AI takes a leap forward in understanding conceptual relationships

For decades, philosophers and cognitive scientists have debated whether AI, powered by artificial neural networks, can achieve this level of cognitive performance. However, recent research has shown promising results, challenging the long-standing belief that machines lack the capacity for compositional thinking.

Scientists at New York University and Spain’s Pompeu Fabra University have made significant progress in advancing the capabilities of artificial intelligence (AI) systems to comprehend and utilise compositional generalisations. Compositional generalisations refer to the ability to learn a new concept and apply it to understand related uses of that concept, a skill humans effortlessly possess.

While the debate surrounding the true extent of AI’s cognitive capabilities persists, this groundbreaking research demonstrates that machines can indeed grasp related concepts after learning only one, marking a significant milestone in the realm of artificial intelligence.

The breakthrough achieved through MLC brings us one step closer to bridging the gap between human cognition and machine learning. As AI continues to evolve, these advancements pave the way for exciting possibilities in various fields, including speech recognition, natural language processing, and beyond.

The researchers have introduced a cutting-edge technique called Meta-learning for Compositionality (MLC), which outperforms existing approaches and rivals human performance in making compositional generalizations. By training neural networks, such as ChatGPT, through practice, MLC enhances their ability to comprehend and apply compositional thinking.

Unlike previous methods that relied on standard training or special-purpose architectures, MLC explicitly focuses on practicing and refining compositional skills, allowing AI systems to unlock new powers. This innovative learning procedure involves continuously updating the neural network’s capabilities over a series of episodes.

During each episode, MLC is presented with a new word and asked to use it compositionally, creating new word combinations and expanding its understanding of the concept. The network then progresses to subsequent episodes with different words, further enhancing its compositional skills.

To evaluate the effectiveness of MLC, the researchers conducted experiments with both human participants and AI systems. Remarkably, MLC performed as well as, and sometimes even better than, human participants in tasks identical to those performed by the AI system. Additionally, MLC outperformed existing language models such as ChatGPT and GPT-4, showcasing its superior compositional learning capabilities.

Brenden Lake, an assistant professor at NYU’s Center for Data Science and Department of Psychology and one of the authors of the study, expressed enthusiasm regarding the findings. “We have shown, for the first time, that a generic neural network can mimic or exceed human systematic generalization in a head-to-head comparison,” Lake said.

Marco Baroni, a researcher at the Catalan Institute for Research and Advanced Studies and professor at Pompeu Fabra University, emphasized the potential of MLC to further enhance the compositional skills of large language models. Baroni said, “Large language models such as ChatGPT still struggle with compositional generalization, though they have gotten better in recent years.”