Chengxi Ye, Grace Chu, Yanfeng Liu, Yichi Zhang, Lukasz Lew, Andrew Howard
{"title":"以任意精度和稀疏度进行神经网络的鲁棒性训练","authors":"Chengxi Ye, Grace Chu, Yanfeng Liu, Yichi Zhang, Lukasz Lew, Andrew Howard","doi":"arxiv-2409.09245","DOIUrl":null,"url":null,"abstract":"The discontinuous operations inherent in quantization and sparsification\nintroduce obstacles to backpropagation. This is particularly challenging when\ntraining deep neural networks in ultra-low precision and sparse regimes. We\npropose a novel, robust, and universal solution: a denoising affine transform\nthat stabilizes training under these challenging conditions. By formulating\nquantization and sparsification as perturbations during training, we derive a\nperturbation-resilient approach based on ridge regression. Our solution employs\na piecewise constant backbone model to ensure a performance lower bound and\nfeatures an inherent noise reduction mechanism to mitigate perturbation-induced\ncorruption. This formulation allows existing models to be trained at\narbitrarily low precision and sparsity levels with off-the-shelf recipes.\nFurthermore, our method provides a novel perspective on training temporal\nbinary neural networks, contributing to ongoing efforts to narrow the gap\nbetween artificial and biological neural networks.","PeriodicalId":501162,"journal":{"name":"arXiv - MATH - Numerical Analysis","volume":"82 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-09-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Robust Training of Neural Networks at Arbitrary Precision and Sparsity\",\"authors\":\"Chengxi Ye, Grace Chu, Yanfeng Liu, Yichi Zhang, Lukasz Lew, Andrew Howard\",\"doi\":\"arxiv-2409.09245\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"The discontinuous operations inherent in quantization and sparsification\\nintroduce obstacles to backpropagation. This is particularly challenging when\\ntraining deep neural networks in ultra-low precision and sparse regimes. We\\npropose a novel, robust, and universal solution: a denoising affine transform\\nthat stabilizes training under these challenging conditions. By formulating\\nquantization and sparsification as perturbations during training, we derive a\\nperturbation-resilient approach based on ridge regression. Our solution employs\\na piecewise constant backbone model to ensure a performance lower bound and\\nfeatures an inherent noise reduction mechanism to mitigate perturbation-induced\\ncorruption. This formulation allows existing models to be trained at\\narbitrarily low precision and sparsity levels with off-the-shelf recipes.\\nFurthermore, our method provides a novel perspective on training temporal\\nbinary neural networks, contributing to ongoing efforts to narrow the gap\\nbetween artificial and biological neural networks.\",\"PeriodicalId\":501162,\"journal\":{\"name\":\"arXiv - MATH - Numerical Analysis\",\"volume\":\"82 1\",\"pages\":\"\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-09-14\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"arXiv - MATH - Numerical Analysis\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/arxiv-2409.09245\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - MATH - Numerical Analysis","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2409.09245","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Robust Training of Neural Networks at Arbitrary Precision and Sparsity
The discontinuous operations inherent in quantization and sparsification
introduce obstacles to backpropagation. This is particularly challenging when
training deep neural networks in ultra-low precision and sparse regimes. We
propose a novel, robust, and universal solution: a denoising affine transform
that stabilizes training under these challenging conditions. By formulating
quantization and sparsification as perturbations during training, we derive a
perturbation-resilient approach based on ridge regression. Our solution employs
a piecewise constant backbone model to ensure a performance lower bound and
features an inherent noise reduction mechanism to mitigate perturbation-induced
corruption. This formulation allows existing models to be trained at
arbitrarily low precision and sparsity levels with off-the-shelf recipes.
Furthermore, our method provides a novel perspective on training temporal
binary neural networks, contributing to ongoing efforts to narrow the gap
between artificial and biological neural networks.