Hellinger Multimodal Variational Autoencoders

The article introduces HELVAE, a novel multimodal variational autoencoder that utilizes Hellinger pooling for efficient multimodal inference. By leveraging a moment-matching approximation, HELVAE avoids sub-sampling and enhances the expressiveness of latent representations as more modalities are incorporated. It demonstrates superior performance in generative coherence and quality compared to existing multimodal VAE models, making it significant for practitioners focused on improving multimodal generative learning.