{"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/rethinking-the-value-of-network-pruning","title":"Rethinking the Value of Network Pruning","arxiv_id":"1810.05270","date":"2018-10-11","proceeding":"ICLR 2019 5","authors":["Zhuang Liu","Ming-Jie Sun","Tinghui Zhou","Gao Huang","Trevor Darrell"],"abstract":"Network pruning is widely used for reducing the heavy inference cost of deep\nmodels in low-resource settings. A typical pruning algorithm is a three-stage\npipeline, i.e., training (a large model), pruning and fine-tuning. During\npruning, according to a certain criterion, redundant weights are pruned and\nimportant weights are kept to best preserve the accuracy. In this work, we make\nseveral surprising observations which contradict common beliefs. For all\nstate-of-the-art structured pruning algorithms we examined, fine-tuning a\npruned model only gives comparable or worse performance than training that\nmodel with randomly initialized weights. For pruning algorithms which assume a\npredefined target network architecture, one can get rid of the full pipeline\nand directly train the target network from scratch. Our observations are\nconsistent for multiple network architectures, datasets, and tasks, which imply\nthat: 1) training a large, over-parameterized model is often not necessary to\nobtain an efficient final model, 2) learned \"important\" weights of the large\nmodel are typically not useful for the small pruned model, 3) the pruned\narchitecture itself, rather than a set of inherited \"important\" weights, is\nmore crucial to the efficiency in the final model, which suggests that in some\ncases pruning can be useful as an architecture search paradigm. Our results\nsuggest the need for more careful baseline evaluations in future research on\nstructured pruning methods. We also compare with the \"Lottery Ticket\nHypothesis\" (Frankle & Carbin 2019), and find that with optimal learning rate,\nthe \"winning ticket\" initialization as used in Frankle & Carbin (2019) does not\nbring improvement over random initialization.","url_abs":"http://arxiv.org/abs/1810.05270v2","url_pdf":"http://arxiv.org/pdf/1810.05270v2.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":[{"paper_slug":"rethinking-the-value-of-network-pruning","repo_url":"https://github.com/Eric-mingjie/rethinking-network-pruning","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"rethinking-the-value-of-network-pruning","repo_url":"https://github.com/liuzhuang13/slimming","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"network-pruning","task_name":"Network Pruning"},{"task_slug":"architecture-search","task_name":"Neural Architecture Search"}],"methods":[{"method_slug":"pruning","method_name":"Pruning"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1810.05270","atlas_url":"https://app.syntology.ai/?focus=1810.05270","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}