Tanja Munz-Körner

Science meets Art

Hidden States in Recurrent Neural Networks

Project Overview

This project focuses on a visual analytics approach designed for machine learning experts who want to analyze the hidden states of layers within recurrent neural networks (RNNs). This technique allows users to interactively explore how hidden states store and process information as the input sequence propagates through the network. By employing this method, users can gain insights into the prediction mechanism, identifying which aspects of the input data have a significant impact on predictions and understanding how the model links specific outputs to various configurations of hidden states. This work builds upon the research conducted by Garcia et al. (Inner-process visualization of hidden states in recurrent neural networks).

My Contribution

I utilized materials from the previous project, including the source code of Jupyter notebooks used for creating visualizations, to develop an interactive visualization system using Python 3, JavaScript, and D3.js. I added new visualizations, provided a range of options for customizing them, and ensured more consistent use of color across different visualizations. Additionally, I updated and expanded the original paper to thoroughly describe all new features and included new examples. This involved creating all illustrations for the paper and producing a video using the interactive system.

Publications

Results of this project are published in the paper Visual Analytics Tool for the Interpretation of Hidden States in Recurrent Neural Networks. Additionally, content is presented in the poster Visual-Explainable AI: The Use Case of Language Models.