To go further, the researchers attempted to replicate the performance of humans and baboons with artificial intelligence, using neural network models inspired by basic mathematical ideas about what a neuron does and how neurons are connected. These models—statistical systems fed by high-dimensional vectors, matrices multiplying layers upon layers of numbers—successfully matched the performance of baboons but not that of humans; they failed to replicate the regularity effect. However, when the researchers created a model inflated with symbolic elements – the model was given a list of geometric regularity properties, such as right angles, parallel lines – it faithfully reproduced human performance.
These findings, in turn, pose a challenge to artificial intelligence. “I love the advancements in AI,” Dr. Dehaene said. ” It is very impressive. But I believe there is a deep aspect missing, which is symbol processing,” that is, the ability to manipulate abstract symbols and concepts, as the human brain does. This is the subject of his latest book,How We Learn: Why Brains Learn Better Than Any Machine…Yet.”
Yoshua Bengio, a computer scientist at the University of Montreal, agreed that current AI lacks anything related to symbols or abstract reasoning. Dr. Dehaene’s work, he said, presents “evidence that human brains are using capabilities we don’t yet find in cutting-edge machine learning.”
This is especially the case, he says, when we combine symbols while composing and recomposing knowledge, which helps us to generalize. This discrepancy could explain the limitations of AI – a self-driving car, for example – and the rigidity of the system when faced with environments or scenarios that differ from the training repertoire. And that’s an indication, Dr. Bengio said, of where AI research needs to go.
Dr. Bengio noted that from the 1950s to the 1980s, symbolic processing strategies dominated “good old-fashioned AI”. But these approaches were driven less by a desire to replicate the capabilities of the human brain than by reasoning based on logic (eg, checking a proof of a theorem). Then came statistical AI and the neural network revolution, which began in the 1990s and gained traction in the 2010s. Dr. Bengio was a pioneer of this deep learning method, directly inspired by the network of human brain neurons.
It’s not impossible to do abstract reasoning with neural networks, he said, “it’s just that we don’t know how to do it yet.” Dr. Bengio has a major project aligned with Dr. Dehaene (and other neuroscientists) to investigate how human conscious processing powers could inspire and empower next-generation AI. ultimately our understanding of how brains do it,” Dr. Bengio said.