Adversarial Attribute Learning by Exploiting negative correlated attributes

Satoshi Tsutsui, Yanwei Fu, David Crandall

Abstract: A typical method for classifying visual attributes in images to use convolutional neural networks (CNNs) with multi-task learning. However, this approach often suffers from negative transfer, which means that classifiers trained together to classify multiple attributes at a time perform worse than classifiers trained separately. Many multi-task learning techniques attempt to circumvent this issue, but we are interested in negative transfer itself from a different point of view: can we take advantage of negative transfer to improve our classifiers? In this paper, we propose adversarial attribute learning (AAL) where two classifiers compete with each other so that the primary classifier can learn a representation that is invariant to an attribute exhibiting negative transfer. Our experiments on human attribute classification datasets demonstrate that our method can take advantage of this negative relationship.