Scientists from Osaka University have discovered the ability of generative artificial intelligence models Vision Transformers (ViT) to spontaneously develop visual processing mechanisms similar to humans.
According to a new study, the right training method allows artificial intelligence to independently recreate visual processing mechanisms similar to humans. Scientists compared human eye tracking data and visual processing models generated by ViT. The artificial intelligence models were trained using a special DINO method without using fixed filters for image analysis.
After training, ViT demonstrated visual information processing close to how adults watch video clips. And models that were trained using fixed filters and algorithms showed unnatural visual processing.
A thorough analysis confirmed that the AI capabilities that brought visual processing closer to humans arose naturally as a result of DINO training.
Recall that
the AI was offered to be threatened so that it would work better.
Only registered users can leave comments