Training BigGAN on an ecologically motivated image dataset
Weronika Kłos, Katja Seeliger, Martin N. Hebart, Max Planck Institute for Cognitive and Brain Sciences, Germany; Piero Coronica, Max Planck Computing and Data Facility, Germany
Posters 2 Poster
Pacific Ballroom H-O
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
Generative adversarial networks (GANs) have been gaining in popularity in cognitive computational neuroscience as an exploratory tool for understanding representations in brain data. Yet scientific conclusions based on these GANs depend on the constraints imposed by the training data. One common training dataset consists of the 1,000 image classes of the 2012 ImageNet competition. However, these image categories were originally selected as an engineering challenge, not to reflect natural category distributions. As a result, some categories are overrepresented (e.g. dogs), while other important categories are lacking completely (e.g. faces and human bodies). Thus, the use of such pretrained GANs may lead to biases in scientific conclusions. With ecoset, recently an ecologically motivated training dataset has been introduced. Here we report on our efforts in using this dataset for training BigGAN, a common highly expressive GAN with a continuous latent space. The resulting models fall short of the quantitative performance of the original (proprietary) version of BigGAN, yet produce recognizable and often highly detailed naturalistic images. We release the entire set of weights and category samples of this ecoset-trained BigGAN, in the hope for this model to be useful for research applications.