Research
Researchers crack continual learning, letting models learn without forgetting
A new training method allows neural networks to absorb new skills for months without erasing old ones — attacking one of deep learning's oldest unsolved problems.
By Priya Sharma, Research Editor — ZURICH
ZURICH — Researchers at a European institute have demonstrated a training method that lets large neural networks keep learning new skills for months without degrading previously mastered ones, according to a peer-reviewed study published this week — a direct assault on the 'catastrophic forgetting' problem that has shadowed deep learning since the 1980s.
The approach continuously identifies which parameters encode consolidated capabilities and shields them during new learning, while routing fresh experience into deliberately plastic regions of the network — a mechanism the authors compare to memory consolidation during sleep.
In evaluations spanning 400 sequential tasks, the method retained 97 percent of original performance on early tasks — against 61 percent for standard fine-tuning — while matching it on new ones.
Enable JavaScript to read the full story on Neural Daily News.