The ability of artificial intelligence to understand concepts has taken a big step forward.

Humans have the ability to learn new concepts and immediately apply them to understand related usage - if children learn "jumping," they understand the meaning of "jumping two rounds in the room" or "jumping with hands raised." But can machines possess this kind of thinking ability? In the late 1980s, philosopher and cognitive scientist Jerry Fodor and Zenon Pylyshyn proposed that artificial neural networks driving artificial intelligence and machine learning cannot perform this type of connection called "compositional generalization." However, scientists have been developing methods for decades to enable neural networks and related technologies to acquire this ability.

Researchers from New York University and Pompeu Fabra University in Spain now report in the journal Nature that they have developed a technique that enhances the ability of tools like ChatGPT to perform compositional generalization.

This technique, known as Meta-learning for Compositionality (MLC), surpasses existing methods in performance and even outperforms human performance in certain cases. MLC focuses on improving the compositional reasoning ability of neural networks (driving ChatGPT and related technologies for speech recognition and natural language processing) through practical training.

Developers of existing systems, including developers of large-scale language models, have been hoping that compositional generalization ability can be developed through standard training methods or by developing architectures for specific purposes to achieve these abilities. In contrast, MLC demonstrates how these systems can acquire new abilities by explicitly practicing these skills, the authors pointed out.

Brenden Lake, director of the Data Science Center and the Program for Mind, Brain, and Machines at New York University and one of the authors of the paper, said, "For the past 35 years, researchers in cognitive science, artificial intelligence, linguistics, and philosophy have been debating whether neural networks can achieve human-like systematic generalization. We have shown for the first time that general-purpose neural networks can imitate or even surpass human systematic generalization abilities in direct comparison."

To explore the possibility of compositional learning in neural networks, researchers created MLC, a novel learning approach that continuously updates neural networks to improve their skills through a series of training tasks. In each task, MLC receives a new word and is asked to apply it in a compositional manner, such as transforming the word "jumping" into new word combinations like "jumping twice" or "jumping around the room." MLC then receives the next task with different words, and so on, improving the network's compositional ability with each iteration.

To test the effectiveness of MLC, Brenden Lake, director of the Program for Mind, Brain, and Machines at New York University, and Marco Baroni, a researcher in the Translation and Language Sciences Department at Pompeu Fabra University, conducted a series of experiments with human participants performing the same tasks as MLC. In addition to learning the meanings of actual words (words already known to humans), participants also had to learn the meanings of meaningless words defined by the researchers (such as "zup" and "dax") and know how to apply them in different ways. MLC performed as well as human participants and even outperformed humans in certain cases. MLC and humans also outperformed ChatGPT and GPT-4, despite the latter's impressive general capabilities, which proved challenging in this learning task.

Marco Baroni, a member of the Computational Linguistics and Language Theory Research Group at Pompeu Fabra University, pointed out, "Large language models like ChatGPT still struggle with compositional generalization, although there have been improvements in recent years. However, we believe that MLC can further enhance the compositional ability of large language models."