any-dimensional

This paper develops a systematic framework for establishing universality of machine learning models that handle inputs of varying dimensions (e.g., graphs with different node counts). It shows that many existing architectures fail to be universal and proposes simple modifications to restore universality.