GAN you train your network

Abstract

Zero-shot classifiers identify unseen classes — classes not seen during training. Specifically, zero-shot models classify attribute information associated with classes (e.g., a zebra has stripes but a lion does not). Lately, the usage of generative adversarial networks (GAN) for zero-shot learning has significantly improved the recognition accuracy of unseen classes by producing visual features on any class. Here, I investigate how similar visual features obtained from images of a class are to the visual features generated by a GAN. I find that, regardless of metric, both sets of visual features are disjointed. I also fine-tune a ResNet so that it produces visual features that are similar to the visual features generated by a GAN — this is novel because all standard approaches do the opposite: they train the GAN to match the output of the model. I conclude that these experiments emphasize the need to establish a standard input pipeline in zero-shot learning because of the mismatch of generated and real features, as well as the variation in features (and subsequent GAN performance) from different implementations of models such as ResNet-101.