New Study Led by UW Shows Modular Computer Brains Are Less Forgetful

A University of Wyoming researcher directed a study that suggests that, when brains are organized into modules, they are better at learning new information without forgetting prior knowledge.

The work not only sheds light on the evolution of intelligence in natural animals, but also will accelerate attempts to create artificial intelligence. The group used simulations of evolving computational brain models, called artificial neural networks, to show that more modular brains learn more and forget less, according to a Public Library of Science (PLOS) Computational Biology media release.

The brains of animals, including humans, are modular, which means they have many separate units, such as those for language and facial recognition. While natural animals tend to forget gradually, artificial neural networks currently exhibit what is called “catastrophic forgetting,” according to the PLOS release. That’s because the artificial neural networks rapidly overwrite previously acquired knowledge when learning a new skill.

“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,” says Jeff Clune, a University of Wyoming assistant professor in the Department of Computer Science[BROKEN LINK] who directed the study. “We must solve the problem of ‘catastrophic forgetting’ (in AI) to realize that goal. This work is an important step in that direction, but is just one step in a long journey.”

The findings are published in a paper, titled “Neural Modularity Helps Organisms Evolve to Learn New Skills Without Forgetting Old Skills,” in today’s (April 2) issue of PLOS Computational Biology. The peer-reviewed, open-access journal features works of significance that further the understanding of living systems -- at all scales -- through the application of computational methods.

Kai Olaf Ellefson, a then-doctoral student at the Norwegian University of Science and Technology who visited Clune’s Evolving Artificial Laboratory at UW, led the study. He is now a postdoctoral researcher at Brazil Institute of Robotics. Jean-Baptiste Mouret, an assistant professor of artificial intelligence and robotics at Pierre & Marie Curie University in Paris, France, and a longtime collaborator of Clune’s, also contributed to the report.

Animals, including humans, have the ability to learn many different skills, because learning something new does not cause the other skills to be lost.

“If I learn how to ride a bike, that does not cause me to forget how to play chess,” Clune says.

The opposite is true for the current computer brain models that underpin artificial intelligence, he says. Artificial neural networks, as these computer brains are called, can learn one skill. But, when they start learning a second skill, they do so by overwriting everything they learned about the first skill.

In the future, Clune and his colleagues plan to dramatically scale up the complexity of the artificial brain models they work with and the difficulty of the tasks they ask the neural networks to learn.

Clune explains that we ultimately want robots to be able to perform multiple tasks, not just one. For example, he says a home robot should be able to do the laundry, vacuum, wash the dishes and take out the trash.

“We also will want robots to perform the variety of tasks associated with caring for the elderly or putting out forest fires,” Clune says. “In all cases, we need robots that can learn many different skills. But, that is challenging for current artificial intelligence learning algorithms.

“Robots, currently, are mostly one-trick ponies,” Clune continues. “They forget everything they just learned when they go on to learn a new task. We show, in this paper, that encouraging neural networks to be modular is one way to help robots learn multiple, different skills.”