The author proposes a novel multimodal VAE, MM-VAMP VAE, with a data-dependent mixture-of-experts prior to improve representation learning and generative quality.
