Teaching a robot to teach other robots

(Nanowerk News) You are a poker wizard. A friend knows all about French cuisine. Another friend is a Mozart connoisseur.

The three of you get together and share knowledge about your respective expertise. You each go learn something from the other two.

People learn a lot by sharing and exchanging information. Can computers do the same thing as other computers—can robots, in effect, teach other robots how to learn by sharing knowledge?

A research team led by computer science Professor Laurent Itti and one of the Ph.D. student, Yunhao Ge, answers this question in a paper published in the journal Transactions on Machine Learning Research (“Light Learner for Lifelong Learning Shared Knowledge”).

They came up with an unequivocal answer:

Yes.

Their paper describes a new approach to the growing area of ​​machine learning (ML) research known as Lifelong Learning (LL), in which AI agents continue to learn as they face new tasks while retaining knowledge of previous tasks.

Itti and Ge describe in a paper a tool they created, SKILLS (for Lifelong Learning Shared Knowledge), in which AI learns 102 different tasks – for example, categorizing tens of thousands of images of cars by model (Ferrari, Jeep, Cadillac) or flowers by species or chest X-rays by disease.

The AI ​​then shares their knowledge through a decentralized communication network and eventually masters the knowledge of all 102 tasks.

“It’s as if each robot is teaching a class in its specialty, and all the other robots are attentive students,” said Ge. “They share knowledge through a digital network that connects them all, like their own private internet.”

Itti and Ge call their work a new direction in LL research.

Most of LL’s current research, they explain, involves a single AI agent learning a task sequentially – an inherently slow process.

Their SKILL tool involves a series of algorithms that make the process go faster, they say, because agents learn at the same time in parallel. Their research showed that if the 102 agents each learned a task and then shared it, the amount of time taken was reduced by a factor of 101.5 after taking into account the necessary communication and consolidation of knowledge among agents.

“Usually,” explains Itti, “you first collect all the data that the AI ​​wants to learn, and then you train the AI ​​to learn it. But like humans, we try to create AI agents who can continue to learn after they discover new things.”

Increase

Itti believes SKILL, the result of research funded, in part, by the Defense Advanced Research Project Agency (DARPA), is a promising starting point for advances in the field of LL.

No previous research has involved so many natural tasks, say Itti and Ge. And this is just the beginning.

“We believe this research, in the future, can be scaled up to thousands or millions of tasks,” said Itti.

When that happens – in just a few years, Itti predicts – LL could have the ability to transform many aspects of our lives and bring humanity closer to achieving a “truly connected, intelligent and efficient global community.”

For example, in the medical field, different AI systems can specialize in studying different diseases, treatments, patient care techniques and the latest research, said Itti.

After consolidating their knowledge, Itti and Ge explain, this AI can serve as a comprehensive medical assistant, providing doctors with the latest and most accurate information in all fields of medicine.

Or imagine that every smartphone user is a local tour guide in the city he or she visits. Each user takes a photo and provides details about landmarks, shops, products and local cuisine.

Once this data is shared across the network, each user will have an advanced digital tour guide in his pocket.

“In essence,” says Ge, “any profession that requires broad and diverse knowledge or deals with complex systems can benefit significantly from SKILL technology.”

Beyond recognition

The SKILL tool tests an AI’s ability to simply recognize what’s in an image, notes Ge.

“Confession is a good starting point,” he says. “But future research will look at AI being used to perform more sophisticated tasks.”

Itti and Ge say crowdsourcing concepts – for example, online restaurant reviews – are comparable to the ideas described in their paper.

“In crowdsourcing,” says Itti, “a lot of people are tackling a problem and when knowledge is shared you have a solution. Now we can do the same with AI agents.”

“What if you, as one person, had to relearn all human knowledge?” Itti added. “That would be an insurmountable task. Humans have a means of sharing information. We are now pushing that idea into the AI ​​domain.”