arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2004.12538cs.LGcs.NE
收藏

通过内部激活节点的选择性可视化来解释深层时间表示

Interpretation of Deep Temporal Representations by Selective Visualization of Internally Activated Nodes

KAIST(韩国科学技术院);UNIST(蔚山科学技术院)

Sohee Cho, Ginkyeng Lee, Wonjoon Chang, Jaesik Choi

摘要:

最近,深度神经网络在许多时间或序列数据的分类和回归任务中表现出竞争力。然而,时间深度神经网络的分类机制仍然很难理解。在本文中,我们提出了两个新的框架来可视化从深度神经网络学习的时间表示。给定输入数据和输出,我们的算法解释的时间神经网络的决定,通过提取高度活跃的时期和可视化的子序列的输入数据,有助于激活的单位。此外,我们的特点,这样的子序列与聚类和计算的不确定性,建议的类型和实际数据。我们还建议从一个单元的输出,而不是从最终的输出,逐层的相关性与向后蒙特卡洛辍学,以显示每个输入点的相关性得分,以激活单元,提供有关这种影响的不确定性的视觉表示。

英文摘要:

Recently deep neural networks demonstrate competitive performances in classification and regression tasks for many temporal or sequential data. However, it is still hard to understand the classification mechanisms of temporal deep neural networks. In this paper, we propose two new frameworks to visualize temporal representations learned from deep neural networks. Given input data and output, our algorithm interprets the decision of temporal neural network by extracting highly activated periods and visualizes a sub-sequence of input data which contributes to activate the units. Furthermore, we characterize such sub-sequences with clustering and calculate the uncertainty of the suggested type and actual data. We also suggest Layer-wise Relevance from the output of a unit, not from the final output, with backward Monte-Carlo dropout to show the relevance scores of each input point to activate units with providing a visual representation of the uncertainty about this impact.

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。

↑