Despite the rapid advances in artificial intelligence in recent years, the humble human brain still has the edge over computers in its ability to transfer skills and learn across tasks. A new study reveals how we likely do this.
Led by a team from Princeton University, the researchers behind the new study didn't actually run tests on humans, but instead used animals that are very close to us in terms of biology and brain function: rhesus macaques (Macaca mulatta).
These monkeys were asked to identify shapes and colors on a screen, and to look in particular directions to give their answers. While this was happening, brain scans were used to check for overlapping patterns and shared areas of activity in the animals' brains.
Related: Computers Made From Human Brain Tissue Are Coming. Are We Prepared?
Those scans showed the monkey brains using different blocks of neurons – 'cognitive Legos', in the words of the researchers – across tasks. Existing blocks can be repurposed and recombined across new tasks, showing a neural flexibility that even the best AI models can't compete with.
"State-of-the-art AI models can reach human, or even super-human, performance on individual tasks," says neuroscientist Tim Buschman, from Princeton University. "But they struggle to learn and perform many different tasks."
"We found that the brain is flexible because it can reuse components of cognition in many different tasks. By snapping together these 'cognitive Legos', the brain is able to build new tasks."
As you can see in the video below, the animals had to discriminate between shapes and colors in three separate but related tasks that required the animals to continually learn and apply what they knew from one task to the next.
The cognitive Lego blocks the researchers identified were concentrated in the brain's prefrontal cortex. The region is linked to higher cognition – solving problems, planning, making decisions – and seems to play an important role in cognitive flexibility.
The researchers also found that when certain cognitive blocks weren't needed, activity in them was reduced, suggesting the brain can file away the neural Legos that it doesn't immediately need to better focus on the task at hand.
"I think about a cognitive block like a function in a computer program," says Buschman.
"One set of neurons might discriminate color, and its output can be mapped onto another function that drives an action. That organization allows the brain to perform a task by sequentially performing each component of that task."
It explains how monkeys and possibly humans can adapt to challenges and tasks they haven't seen before, and use existing knowledge to tackle them – something that artificial intelligence in its current form struggles with.
Further down the line, the researchers suggest their findings could help train AIs to be more adaptable to new tasks. Their work could also be useful in developing treatments for neurological and psychiatric disorders where people struggle to apply skills to new settings.
For now, these cognitive Legos show, on a fundamental level, how our brains are more flexible and adaptable than AI models, which exhibit so-called catastrophic forgetting: a weakness that means neural networks can't learn consecutive tasks without forgetting how to perform the last one they trained on.
While task-switching isn't exactly great for our brains, applying what we know from one task to another can be a useful shortcut.
"If, as suggested by our results, the brain can reuse representations and computations across tasks, then this could allow one to rapidly adapt to changes in the environment, either by learning the appropriate task representation through reward feedback or by recalling it from long-term memory," the researchers conclude.
The research has been published in Nature.