{"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/convolutional-neural-networks-using","title":"Convolutional Neural Networks using Logarithmic Data Representation","arxiv_id":"1603.01025","date":"2016-03-03","proceeding":null,"authors":["Daisuke Miyashita","Edward H. Lee","Boris Murmann"],"abstract":"Recent advances in convolutional neural networks have considered model\ncomplexity and hardware efficiency to enable deployment onto embedded systems\nand mobile devices. For example, it is now well-known that the arithmetic\noperations of deep networks can be encoded down to 8-bit fixed-point without\nsignificant deterioration in performance. However, further reduction in\nprecision down to as low as 3-bit fixed-point results in significant losses in\nperformance. In this paper we propose a new data representation that enables\nstate-of-the-art networks to be encoded to 3 bits with negligible loss in\nclassification performance. To perform this, we take advantage of the fact that\nthe weights and activations in a trained network naturally have non-uniform\ndistributions. Using non-uniform, base-2 logarithmic representation to encode\nweights, communicate activations, and perform dot-products enables networks to\n1) achieve higher classification accuracies than fixed-point at the same\nresolution and 2) eliminate bulky digital multipliers. Finally, we propose an\nend-to-end training procedure that uses log representation at 5-bits, which\nachieves higher final test accuracy than linear at 5-bits.","url_abs":"http://arxiv.org/abs/1603.01025v2","url_pdf":"http://arxiv.org/pdf/1603.01025v2.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":"convolutional-neural-networks-using","repo_url":"https://github.com/Enderdead/Pytorch_Quantize_impls","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"classification","task_name":"General Classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1603.01025","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}