It is has long been known that organic brains, including yours, mine and every other creature’s on this planet, are modular, meaning they have separate sections for dealing with different things.
However, new research published in PLOS Computational Biology has found that this same modular structure could be key to improving artificial intelligence, or more specifically its memory.
Humans and animals gradually forget things over time, which might make remembering what you were doing when you were five a bit tricky, but is a lot better than what happens to the memories of their artificial equivalents.
Artificial neural networks – in effect, artificial brains – suffer from what is known as “catastrophic forgetting”. When they learn something new, they will often completely and quickly overwrite existing memory, which is not an ideal situation if you want artificial intelligence (AI) to learn and retain knowledge.
However, researchers from the Norwegian University of Science and Technology, Pierre & Marie Curie University and the University of Wyoming have found that introducing modules into artificial neural networks cut the rate of catastrophic forgetting considerably.
They came to this conclusion by simulating the evolution of artificial neural networks, and studying the different rates of skill acquisition and memory loss between modular and non-modular artificial brains.
“Building models that incorporate both evolution and learning is critical to understanding the evolution of the animal nervous system,” explained Jean-Baptiste Mouret, from Pierre & Marie Curie University.
Not only did the modular artificial brains forget less, but they also learnt more, something that will be essential to the development of advanced AI.
“The ultimate goal of artificial intelligence research is to produce AI that can learn many different skills and get better at each of them over time, just as humans and animals do,” said Jeff Clune, from the University of Wyoming.
“We must solve the problem of catastrophic forgetting to realise that goal.”
However, he was keen to stress that there is much to do before the research can be beneficial to AI.
“This work is an important step in that direction, but it is just one step in a long journey,” he said.
Now the researchers are planning to significantly scale up their work, to develop far more complex brain models than can learn detailed tasks and skills.