{"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/high-accuracy-low-precision-training","title":"High-Accuracy Low-Precision Training","arxiv_id":"1803.03383","date":"2018-03-09","proceeding":null,"authors":["Christopher De Sa","Megan Leszczynski","Jian Zhang","Alana Marzoev","Christopher R. Aberger","Kunle Olukotun","Christopher Ré"],"abstract":"Low-precision computation is often used to lower the time and energy cost of\nmachine learning, and recently hardware accelerators have been developed to\nsupport it. Still, it has been used primarily for inference - not training.\nPrevious low-precision training algorithms suffered from a fundamental\ntradeoff: as the number of bits of precision is lowered, quantization noise is\nadded to the model, which limits statistical accuracy. To address this issue,\nwe describe a simple low-precision stochastic gradient descent variant called\nHALP. HALP converges at the same theoretical rate as full-precision algorithms\ndespite the noise introduced by using low precision throughout execution. The\nkey idea is to use SVRG to reduce gradient variance, and to combine this with a\nnovel technique called bit centering to reduce quantization error. We show that\non the CPU, HALP can run up to $4 \\times$ faster than full-precision SVRG and\ncan match its convergence trajectory. We implemented HALP in TensorQuant, and\nshow that it exceeds the validation performance of plain low-precision SGD on\ntwo deep learning tasks.","url_abs":"http://arxiv.org/abs/1803.03383v1","url_pdf":"http://arxiv.org/pdf/1803.03383v1.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":"high-accuracy-low-precision-training","repo_url":"https://github.com/HazyResearch/lp_rffs","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[{"task_slug":null,"task_name":"CPU"},{"task_slug":"quantization","task_name":"Quantization"},{"task_slug":"high","task_name":"Vocal Bursts Intensity Prediction"}],"methods":[{"method_slug":"sgd","method_name":"SGD"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1803.03383","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}