HyperMM: Robust multimodal learning with varying-sized inputs