Doubt can make artificial intelligence more reliable

A research team led by Gergő Orbán at HUN-REN investigates the parallels between neural systems and artificial intelligence, focusing on uncertainty estimation in AI models. Their study introduces the Explaining-Away Variational Autoencoder (EA-VAE), which improves uncertainty assessments in image processing tasks. This model incorporates a computational mechanism similar to biological neural networks, enhancing AI's reliability in critical applications like medical imaging. The findings suggest a dual benefit: advancing AI capabilities while providing insights into biological vision systems. The research was published in the IEEE Transactions on Pattern Analysis and Machine Intelligence.