Data Selection for Deep Learning via diversity visualization and scoring

Abstract

Data diversity is a key ingredient for robust deep learning models, especially in the medical domain. We present a diversity visualization and quantification scheme which enables decisions on data selection different enough from already existing data. Out experiments amply validate the usefulness of the proposed diversity metric in terms of enhancement in accuracy of models resulting from using them in data selection decision process with accuracy improvement from 3%->10% across different sites.

Ähnliche Arbeiten