Abstract
An open problem on the path to artificial intelligence is generalization fromthe known to the unknown, which is instantiated as Generalized Zero-ShotLearning (GZSL) task. In this work, we propose a novel Evolutionary GeneralizedZero-Shot Learning setting, which (i) avoids the domain shift problem ininductive GZSL, and (ii) is more in line with the needs of real-worlddeployments than transductive GZSL. In the proposed setting, a zero-shot modelwith poor initial performance is able to achieve online evolution duringapplication. We elaborate on three challenges of this special task, i.e.,catastrophic forgetting, initial prediction bias, and evolutionary data classbias. Moreover, we propose targeted solutions for each challenge, resulting ina generic method capable of continuing to evolve on a given initial IGZSLmodel. Experiments on three popular GZSL benchmark datasets show that our modelcan learn from the test data stream while other baselines fail.