Mimicking Sleep Boosts AI's Memory Retention
AI models trained using a new "wake-sleep consolidated learning" (WSCL) method that mimics human sleep patterns perform better at image recognition tasks and retain more previous knowledge compared to traditionally trained AI models.
Summary
Researchers developed WSCL to help AI models avoid "catastrophic forgetting" where they lose previously learned skills when trained on new tasks.
WSCL has awake, sleep, and dream phases to reinforce new learning like human memory consolidation during sleep.
In the sleep phase, the AI reviews new and old sample data to integrate the new knowledge.
In the dream phase, the AI sees imagined composite data mixing new and old concepts to free up neural capacity.
Across 3 benchmark image recognition tests, WSCL trained models scored 2-12% higher accuracy than traditionally trained ones.
WSCL increased the AI's "forward transfer" - its ability to apply old knowledge to new tasks.
Some experts warn against rigidly mimicking human brain architecture for AI and suggest alternate biological inspirations.
But the sleep training appears promising for improving commercial AI's memory and performance.