2023-10-25 17:00:56
Updated Wednesday, October 25, 2023 – 17:00
Researchers from New York University and Pompeu Fabra have developed a technique to improve the ability of artificial intelligence tools to make generalizations from what they teach it.
An example of compositional generalization: If a person knows the meaning of “hula-hoop”, “juggling” and “skating”, he can understand what it means to do all three togetherMikhail Voitik
The ability of people tograsp a concept and apply it to other uses It is one of the main manifestations of the uniqueness of human intelligence. Integrate a specific teaching and then be able to generalize it for use in different conditions. In the 1980s, a group of philosophers and cognitive scientists pointed out that this human capacity was an insurmountable barrier to artificial intelligence (AI); artificial neural networks – engines that drive AI and machine learning – may not be able to establish these connections.
In recent decades, scientists have developed lines of research to test these limits and incorporate the ability to make “compositional generalizations” to new technologies. This week researchers from New York University (NYU) and Pompeu Fabra University in Barcelona describe one of those lines in the journal Nature. Called Meta-learning for Compositionality (MLC), this approach surpasses existing ones, according to its authors, with results similar to – and, in some cases, superior – to human performance.
The method focuses on the training artificial neural networks, the systems behind many of today’s technologies, for example those related to speech recognition and natural language processing. Neural networks are trained to improve compositional generalization through practice and continuous updating.
Some of the designers of current AI systems had theorized that the ability to generalize could arise spontaneously with standard training methods, while others have attempted to develop specific architectures to acquire these capabilities.
The authors of the study in Nature point out that the MLC demonstrates that the explicit practice of these skills allows these systems to unlock new faculties. “For 35 years, researchers in cognitive science, artificial intelligence, linguistics and philosophy have debated whether neural networks can achieve systematic human-like generalization,” explains Brenden Lake, one of the authors of the work, professor from the Data Science Center and the Department of Psychology at NYU. “We have demonstrated, for the first time, that “A generic neural network can mimic or surpass human systematic generalization in a direct comparison.”
innovative research
To test the effectiveness of the method, Lake and Marco Baroni, researcher at the Instituci Catalana de Recerca i Estudis Avanats and professor at the Department of Translation and Language Sciences at the Pompeu Fabra University, conducted a series of experiments in which human participants and the MLC performed identical tasks. However, instead of having to learn the meaning of real words – terms that humans might already know – the participants and the AI had to learn the meaning of terms created by the researchers and apply them as they had designed them.
The AI’s performance in testing was generally as good as that of human participants and, in some tests, better. AND both groups outperformed ChatGPT and GPT-4. “This is an ingenious article when comparing the prediction capacity of generative neural networks with that of humans,” Teodoro Calonge, professor of the Department of Computer Science at the University of Valladolid, analyzes in statements to the Science Media Center Spain (SMC). . “To do this, they have designed a battery of tests based on psychology on intelligence.”
However, Calonge points out that this is a first work and It is still too early to predict the path that this new method of relating concepts may have. “I can’t say if it is a line of research that is going to offer great advances in the short or medium term,” he clarifies. “Of course, I do not believe that it will answer the questions that are currently being raised in the field of ‘explainability of artificial intelligence’ and, in particular, in the field of artificial intelligence and medicine, where the main point of suspicion”.
#advance #artificial #intelligence #human #begins #generalize #learned