For at least thirty years, research groups from all over the world have been working on the development of artificial intelligence tools. Of course, until now, there was at least one premise that was clear: a machine (or in this case, rather, a network of artificial neurons) can never think like the human mind. Well, a study published this Wednesday in the scientific journal ‘Nature’, and which has included the participation of several Spanish researchers, has managed to find a method so that artificial intelligence networks can combine concepts equally (or better) than the humans.
The study, led by the University of New York and Pompeu Fabra in Barcelona, has managed to ‘transfer’ to a machine the human ability to learn new conceptscombine them with other existing ones and, finally, continue developing new ideas based on these. This is the same mechanism that allows, for example, a child to learn to jump and then continue jumping around a room, start bouncing by moving his arms, or simply experiment with different movements of this style.
Until now, these types of skills They were the exclusive heritage of humans. Machines, even in their most sophisticated version, could store practically infinite amounts of information but could not ‘reason’ according to this logic. To overcome this obstacle, researchers Brenden Lake and Marco Baroni developed a technique that combines constant training of artificial neural networks with other related technologies with speech recognition and natural language processing. The result, they explain, is superior to the capacity of tools like ChatGPT.
better than humans
The technique has been named ‘Meta-learning for Compositionality (MLC)’. According to its creators, in the experiments carried out to date it has been shown that by applying it you can emulate the human ability to reason and, in some cases, go even further. “We have shown, for the first time, that a generic neural network can imitate or overcome generalization human systematics in a head-to-head comparison,” said Brenden Lake, an associate professor in the Center for Data Science and the Department of Psychology at New York University.
Related news
Testing to date also shows that these ‘super-trained’ neural networks and humans reason better than AI tools like ChatGPT. In this sense, Marco Baroni, professor at the Department of Translation and Language Sciences at the Pompeu Fabra University, remembers thatThese types of ‘chatbots’ “continue to have difficulties with compositional generalization” although, thanks to this type of progress, everything indicates that these obstacles will soon be overcome.
According to the promoters of this work, this type of machine learning techniques could achieve artificial intelligence tools like ChatGPTamong others, can learn faster, more efficiently and at a lower cost. Until now, training chatbots required dumping huge amounts of information, training them for several months and spending a lot of electricity (and water) in the process. If machines began to ‘reason’ by imitating human logic, It is possible that part of this process can be saved. Or at least that is what, for now, experts propose after the publication of this milestone.