Scientists from Italy and Japan, working together as a research team, have improved the Daydreaming algorithm, which removes information noise and hallucinations from neural networks. The key innovation is that the algorithm compares not the data itself, but the differences between data points. This approach enables neural networks to ignore information noise and work more accurately with real-world data.
The project was developed for classical Hopfield networks—models that mimic the associative memory of the human brain. Without proper cleaning, false memories accumulate rapidly, reducing performance and leading to the loss of up to 87% of useful memory capacity.
The researchers believe that this upgrade could help create more transparent, reliable, and efficient artificial intelligence systems.