This paper addresses the long-standing challenge of whether artificial neural networks can achieve human-like systematic generalization, a capacity that involves understanding and producing novel combinations from known components. The authors introduce a meta-learning for compositionality (MLC) approach, demonstrating that neural networks, when optimized for compositional skills using MLC, can achieve human-like systematicity and flexibility in generalization. Human behavioral experiments using an instruction learning paradigm are conducted for comparison, revealing that MLC surpasses both rigid symbolic models and flexible but unsystematic neural networks in achieving both systematicity and flexibility. MLC also improves performance on several systematic generalization benchmarks.

Brenden M. Lake, Marco Baroni