|
Abstract:
|
In the past decade, variational autoencoders (VAEs), generative adversarial networks (GANs) and their variations have become popular techniques for generating new images. Powered by the availability of large number of training samples, VAEs and GANs have achieved amazing successes in many applications. However, their performance under smaller sample size, as well as the associated adjustment in batch size, learning rates, epochs, etc. has not been systematically studied. Moreover, for a given sample size, how the complexity of images impacts the performance of VAEs or GANs remains unknown. Such information is very important for medical applications as the training sample size tends to be limited and complexity of images also vary in many medical studies. In this work, we’ll conduct an extensive simulation study to evaluate the performance of VAEs and GANs in different settings, with the focus on the effect of sample size and image complexity. Some practical recommendations will be made based on results from the simulation study.
|
