{"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/improved-training-of-binary-networks-for","title":"Improved training of binary networks for human pose estimation and image recognition","arxiv_id":"1904.05868","date":"2019-04-11","proceeding":null,"authors":["Adrian Bulat","Georgios Tzimiropoulos","Jean Kossaifi","Maja Pantic"],"abstract":"Big neural networks trained on large datasets have advanced the\nstate-of-the-art for a large variety of challenging problems, improving\nperformance by a large margin. However, under low memory and limited\ncomputational power constraints, the accuracy on the same problems drops\nconsiderable. In this paper, we propose a series of techniques that\nsignificantly improve the accuracy of binarized neural networks (i.e networks\nwhere both the features and the weights are binary). We evaluate the proposed\nimprovements on two diverse tasks: fine-grained recognition (human pose\nestimation) and large-scale image recognition (ImageNet classification).\nSpecifically, we introduce a series of novel methodological changes including:\n(a) more appropriate activation functions, (b) reverse-order initialization,\n(c) progressive quantization, and (d) network stacking and show that these\nadditions improve existing state-of-the-art network binarization techniques,\nsignificantly. Additionally, for the first time, we also investigate the extent\nto which network binarization and knowledge distillation can be combined. When\ntested on the challenging MPII dataset, our method shows a performance\nimprovement of more than 4% in absolute terms. Finally, we further validate our\nfindings by applying the proposed techniques for large-scale object recognition\non the Imagenet dataset, on which we report a reduction of error rate by 4%.","url_abs":"http://arxiv.org/abs/1904.05868v1","url_pdf":"http://arxiv.org/pdf/1904.05868v1.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":"improved-training-of-binary-networks-for","repo_url":"https://github.com/1adrianb/binary-networks-pytorch","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"binarization","task_name":"Binarization"},{"task_slug":"classification-with-binary-neural-network","task_name":"Classification with Binary Neural Network"},{"task_slug":"knowledge-distillation","task_name":"Knowledge Distillation"},{"task_slug":"object-recognition","task_name":"Object Recognition"},{"task_slug":"pose-estimation","task_name":"Pose Estimation"},{"task_slug":"quantization","task_name":"Quantization"}],"methods":[{"method_slug":"knowledge-distillation","method_name":"Knowledge Distillation"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/pose-estimation-on-mpii-human-pose","task":"Pose Estimation","dataset":"MPII Human Pose","model":"Improved Binary Network (HourGlass)","rank_in_archive_order":46,"of":46,"metrics":{"PCKh-0.5":"80.9"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.05868","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}