philip lelyveld The world of entertainment technology

This AI Learns Continuously From New Experiences—Without Forgetting Its Past

...

Tackling a new task sometimes requires a whole new round of training and learning, which erases what came before and costs millions of dollars. For ChatGPT and other AI tools, this means they become increasingly outdated over time.

This week, Dohare and colleagues found a way to solve the problem. The key is to selectively reset some artificial neurons after a task, but without substantially changing the entire network—a bit like what happens in the brain as we sleep. ...

Called continual back propagation, the strategy is “among the first of a large and fast-growing set of methods” to deal with the continuous learning problem, wrote Drs. Clare Lyle and Razvan Pascanu at Google DeepMind, who were not involved in the study. ...

It still uses back propagation, but with a small difference. A tiny portion of artificial neurons are wiped clean during learning in every cycle. To prevent disrupting whole networks, only artificial neurons that are used less get reset. The upgrade allowed the algorithm to tackle up to 5,000 different image recognition tasks with over 90 percent accuracy throughout. ...

AI networks that can no longer learn could also be due to network interactions that destabilize the way the AI learns. Scientists are still only scratching the surface of the phenomenon.

Meanwhile, for practical uses, when it comes to AIs, “you want them to keep with the times,” said Dohare. ...

“These capabilities are crucial to the development of truly adaptive AI systems that can continue to train indefinitely, responding to changes in the world and learning new skills and abilities,” wrote Lyle and Pascanu.

See the full story here: https://singularityhub.com/2024/08/22/this-ai-learns-continuously-from-new-experiences-without-forgetting-its-past/