Visualizing Gate Dynamics in Long Short-term Memory Neural Networks (LSTMs)
  • Designed and implemented a novel interactive visualization system to study of the gate dynamics of LSTMs
  • Formulated the gate dynamics as multiple time series and combined several visual analytical techniques to allow flexible exploration and pattern discovery
  • Exposed strong patterns of a stacked two-layer language model and confirmed the common belief that LSTMs can selectively carry long-term information
  • Implemented with d3, Babel, webpack, Flask and TensorFlow
Visualizing Convolutional Neural Networks for Text Analytics
  • Designed and implemented a novel interactive visualization system to study of the internal mechanisms of CNNs in the text domain
  • Integrated multiple visualization paradigms, proposed a novel aggregated animation to expose patterns and a novel visual design for large networks
  • Supported instance- and subset-based analysis and comparison
  • Scaled to large networks by clustering, compression and summarization
  • Revealed multiple patterns that facilitated deep learning researchers to understand and improve the performance of multiple part-of-speech classification models
  • Implemented with d3, Flask and TensorFlow
Rapid Sequence Matching for Visualization Recommender Systems
  • Adapted and implemented locality sensitive hashing for rapid visualization matching
  • Proposed a representation of visualizations as set notations, applied MinHash and locality sensitive hashing for rapid matching and proposed multiple metrics to rank recommendations based on the sequence graph
  • Achieved constant time performance with high accuracy on simulated large databases
  • Implemented with Java and Neo4j