A new study from the University of Alberta is in the news. The finding revealed a critical limitation in artificial intelligence (AI) models. It highlighted the inability of AI models to effectively learn new information without starting from scratch. It exposed a fundamental flaw in the way the AI systems are currently operating and particularly such models which are trained by using deep learning techniques.
Deep learning involves training algorithms on vast datasets to recognize patterns and thereafter make predictions. However, the models struggle when new information is introduced. The study found that “neurons” of AI models lost ability to function when it was attempted to teach new concepts. It caused a complete loss of “plasticity.”
Lead author Shibhansh Dohare said that the loss of plasticity means they can’t adapt to new information without being retrained entirely. Hence, it is time-consuming and extremely costly too. Training advanced AI models can cost millions of dollars each time they need to be updated.
The inability to adapt presents a significant hurdle for achieving the long-term goal of artificial general intelligence (AGI) where an AI would be as versatile and intelligent as a human.
However, the researchers have also developed an algorithm that can randomly “revive” some of the inactive neurons. It means some promises are there in addressing the plasticity problem.
Finding a practical solution still remains a billion-dollar question and a breakthrough could drastically reduce the costs associated with the training of AI models. Well, AI remains a powerful tool now but with a significant limitation that it can learn but not as flexibly or as efficiently as it is hoped.
The finding was published in the Nature journal.
