【阅读笔记】Layer-wise relevance propagation for neural networks with local renormalization layers

本文主要是介绍【阅读笔记】Layer-wise relevance propagation for neural networks with local renormalization layers，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

Binder, Alexander, et al. “Layer-wise relevance propagation for neural networks with local renormalization layers.” International Conference on Artificial Neural Networks. Springer, Cham, 2016.

本文是探究的是图片上的像素与最终结果的相关性。创新点是把 Layer-wise Relevance Propagation (LRP) 扩展到了非线性映射上。

Layer-wise Relevance Propagation for Neural Networks

在神经网络中，每一层的输出$x_j$都是输入$x_i$和激活函数$g$的结果（作者写着主要是约定一下各个符号的意思）：
$$x_j=g(\sum _I{w}_{ij}{x}_i+b)$$

给定一个图像$x$和分类器$f$，$x$上每一个像素$p$都有一个 pixel-wise relevance score $R_p^{(1)}$使得：
$$f(x)=\sum _p{R}_p^{(1)}$$

这篇关于【阅读笔记】Layer-wise relevance propagation for neural networks with local renormalization layers的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！

---