Tabla de Contenido
AI takes a step closer to human-like generalization:
In a groundbreaking development, researchers from New York University and Pompeu Fabra University have unlocked the potential for artificial intelligence (AI) tools to mimic human-like generalization based on what they learn.
The essence of human intelligence lies in the ability to grasp a concept and apply it in different contexts,
A feat that demonstrates the uniqueness of human thought. For decades, scientists and philosophers debated whether artificial neural networks-the core engines behind AI and machine learning-could ever bridge this gap. The prevailing view was often that these networks couldn’t make such connections.
In recent years, however, these limits have been relentlessly explored, and efforts have been made to equip AI with the ability to perform “compositional generalization”. A recent study published in the journal Nature by researchers at New York University (NYU) and Pompeu Fabra University in Barcelona sheds light on one of these crucial advances.
This groundbreaking approach, known as “Meta-learning for Compositionality” (MLC), is already outperforming existing methods in the field. The approach demonstrates AI capabilities that are comparable to, and in some cases superior to, human performance.
MLC focuses on training artificial neural networks, the engine behind cutting-edge technologies such as speech recognition and natural language processing. These networks are meticulously trained to increase their compositional generalization through rigorous practice and continuous improvement.
While some AI architects have speculated that the ability to generalize might emerge spontaneously through standard training methods, others have tried to design specialized architectures to achieve it.
The researchers of the Nature study claim that the MLC shows that explicit practice of these skills enables AI systems to unlock new capabilities. One of the study’s authors, Brenden Lake, a professor in NYU’s Data Science Center and Department of Psychology, articulates the significance: “For 35 years, cognitive scientists, AI experts, linguists, and philosophers have debated whether neural networks can achieve human-like systematic generalization. We have shown for the first time that a generic neural network can mimic or even outperform human systematic generalization in a head-to-head comparison.
To evaluate the effectiveness of the MLC method, Lake and Marco Baroni, a researcher at the Institució Catalana de Recerca i Estudis Avançats and a professor in the Department of Translation and Language Sciences at Pompeu Fabra University, conducted a series of experiments. These experiments involved both human and MLC participants in identical tasks. However, instead of using real-world terms that humans might be familiar with, the participants and the AI had to learn the meanings of terms created by the researchers and apply them as intended.
The results of these experiments showed that the AI’s performance generally matched, and in some cases exceeded, that of the human participants. This novel approach offers profound insights into the ability of AI to understand and generalize concepts more closely to how humans do.
But Teodoro Calonge, a professor at the University of Valladolid, warns that this is just the beginning. It is still too early to predict the full potential of this new method for making connections between concepts. While the initial results are promising, the implications for the field of “explainable AI” and, in particular, for AI in medicine, remain uncertain. This groundbreaking research is an important step toward improving the understanding and learning capabilities of AI, bringing it closer to human thinking.
Read other news: