{"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/ampnet-asynchronous-model-parallel-training","title":"AMPNet: Asynchronous Model-Parallel Training for Dynamic Neural Networks","arxiv_id":"1705.09786","date":"2017-05-27","proceeding":"ICLR 2018 1","authors":["Alexander L. Gaunt","Matthew A. Johnson","Maik Riechert","Daniel Tarlow","Ryota Tomioka","Dimitrios Vytiniotis","Sam Webster"],"abstract":"New types of machine learning hardware in development and entering the market\nhold the promise of revolutionizing deep learning in a manner as profound as\nGPUs. However, existing software frameworks and training algorithms for deep\nlearning have yet to evolve to fully leverage the capability of the new wave of\nsilicon. We already see the limitations of existing algorithms for models that\nexploit structured input via complex and instance-dependent control flow, which\nprohibits minibatching. We present an asynchronous model-parallel (AMP)\ntraining algorithm that is specifically motivated by training on networks of\ninterconnected devices. Through an implementation on multi-core CPUs, we show\nthat AMP training converges to the same accuracy as conventional synchronous\ntraining algorithms in a similar number of epochs, but utilizes the available\nhardware more efficiently even for small minibatch sizes, resulting in\nsignificantly shorter overall training times. Our framework opens the door for\nscaling up a new class of deep learning models that cannot be efficiently\ntrained today.","url_abs":"http://arxiv.org/abs/1705.09786v3","url_pdf":"http://arxiv.org/pdf/1705.09786v3.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":"ampnet-asynchronous-model-parallel-training","repo_url":"https://github.com/facebookresearch/fairscale","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"deep-learning","task_name":"Deep Learning"},{"task_slug":"dynamic-neural-networks","task_name":"Dynamic neural networks"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1705.09786","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}