神经网络可解释性可视化方法研究

Research on Interpretable Visualization Methods of Neural Networks

深度学习的发展,促使神经网络在现实各个领域得到广泛应用。神经网络可解释性的缺乏是的其在安全性、可靠性要求较高的行业没有得到实质性应用。可视化方法立足对神经网络结构、特征的诠释,是一种很好的神经网络解释性方法。依据神经网络结构、训练阶段,归纳可视化方法为:特征可视化、关系可视化和过程可视化。最后利用Grad-CAM方法举例了特征可视化,描述了其工作原理。

The lack of interpretability of neural network makes it not be applied substantively in industries requiring high se⁃curity and reliability.The visualization method is based on the interpretation of the structure and characteristics of neural network,which is a good interpretation method of neural network.According to neural network structure and training stage,the visualization methods are summarized as:feature visualization,relationship visualization and process visualization.Finally,the Grad-CAM method is used to illustrate the feature visualization and describe its working principle.