Visualizing examples of deep neural networks at scale

Abstract

Many programmers want to use deep learning due to its superior accuracy in many challenging domains. Yet our formative study flter the corpus down to projects tackling similar tasks and comwith ten programmers indicated that, when constructing their own pare design choices made by others. We evaluated ExampleNet in deep neural networks (DNNs), they often had a difcult time choos- a within-subjects study with sixteen participants. Compared with ing appropriate model structures and hyperparameter values. This the control condition (i.e., online search), participants using Expaper presents ExampleNet-a novel interactive visualization sys- ampleNet were able to inspect more online examples, make more tem for exploring common and uncommon design choices in a tem for exploring common and uncommon design choices in a data-driven design decisions, and make fewer design mistakes.