“A solution to continual learning is literally a billion-dollar question.”
Learning Gaze
A new study highlights a glaring hole in AI models’ ability to learn new information: turns out, they can’t!
According to the study, conducted by a team of scientists at Canada’s University of Alberta and published this week in the journal Nature, AI algorithms trained via deep learning — in short, AI models like large language models built by finding patterns in heaps of data — fail to work in “continual learning settings,” or when new concepts are introduced to a model’s existing training.
In other words, if you want to teach an existing deep learning model something new, you’ll likely have to retrain it from the ground up — otherwise, according to the research, the artificial neurons in their proverbial minds will sink to a value of zero. This results in a loss of “plasticity,” or their ability to learn at all.
“If you think of it like your brain, then it’ll be like 90 percent of the neurons are dead,” University of Alberta computer scientist and lead study author Shibhansh Dohare told New Scientist. “There’s just not enough left for you to learn.”
And training advanced AI models, as the researchers point out, is a cumbersome and wildly expensive process — making this a major financial obstacle for AI companies, which burn through a ton of cash as it is.
“When the network is a large language model and the data are a substantial portion of the internet,” reads the study, “then each retraining may cost millions of dollars in computation.”
Obstacles Course
This phenomenon of plasticity loss is also a major moat between current AI models and the imagined “artificial general intelligence,” or a theoretical AI that would be considered generally as intelligent as humans. After all, in human terms, this would be like if we had to fully reboot our brains from scratch every time we took a new college course, lest we nuke most of our neurons.
If there’s any bright spot for AI companies? Excitingly, the study authors were able to create an algorithm with the power to randomly revive certain damaged or “dead” AI neurons, which showed some success in countering the plasticity problem.
Still, as it stands, a practical solution is still out of reach.
“A solution to continual learning is literally a billion-dollar question,” Dohare told New Scientist. “A real, comprehensive solution that would allow you to continuously update a model would reduce the cost of training these models significantly.”
More on AI training: When AI Is Trained with AI-Generated Data, It Starts Spouting Gibberish
