Search Papers | Poster Sessions | All Posters
Poster C103 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Evaluating the Category Representations in Conditional Generative Adversarial Networks via Human Category Learning
Victor Navarro1 (), Christoph Teufel1,2; 1School of Psychology, Cardiff University, 2Cardiff University Brain Research Imaging Centre (CUBRIC)
Generative adversarial networks (GANs) hold great potential as a tool for cognitive scientists, equipping researchers with the ability to generate a theoretically infinite number of complex stimuli. For this purpose, a good understanding of the correspondence between human and GAN representations is critical. In the present work, we evaluated the category representations developed by conditional GANs using human category learning. Specifically, we asked whether humans can learn to categorize class-specific GAN-generated samples, and if so, whether they can generalize that knowledge to real samples. Two groups of participants first learned to categorize either real or GAN-generated histology samples depicting benign or malignant breast cancer. Then all participants were probed for generalization to novel samples from both image sources. Categorization performance, as characterized by sensitivity and bias, showed no reliable differences between groups during training. During generalization, categorization performance with samples matching the image source seen during training was maintained. Most critically, categorization performance generalized across image sources with no loss: participants trained with GAN-generated samples were as sensitive and unbiased in categorizing real samples as those trained with real samples, and vice versa. Our results thus support a close correspondence between how humans and deep networks represent natural categories.
Keywords: generative adversarial networks deep neural networks breast cancer categorization