Scalable multimodal variational autoencoders with surrogate joint posterior

To obtain a joint representation from multimodal data in variational autoencoders (VAEs), it is important to infer the representation from arbitrary subsets of modalities after learning. A scalable way to achieve this is to aggregate the inferences of each modality as experts. A state-of-the-art approach to learning this aggregation of experts is to encourage all modalities to be reconstructed and cross-generated from arbitrary subsets. However, this learning may be insufficient if cross-generation is difficult. Furthermore, to evaluate its objective function, exponential generation paths concerning the number of modalities are required. To alleviate these problems, we propose to explicitly minimize the divergence between inferences from arbitrary subsets and the surrogate joint posterior that approximates the true joint posterior. We also proposed using a gradient origin network, a deep generative model that learns inferences without using an inference network, thereby reducing the need for additional parameters by introducing the surrogate posterior. We demonstrate that our method performs better than existing scalable multimodal VAEs in inference and generation.