人工智能培训

搜索

深度学习论文:贝叶斯量化神经网络的无样本学习(Sampling-Free Learning of Bayesian Quantized Neural Netwo

[复制链接]
shanhawk 发表于 2019-12-9 13:31:30 | 显示全部楼层 |阅读模式
shanhawk 2019-12-9 13:31:30 178 0 显示全部楼层
深度学习论文:贝叶斯量化神经网络的无样本学习(Sampling-Free Learning of Bayesian Quantized Neural Networks)在神经网络中模型参数的贝叶斯学习很重要,在这种情况下,具有良好校准不确定性的估计很重要。在本文中,我们提出了贝叶斯量化网络(BQN),量化神经网络(QNN),为此我们了解了它们的离散参数的后验分布。我们提供了一组用于BQN中学习和预测的高效算法,而无需从其参数或激活中进行采样,这不仅允许QNN中的差异学习,而且还减小了梯度的方差。我们评估了MNIST,Fashion-MNIST,KMNIST和CIFAR10图像分类数据集上的BQN,并与QNN(E-QNN)的自举集合进行了比较。我们证明,与E-QNN相比,BQN可以实现更低的预测误差和更好的校准不确定性(负对数可能性小于20%)。
Bayesian learning of model parameters in neural networks is important inscenarios where estimates with well-calibrated uncertainty are important.Inthis paper, we propose Bayesian quantized networks (BQNs), quantized neuralnetworks (QNNs) for which we learn a posterior distribution over their discreteparameters.We provide a set of efficient algorithms for learning andprediction in BQNs without the need to sample from their parameters oractivations, which not only allows for differentiable learning in QNNs, butalso reduces the variance in gradients.We evaluate BQNs on MNIST,Fashion-MNIST, KMNIST and CIFAR10 image classification datasets, comparedagainst bootstrap ensemble of QNNs (E-QNN).We demonstrate BQNs achieve bothlower predictive errors and better-calibrated uncertainties than E-QNN (withless than 20% of the negative log-likelihood).深度学习论文:贝叶斯量化神经网络的无样本学习(Sampling-Free Learning of Bayesian Quantized Neural Networks)
URL地址:https://arxiv.org/abs/1912.02992     ----pdf下载地址:https://arxiv.org/pdf/1912.02992    ----深度学习论文:贝叶斯量化神经网络的无样本学习(Sampling-Free Learning of Bayesian Quantized Neural Networks)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

shanhawk当前离线
新手上路

查看:178 | 回复:0

快速回复 返回顶部 返回列表