Implementation of the neural network architecture described in https://arxiv.org/abs/1504.07225.
This architecture is an extension of Multimodal Autoencoders that expects the vector representations of two modalities as inputs and:
- Creates a common representation of the two modalities (fusion)
- Reconstructs one modality given the other (crossmodal mapping)
Demo available in the corrnet.ipynb
notebook.