Enhancing Classification in Zero-Shot Learning with the Aid of Perceptron

Since it is costly to obtain labeled instances for each class, and new classes emerge over time, there are no instances in training set for some classes in image classification. These classes are called test classes and to classify them, Zero-Shot Learning (ZSL) was developed. However, ZSL makes use of training classes to classify the test classes, which raises the domain shift problem. To deal with the domain shift problem, a new algorithm called PPG was developed in this study, which has its roots in the perceptron algorithm. PPG is able to update the prototypes of the test classes considering the transfer matrix learned using the training classes. By integrating PPG into the state-of-the-art ZSL methods, a better classification of the test classes was achieved. © 2022 IEEE.