06-11","proceeding":"NeurIPS 2015 12","authors":["Oren Rippel","Jasper Snoek","Ryan P. Adams"],"abstract":"Discrete Fourier transforms provide a significant speedup in the computation\nof convolutions in deep learning. In this work, we demonstrate that, beyond its\nadvantages for efficient computation, the spectral domain also provides a\npowerful representation in which to model and train convolutional neural\nnetworks (CNNs).\n  We employ spectral representations to introduce a number of innovations to\nCNN design. First, we propose spectral pooling, which performs dimensionality\nreduction by truncating the representation in the frequency domain. This\napproach preserves considerably more information per parameter than other\npooling strategies and enables flexibility in the choice of pooling output\ndimensionality. This representation also enables a new form of stochastic\nregularization by randomized modification of resolution. We show that these\nmethods achieve competitive results on classification and approximation tasks,\nwithout using any dropout or max-pooling.\n  Finally, we demonstrate the effectiveness of complex-coefficient spectral\nparameterization of convolutional filters. While this leaves the underlying\nmodel unchanged, it results in a representation that greatly facilitates\noptimization. We observe on a variety of popular CNN configurations that this\nleads to significantly faster convergence during training.","url_abs":"http://arxiv.org/abs/1506.03767v1","url_pdf":"http://arxiv.org/pdf/1506.03767v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[],"tasks":[{"task_slug":"dimensionality-reduction","task_name":"Dimensionality Reduction"}],"methods":[{"method_slug":"dropout","method_name":"Dropout"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-classification-on-cifar-10","task":"Image Classification","dataset":"CIFAR-10","model":"Spectral Representations for Convolutional Neural Networks","rank_in_archive_order":192,"of":265,"metrics":{"Percentage correct":"91.4"},"uses_additional_data":false},{"leaderboard":"/sota/image-classification-on-cifar-100","task":"Image Classification","dataset":"CIFAR-100","model":"Spectral Representations for Convolutional Neural Networks","rank_in_archive_order":180,"of":211,"metrics":{"Percentage correct":"68.4"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1506.03767","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}