Here, the N(µwhat,σ2what) is chosen to be a standard normal distribution.

However, most of these models do not support the other crucial ability ofagenerativemodel: generating imaginary observations bylearning thedensity of theobserveddata. Although thisability toimagine according tothedensity ofthepossible worlds plays a crucial role, e.g., in world models required for planning and model-based reinforcement

Nevertheless, two issues impede the usage of GNNs in real scenarios: (1)insufficient labeled molecules forsupervised training; (2)poorgeneralization capability to new-synthesized molecules.

Many PPS, such as Stan [1] or PyMC3 [2], limit the expressiveness of their language to ensure that the programs in their language always correspond tomodels with static support--i.e.

Akeyrequirement forthe success ofsupervised deep learning isalargelabeled dataset - a condition that is difficult to meet in medical image analysis.

Frobenius norm of M λ ( M) Spectrum of M σ ( M) Singular values of M E ( x) Dirichlet energy computed on x H ( G) Homophily coefficient of the graph G A B Kronecker product between A and B vec( M) V ector obtained stacking columns of M . In this section, we give the details on the numerical results in Section 6 . On the contrary, the graph layers do not use any dropout nor non-linearity. A sketch of the algorithm is reported in fLode . Cora, Citeseer, and Pubmed are already undirected graphs: to these, we added self-loops.