Learning models using similarity based and one vs previous paradigms

Abstract

Traditionally, when building machine learning models for multi-class classifi cation, it is common practice to build a model consisting of an ensemble of binary classifiers using some learning paradigm which dictates how the binary classifi ers work together to discriminate between the individual classes. As new data comes in and the model needs updating, these models would often need to be retrained from scratch. This work considers three new learning paradigms which provide a way for the trained models to update without the need of retraining the entire system from scratch. Through training class by class, we utilize previous classifi ers for previous classes to build more efficient classifi ers for future classes, which gives the paradigms applications in Lifelong Machine Learning, by avoiding training against the examples of classes which would provide no bene t to our new classifi er's classification performance. The goal is to create models which are reusable and take less time to train while retaining classifi cation performance. Results show that the new paradigms are promising in different scenarios with regards to their goal.