Multitask learning etc. ?

[schinto]

Can Graph ODEs be used for multitask and multilabel learning (in the context of molecular property prediction ; see MoleculeNet.ai) ? Can an attention mechanism be incorporated in Graph ODEs (similar to AttentiveFP) ? How robust are Graph ODEs towards hyperparameter changes? How much training data is required to build robust graph ODEs? What about their interpretability? How is the uncertainty in the prediction of new samples estimated? How good are Graph ODEs on node-level prediction tasks (like 13C NMR prediction; see Expressive Graph Informer Networks) ? Can wide & deep graph ODEs be combined to improve prediction performance? Can graph ODEs be used for low data problems (often occurring in drug discovery)?