Deep Learning in Practice: Guidelines for Model Selection
Ganesan Ramachandran, Vedant Bhatia · 2019
Deep Learning models are becoming the de facto standard in most image, text and speech understanding tasks. Model selection based only on numerical metrics such as accuracy and inference time is not sufficient especially in highly regulated industries like healthcare or autonomous driving where lives are at stake. In order to trust the model, interpreting the reasons behind the model's decisions is essential. We propose a three pronged approach in selecting models based on accuracy, inference time and on whether they learn the right features.