New AI Training Method Mimics Human Sleep to Boost Learning
-
Researchers developed a new AI training method called wake-sleep consolidated learning (WSCL) that mimics how human brains reinforce memories during sleep. This helps AIs retain abilities over time.
-
In the "awake" phase, AIs are trained on new data. In the "sleep" phase, they review new data plus highlights from previous training.
-
There is also a "dreaming" phase where AIs view mashed-up abstract images combining old and new concepts. This helps merge neural pathways and primes unused parts.
-
Testing shows WSCL boosts AI accuracy by 2-12% on image recognition tasks compared to traditional training. It also improves retention of old abilities when learning new ones.
-
Some experts say mimicking human brains may not be the ideal path for AI. An approach inspired by dolphin sleep could allow constant waking operation.