Practical Aspects of Zero-Shot Learning

TL;DR

This paper compares various zero-shot learning methods and proposes meta-classifiers to combine their strengths, aiming to outperform individual algorithms on benchmark datasets.

Contribution

It provides a comprehensive comparison of state-of-the-art zero-shot learning methods and introduces meta-classifiers that enhance performance by combining multiple approaches.

Findings

Meta-classifiers outperform individual zero-shot methods

Certain zero-shot algorithms perform better on specific datasets

Meta-classification improves overall zero-shot learning accuracy

Abstract

One of important areas of machine learning research is zero-shot learning. It is applied when properly labeled training data set is not available. A number of zero-shot algorithms have been proposed and experimented with. However, none of them seems to be the "overall winner". In situations like this, it may be possible to develop a meta-classifier that would combine "best aspects" of individual classifiers and outperform all of them. In this context, the goal of this contribution is twofold. First, multiple state-of-the-art zero-shot learning methods are compared for standard benchmark datasets. Second, multiple meta-classifiers are suggested and experimentally compared (for the same datasets).