arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2102.06571stat.MLcs.LG
收藏

贝叶斯神经网络先验知识再探

Bayesian Neural Network Priors Revisited

ETH Zürich(苏黎世联邦理工学院);University of Cambridge(剑桥大学);Google AI Berlin(谷歌柏林人工智能研究院);Imperial College London(伦敦帝国理工学院);University of Bristol(布里斯托尔大学)

Vincent Fortuin, Adrià Garriga-Alonso, Sebastian W. Ober, Florian Wenzel, Gunnar Rätsch, Richard E. Turner, Mark van der Wilk, Laurence Aitchison

摘要:

各向同性高斯先验是现代贝叶斯神经网络推理的事实标准。然而,目前还不清楚这些先验是否准确地反映了我们对权重分布的真实信念或给出最佳性能。为了找到更好的先验,我们研究了使用随机梯度下降(SGD)训练的网络中神经网络权重的汇总统计。我们发现卷积神经网络(CNN)和ResNet权重显示出很强的空间相关性,而全连接网络(FCNN)显示出重尾权重分布。我们表明,将这些观察结果构建到先验中可以提高各种图像分类数据集的性能。令人惊讶的是,这些先验减轻了FCNN中的冷后验效应,但略微增加了ResNets中的冷后验效应。

英文摘要:

Isotropic Gaussian priors are the de facto standard for modern Bayesian neural network inference. However, it is unclear whether these priors accurately reflect our true beliefs about the weight distributions or give optimal performance. To find better priors, we study summary statistics of neural network weights in networks trained using stochastic gradient descent (SGD). We find that convolutional neural network (CNN) and ResNet weights display strong spatial correlations, while fully connected networks (FCNNs) display heavy-tailed weight distributions. We show that building these observations into priors can lead to improved performance on a variety of image classification datasets. Surprisingly, these priors mitigate the cold posterior effect in FCNNs, but slightly increase the cold posterior effect in ResNets.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑