Fig. 4: MODES learns a disentangled latent space.
From: A Representation Fusion Framework for Decoupling Diagnostic Information in Multimodal Learning
a, b Kernel regression on different components of the representations (shared and modality-specific) shows which component is more predictive of a downstream prediction task, such as for (a) clinical phenotypes or (b) diagnostic labels. c Scatter plots of the 2-dimensional projection of the different components of the learned representation shows what features are more separable in each component. d Samples from extreme ends of the first 2 principle components of the modality-specific representations demonstrate what features these subspaces capture.
