Modular meta-learning in abstract graph networks for combinatorial generalization

TL;DR

This paper introduces abstract graph networks combined with modular meta-learning to enable flexible, combinatorial generalization to new tasks, demonstrated by modeling object pushing with minimal training data.

Contribution

It proposes a novel framework that uses graphs as abstractions without fixed node assignments, enhancing modular meta-learning for better generalization.

Findings

Effective modeling of object pushing with minimal data

Flexible generalization to unseen tasks

Novel combination of graph abstractions with meta-learning

Abstract

Modular meta-learning is a new framework that generalizes to unseen datasets by combining a small set of neural modules in different ways. In this work we propose abstract graph networks: using graphs as abstractions of a system's subparts without a fixed assignment of nodes to system subparts, for which we would need supervision. We combine this idea with modular meta-learning to get a flexible framework with combinatorial generalization to new tasks built in. We then use it to model the pushing of arbitrarily shaped objects from little or no training data.