2023-10-30 07:00:10
Artificial neural networks are getting closer and closer to how the human brain works, according to a new study published in the journal Nature. This advancement might change the way we view artificial intelligence (AI) and its potential to mimic human intelligence.
The study, published in the journal Nature, signals a turning point in a decades-long debate in cognitive science. Historically, some experts believed that the architecture of neural networks was not capable of reproducing key aspects of human thought. However, thanks to specific training, these networks now seem capable of this feat.
Brenden Lake, assistant professor of psychology and data science at New York University (New York University (in English: New York University: NYU,…), explains that their work shows that a crucial aspect of human intelligence can be acquired through a model often criticized for its limitations. In short, these networks can now combine known concepts into new concepts, an ability called “systematic compositionality”.
In tests using an invented language, human volunteers and AI models had to identify underlying grammatical rules from sequences of words. Human participants were successful regarding 80% of the time, making consistent errors. A new method, called “meta-learning for compositionality” (MLC), allowed a network of neurons to surpass this performance.
Paul Smolensky, professor of cognitive sciences at Johns Hopkins, welcomes these advances but emphasizes that the model still has limits in terms of generalization. Strengthening this capacity for compositional generalization therefore represents the next important step of the project (A project is an irreversible commitment with an uncertain, non-reproducible result…).
1698653133
#time #achieves #surpasses #crucial #aspect #human #intelligence