Neural Trajectory Analysis of Recurrent Neural Network In Handwriting Synthesis

  • 2018-04-13 11:31:06
  • Kristof B. Charbonneau, Osamu Shouno
  • 3

Abstract

Recurrent neural networks (RNNs) are capable of learning to generate highlyrealistic, online handwritings in a wide variety of styles from a given textsequence. Furthermore, the networks can generate handwritings in the style of aparticular writer when the network states are primed with a real sequence ofpen movements from the writer. However, how populations of neurons in the RNNcollectively achieve such performance still remains poorly understood. Totackle this problem, we investigated learned representations in RNNs byextracting low-dimensional, neural trajectories that summarize the activity ofa population of neurons in the network during individual syntheses ofhandwritings. The neural trajectories show that different writing styles areencoded in different subspaces inside an internal space of the network. Withineach subspace, different characters of the same style are represented asdifferent state dynamics. These results demonstrate the effectiveness ofanalyzing the neural trajectory for intuitive understanding of how the RNNswork.

 

Introduction (beta)

None

 

Conclusion (beta)

None