Garcia Rafael, Munz Tanja, Weiskopf Daniel
VISUS, University of Stuttgart, 70569, Stuttgart, Germany.
Vis Comput Ind Biomed Art. 2021 Sep 29;4(1):24. doi: 10.1186/s42492-021-00090-0.
In this paper, we introduce a visual analytics approach aimed at helping machine learning experts analyze the hidden states of layers in recurrent neural networks. Our technique allows the user to interactively inspect how hidden states store and process information throughout the feeding of an input sequence into the network. The technique can help answer questions, such as which parts of the input data have a higher impact on the prediction and how the model correlates each hidden state configuration with a certain output. Our visual analytics approach comprises several components: First, our input visualization shows the input sequence and how it relates to the output (using color coding). In addition, hidden states are visualized through a nonlinear projection into a 2-D visualization space using t-distributed stochastic neighbor embedding to understand the shape of the space of the hidden states. Trajectories are also employed to show the details of the evolution of the hidden state configurations. Finally, a time-multi-class heatmap matrix visualizes the evolution of the expected predictions for multi-class classifiers, and a histogram indicates the distances between the hidden states within the original space. The different visualizations are shown simultaneously in multiple views and support brushing-and-linking to facilitate the analysis of the classifications and debugging for misclassified input sequences. To demonstrate the capability of our approach, we discuss two typical use cases for long short-term memory models applied to two widely used natural language processing datasets.
在本文中，我们介绍了一种可视化分析方法，旨在帮助机器学习专家分析循环神经网络中各层的隐藏状态。我们的技术允许用户交互式地检查在将输入序列输入网络的过程中，隐藏状态是如何存储和处理信息的。该技术有助于回答一些问题，比如输入数据的哪些部分对预测有更大影响，以及模型如何将每个隐藏状态配置与特定输出相关联。我们的可视化分析方法包含几个组件：首先，我们的输入可视化展示了输入序列及其与输出的关系（使用颜色编码）。此外，通过使用t分布随机邻域嵌入将隐藏状态非线性投影到二维可视化空间中，以了解隐藏状态空间的形状。还采用轨迹来展示隐藏状态配置演变的细节。最后，一个时间多类热图矩阵可视化多类分类器预期预测的演变，并且一个直方图表示原始空间内隐藏状态之间的距离。不同的可视化在多个视图中同时显示，并支持刷选和链接，以方便对分类进行分析以及对错误分类的输入序列进行调试。为了展示我们方法的能力，我们讨论了应用于两个广泛使用的自然语言处理数据集的长短期记忆模型的两个典型用例。
