{"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/eraserelu-a-simple-way-to-ease-the-training","title":"EraseReLU: A Simple Way to Ease the Training of Deep Convolution Neural Networks","arxiv_id":"1709.07634","date":"2017-09-22","proceeding":null,"authors":["Xuanyi Dong","Guoliang Kang","Kun Zhan","Yi Yang"],"abstract":"For most state-of-the-art architectures, Rectified Linear Unit (ReLU) becomes\na standard component accompanied with each layer. Although ReLU can ease the\nnetwork training to an extent, the character of blocking negative values may\nsuppress the propagation of useful information and leads to the difficulty of\noptimizing very deep Convolutional Neural Networks (CNNs). Moreover, stacking\nlayers with nonlinear activations is hard to approximate the intrinsic linear\ntransformations between feature representations.\n  In this paper, we investigate the effect of erasing ReLUs of certain layers\nand apply it to various representative architectures following deterministic\nrules. It can ease the optimization and improve the generalization performance\nfor very deep CNN models. We find two key factors being essential to the\nperformance improvement: 1) the location where ReLU should be erased inside the\nbasic module; 2) the proportion of basic modules to erase ReLU; We show that\nerasing the last ReLU layer of all basic modules in a network usually yields\nimproved performance. In experiments, our approach successfully improves the\nperformance of various representative architectures, and we report the improved\nresults on SVHN, CIFAR-10/100, and ImageNet. Moreover, we achieve competitive\nsingle-model performance on CIFAR-100 with 16.53% error rate compared to\nstate-of-the-art.","url_abs":"http://arxiv.org/abs/1709.07634v2","url_pdf":"http://arxiv.org/pdf/1709.07634v2.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":"blocking","task_name":"Blocking"},{"task_slug":"image-classification","task_name":"Image Classification"}],"methods":[{"method_slug":"relu","method_name":"ReLU"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/image-classification-on-svhn","task":"Image Classification","dataset":"SVHN","model":"EraseReLU","rank_in_archive_order":11,"of":62,"metrics":{"Percentage error":"1.54"},"uses_additional_data":false}],"syntology":{"syntology_url":null,"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}