As part of its activity of support to research in AI, the DATAIA Institute organises monthly seminars.
Stanislas Dehaene (Collège de France) leads a session on «Can current neural networks provide a satisfactory model of the human brain?».
Download the TBA
Learn more about
I will argue that, in spite of their major recent successes, both in solving real-world problems and in predicting brain responses in human and non-human primates, most current neural networks are essentially limited. The human brain keeps the upper hand in its ability to (1) learn from a very small number of examples, sometimes a single trial, using Bayesian-style reasoning (“the child as a scientist”); (2) discover compact, abstract, explicit representations of knowledge, in a form which can be shared with others; (3) learn from others and learn with others; (4) learn compositional representations in a “language of thought”. I will describe extremely simple experiments in which the human brain acquires information in a single trial and represents its in a quasi-linguistic form, which is currently hard to capture with present-day neural networks.