A New Structure and New Critiques for Dataset of Image Classification
5 Pages Posted: 24 Jan 2018
Date Written: January 17, 2018
Abstract
Dataset is an important part of image classification. The current dataset structure has following problems: lack of mutually exclusive explicit definition for each class, and the definitions lack of features to distinguish the class, and the instances are incoherent to the definition, and the label can’t present uncertainty, and there is only critique: the accuracy of algorithms. In this paper, I demonstrate a new structure of dataset in image classification, including: mutually exclusive explicit definition of labels and the label of labels, which classify labels into coherent, wrong and uncertain (including multi-objects, mid-object, unknown and unclear) and explore three datasets (MNIST, CIFAR-10 and Fashion-MNIST) to explain why it is necessary. And I define a new critique to assess algorithms performance on uncertainty and anti-attack. At end I provide some advice on how to build such a dataset.
Suggested Citation: Suggested Citation