%0 Conference Paper %1 10.1145/3430036.3430047 %A Garcia, Rafael %A Weiskopf, Daniel %B Proceedings of the 13th International Symposium on Visual Information Communication and Interaction %C New York, NY, USA %D 2020 %I Association for Computing Machinery %K 2020 b01 sfbtrr161 visus visus:weiskopf %R 10.1145/3430036.3430047 %T Inner-Process Visualization of Hidden States in Recurrent Neural Networks %U https://doi.org/10.1145/3430036.3430047 %X In this paper, we introduce a visualization technique aimed to help machine learning experts to analyze the hidden states of layers in recurrent neural networks (RNNs). Our technique allows the user to visually inspect how hidden states store and process information throughout the feeding of an input sequence into the network. It can answer questions such as which parts of the input data had a higher impact on the prediction and how the model correlates each hidden state configuration with a certain output. Our visualization comprises several components: our input visualization shows the input sequence and how it relates to the output (using color coding); hidden states are visualized by nonlinear projection to 2-D visualization space via t-SNE in order to understand the shape of the space of hidden states; time curves are employed to show the details of the evolution of hidden state configurations; and a time-multi-class heatmap matrix visualizes the evolution of expected predictions for multi-class classifiers. To demonstrate the capability of our approach, we discuss two typical use cases for long short-term memory (LSTM) models applied to two widely used natural language processing (NLP) datasets. %@ 9781450387507

@inproceedings{10.1145/3430036.3430047, abstract = {In this paper, we introduce a visualization technique aimed to help machine learning experts to analyze the hidden states of layers in recurrent neural networks (RNNs). Our technique allows the user to visually inspect how hidden states store and process information throughout the feeding of an input sequence into the network. It can answer questions such as which parts of the input data had a higher impact on the prediction and how the model correlates each hidden state configuration with a certain output. Our visualization comprises several components: our input visualization shows the input sequence and how it relates to the output (using color coding); hidden states are visualized by nonlinear projection to 2-D visualization space via t-SNE in order to understand the shape of the space of hidden states; time curves are employed to show the details of the evolution of hidden state configurations; and a time-multi-class heatmap matrix visualizes the evolution of expected predictions for multi-class classifiers. To demonstrate the capability of our approach, we discuss two typical use cases for long short-term memory (LSTM) models applied to two widely used natural language processing (NLP) datasets.}, added-at = {2021-06-23T09:24:03.000+0200}, address = {New York, NY, USA}, articleno = {20}, author = {Garcia, Rafael and Weiskopf, Daniel}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/21787b2075fa8eaefb5e0bdca1400dddc/christinawarren}, booktitle = {Proceedings of the 13th International Symposium on Visual Information Communication and Interaction}, doi = {10.1145/3430036.3430047}, interhash = {d7c1287419fccdcc102bfa30b00f1c07}, intrahash = {1787b2075fa8eaefb5e0bdca1400dddc}, isbn = {9781450387507}, keywords = {2020 b01 sfbtrr161 visus visus:weiskopf}, location = {Eindhoven, Netherlands}, numpages = {5}, publisher = {Association for Computing Machinery}, series = {VINCI '20}, timestamp = {2022-01-13T07:57:15.000+0100}, title = {Inner-Process Visualization of Hidden States in Recurrent Neural Networks}, url = {https://doi.org/10.1145/3430036.3430047}, year = 2020 }