Abstract
The goal of zero-shot learning (ZSL) is to train a model to classify samplesof classes that were not seen during training. To address this challengingtask, most ZSL methods relate unseen test classes to seen(training) classes viaa pre-defined set of attributes that can describe all classes in the samesemantic space, so the knowledge learned on the training classes can be adaptedto unseen classes. In this paper, we aim to optimize the attribute space forZSL by training a propagation mechanism to refine the semantic attributes ofeach class based on its neighbors and related classes on a graph of classes. Weshow that the propagated attributes can produce classifiers for zero-shotclasses with significantly improved performance in different ZSL settings. Thegraph of classes is usually free or very cheap to acquire such as WordNet orImageNet classes. When the graph is not provided, given pre-defined semanticembeddings of the classes, we can learn a mechanism to generate the graph in anend-to-end manner along with the propagation mechanism. However, thisgraph-aided technique has not been well-explored in the literature. In thispaper, we introduce the attribute propagation network (APNet), which iscomposed of 1) a graph propagation model generating attribute vector for eachclass and 2) a parameterized nearest neighbor (NN) classifier categorizing animage to the class with the nearest attribute vector to the image's embedding.For better generalization over unseen classes, different from previous methods,we adopt a meta-learning strategy to train the propagation mechanism and thesimilarity metric for the NN classifier on multiple sub-graphs, each associatedwith a classification task over a subset of training classes. In experimentswith two zero-shot learning settings and five benchmark datasets, APNetachieves either compelling performance or new state-of-the-art results.