DeepVis: A Visual Interactive System for Exploring Performance of Deep Learning Models

Abstract

Nowadays, deep learning (DL) models have been an emerging technology because of their performances and wide acceptance in various fields. However, in most cases, the performance analysis of DL models is not viable to understand how they predict because they are inherently considered black-boxes, and different models have different performance rates. Ad-ditionally, due to a lack of highly technical expertise and domain knowledge, people struggle to choose a proper model for their work. Therefore, to understand and improve the performance of the D L model, careful selection of model, layer, epoch, optimizer, hyperparameter tuning, and model visualization is essential. In this paper, we design an interactive visualization system named DeepVis with a wide range of performance evaluation methods that assist the non-expert in adopting an appropriate model. Finally, we demonstrate use cases and expert opinion using a publicly available dataset to validate the usability and effectiveness of Deep Vis.