Machine Learning Could be Easier if All Data Were MNIST
Abstract
MNIST is a famous image dataset; several researchers evaluated their algorithms using MNIST and provided high accuracy. However, the accuracies were degraded on other datasets. Such an aspect raises the assumption that accuracy can be improved if all data were MNIST. Accordingly, this study proposes a preprocessing algorithm to transform all data into MNIST. In the proposal, an autoencoder (AE) is trained from MNIST, where the hypothesis lies that all decoder outputs are MNIST. Then, decoders are transferred to process feature vectors extracted from arbitrary input datasets. In the experiment, transformed data are compared with the original data in supervised classification. Although the accuracy is not improved, the proposed transformation method shows an advantage regarding privacy protection.