2023-10-26 13:36:12
Humans have the ability to learn a new concept and, once acquired, understand the different contexts and expressions in which it is used. For example, when a boy or girl learns what “jump” means, she is also able to immediately know what it means to “jump around the room twice” or “jump with your hands up.” This ability is called compositional generalization.
Now, researchers from New York University (NYU) and Pompeu Fabra University (UPF) have just created a pioneering technique that has the potential to develop compositional generalization in computational systems at the same level as in humans or, in in some cases, even at a higher level.
The authors point out that this discovery, which can serve to improve the capabilities of generative artificial intelligence (AI) tools such as ChatGPT, is presented in a article published in Nature.
The technology developed by NYU and UPF researchers comes after nearly four decades of work by the scientific community to develop the capacity for compositional generalization among machines. At the end of the eighties, Jerry Fodor y Zeno Pylyshynphilosophers and cognitive scientists, have already proposed that artificial neural networks – the engines that drive artificial intelligence and machine learning – are not capable of making these connections or compositional generalizations.
Since then, different ways have been developed to encourage this capacity in neural networks and related technologies, but with uneven results. Thus, to this day, the debate on how to achieve this is still alive.
We have shown, for the first time, that a generic neural network can mimic or outperform human systematic generalization in a head-to-head comparison.
Brenden Lake, co-author of the study (New York University)
As explained Brenden Lakeprofessor at the Center for Data Science and the Department of Psychology at NYU and co-author of the work, in the new study “we have shown, for the first time, that a generic neural network can imitate or outperform human systematic generalization in an expensive comparison. to face”.
A pioneering AI training method
The developed technique is called Meta-learning for Compositionality (MLC)which focuses on training neural technological networks to improve the compositional generalization of computational systems through practice.
The creators of the existing systems, including the linguistic models used by generative AI technologies, assumed that the compositional generalization of the technologies would arise from standard training methods, or they had developed special architectures to make the machines acquire these capabilities. On the other hand, the MLC shows that machines can develop compositional generalization skills through the explicit practice of exercises that help them acquire it.
MLC technology is based on an innovative learning procedure in which a neural network is continually updated to improve its skills over a series of phases.
Marco Baroniresearcher, professor and ICREA researcher in Language Sciences at UPF and another co-author of the study, believes that “MLC technology can further improve the compositional abilities of large linguistic models such as ChatGPT, which continue to have problems with compositional generalization, even though they have improved in recent years.
MLC is based on an innovative learning procedure in which a neural network is continuously updated to improve your skills over a series of phases. In one phase, the technology is given a new word and asked to use it in new compositions. For example, you are asked to take the word “jump” and then create new combinations, such as “jump twice” or “jump right twice.” Next, the MLC receives, in a new phase, a different word, and so on, each time improving the compositional abilities of the network.
Experiments with human participants
To test the effectiveness of this technology, Lake and Baroni have carried out a series of experiments with human participants where they were given tasks identical to those performed by the MLC system.
The technology’s performance was as good, and in some cases better, than that of human participants. Both MLC and people also outperformed ChatGPT and GPT-4
Furthermore, instead of learning the meaning of real words—terms that people would already know—they had to learn the meaning of meaningless terms (for example, “zup” and “dax”), defined by researchers, and learn to apply them in different ways.
The MLC’s performance was as good, and in some cases better, than that of human participants. Both MLC and people also outperformed ChatGPT and GPT-4, which, despite showing surprising capabilities in general terms, showed difficulties with this learning task linked to compositional generalization.
Reference:
Lake, B.M., Baroni, M. “Human-like systematic generalization through a meta-learning neural network”. Nature (2023).
Rights: Creative Commons.
#technique #combine #concepts #human #mind