Konjak: Live Visualization in Deep Neural Network Programming as a Learning Tool

Chunqi Zhao, Tsukasa Fukusato, Jun Kato, Takeo Igarashi

Keywords: deep neural network, visualization, live programming, software engineering, human-computer interaction

Abstract: Visualization in deep neural networks~(DNNs) development could play a key role in helping novice programmers to inspect and understand a network structure. However, these visualizations are usually available only after the implementation of the DNN program. We propose combining a code editor with a live visualization of the DNN structure to assist machine learning novice users during the DNN code development. The user assigns operation blocks and input data size in the code editor, and the system continuously updates the network visualization. The visualization is also editable, where the user can directly use drag-and-drop operations to build a network. To our knowledge, we are the first to tightly combine text-based programming editor with live and editable visualization as an educational tool in DNN programming, which can help understand the concept of shape consistency. This paper describes the system's design rationale and presents an exploratory user study to evaluate its effectiveness as a learning environment.

Confirm No Double Submission: Yes

Withdraw On Reject: Yes