Train to disentanglement MNIST (or CIFAR-10) information from a dataset than combines CIFAR-10 and MNIST images. Experiences :
- Freeze or not the weak encoder during strong encoder training
- Replace the weak and strong-common encoders by a unique encoder
- Try to compute jem loss between strong-specific & weak and/or strong-specific & strong-common
- Train a decoder to retrieve the weak modality (MNIST is this case)
Train to separate specific-VBM information to information common between VBM and sulcus skeletons