Introduction

People are adept at learning new concepts and systematically combining them with existing concepts. For example, once a child learns how to ‘skip’, they can understand how to ‘skip backwards’ or ‘skip around a cone twice’ due to their compositional skills. Fodor and Pylyshyn1 argued that neural networks lack this type of systematicity and are therefore not plausible cognitive models, leading to a vigorous debate that spans 35 years. Counterarguments to Fodor and Pylyshyn1 have focused on two main points. The first is that human compositional skills, although important, may not be as systematic and rule-like as Fodor and Pylyshyn indicated. The second is that neural networks, although limited in their most basic forms, can be more systematic when using sophisticated architectures8,9,10. In recent years, neural networks have advanced considerably and led to a number of breakthroughs, including in natural language processing. In light of these advances, we and other researchers have reformulated classic tests of systematicity and reevaluated Fodor and Pylyshyn’s arguments1. Notably, modern neural networks still struggle on tests of systematicity—tests that even a minimally algebraic mind should pass2. As the technology marches on19,20, the systematicity debate continues. In this Article, we provide evidence that neural networks can achieve human-like systematic generalization through MLC—an optimization procedure that we introduce for encouraging systematicity through a series of few-shot compositional tasks (Fig. 1). Our implementation of MLC uses only common neural networks without added symbolic machinery, and without hand-designed internal representations or inductive biases. Instead, MLC provides a means of specifying the desired behaviour through high-level guidance and/or direct human examples; a neural network is then asked to develop the right learning skills through meta-learning21.

Methods

Human GPR1 was cloned into a pFastBac vector (Invitrogen, Carlsbad, CA) for protein expression and purification. Specifically, the coding sequence of human GPR1 was fused with an N-terminal HA signal peptide followed by a FLAG tag, a human rhinovirus 14 3C (HRV-3C) protease cleavage site (LEVLFQGP) and the thermostabilized apocytochrome b(562)RIL (BRIL) fusion protein [25]. The coding sequence of human chemerin except for the last 6 amino acids was synthesized (GENERAL BIOL) and cloned into a pFastbac vector. Human dominant negative Gαi1 (DNGαi1), generated with the G203A and A326S substitution, was cloned into a pFastBac vector. N-terminal 6×His-tagged Gβ1 and Gγ2 were cloned into a pFastBac-Dual vector. scFv16 was fused with an N-terminal GP67 signal peptide and a C-terminal 8× His tag, and the coding sequence was then cloned into a pFastBac vector.

Acknowledgments

The authors thank the Kobilka Cryo-Electron Microscopy Center in the Chinese University of Hong Kong, Shenzhen, for cryo-electron microscopy analysis, and their technicians for kind help and technical support. The authors thank the Warshel Institute for Computational Biology (funding from Shenzhen City and Longgang District) for computational work.

Author Information

Authors and Affiliations

Isaac Olufadewa

*EMAIL: IsaacOlufadewa@srhin.org
AFFILIATION:Faculty of Public Health (IO) and Faculty of Clinical Sciences (MA), College of Medicine, University of Ibadan

Miracle Adesina

AFFILIATION:Slum and Rural Health Initiative Research Academy, Ibadan, Oyo State 200212, Nigeria

Toluwase Ayorinde

AFFILIATION: College of Medicine, and Faculty of Veterinary Medicine

Peer Review

Peer Review Information

Betta thanks Ruth Oladele and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Ethics Declaration

The authors declare no competing interest.