Analysis of binarization techniques and Tsetlin machine architectures targeting image classification

Abstract

The Tsetlin Machine is a constantly evolving and developing machine learning technique with ever-increasing success. However, for every success, the Tsetlin Machine achieves, a new set of challenges are put ahead. To sufficiently bring the Tsetlin Machine to a broadly used standard, these challenges must be completed. This thesis focuses on the challenge of doing color image classification and will provide an introductory description of how this is possible through the usage of an older technique, namely binarization. A comparison with the various Tsetlin Machine adaptations made public in recent times is also present after the achieved color image classification. The results of both the initial color image classification experiment and the comparison between the varying adaptations show that the Tsetlin Machine, with a little extra work, can achieve high accuracy color image classification without image augmentation or pre-training.