Visual Analytics System for Hidden States in Recurrent Neural Networks

A visual analytics system and demonstration data for interpreting hidden states in Long Short-Term Memory (LSTM) models. The project includes source code for preprocessing and visualization, with precomputed data for immediate use. It was created by Tanja Munz at the University of Stuttgart.

Use Cases

• Visualizing hidden state dynamics in LSTM models based on the described visual analytics system.
• Analyzing model behavior on text classification tasks using the provided IMDB and Reuters demonstration data.
• Generating custom data for hidden state exploration using the included data preparation scripts.

Strengths

• Includes precomputed demonstration data for immediate use without running preprocessing steps.
• Provides source code for both data preparation (training LSTM models) and the visual analytics system.
• Supports analysis of two standard text datasets (IMDB and Reuters) available in Keras.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment.

Provenance

Source

University of Stuttgart

Collection Method

Data likely generated by training LSTM models on Keras datasets (IMDB, Reuters).