Zero-shot learning (ZSL) is a recent promising learning approach that is similar to human vision systems. ZSL essentially allows machines to categorize objects without requiring labeled training data. In principle, ZSL proposes a novel recognition model by specifying merely the attributes of the category. Recently, several sophisticated approaches have been introduced to address the challenges regarding this problem. Embarrassingly simple approach to zeroshot learning (ESZSL) is one of the critical of those approaches that basically proposes a simple but efficient linear code solution. However, the performance of the ESZSL model mainly depends on parameter selection. Metaheuristic algorithms are considered as one the most sophisticated computational intelligence paradigms that allows to approximate optimization problems with high success. This paper addresses this problem by adapting leading metaheuristic algorithms to automatically train the parameters of a linear ESZSL model. The model is statistically validated by performing a series of experiments with benchmark datasets.
ÖZSARI, ŞİFA; GÜZEL, MEHMET SERDAR; BOSTANCI, GAZİ ERKAN; and AYDIN, AYHAN
"Adaptation of metaheuristic algorithms to improve training performance of anESZSL model,"
Turkish Journal of Electrical Engineering and Computer Sciences: Vol. 29:
3, Article 30.
Available at: https://journals.tubitak.gov.tr/elektrik/vol29/iss3/30