Artificial intelligence (AI) that mimics human behavior is more than just mimicking human behavior. Technology must process information and ‘think’ like humans to be trusted.

The University of Glasgow’s School of Psychology and Neuroscience published new research in the journal patterns. It uses 3D modeling for analysis of how Deep Neural Networks process information. This is part of a more prominent family of machine learning.

This new work is expected to lead to more reliable AI technology that can process information as humans and make predictions that are easy to understand.

AI development faces many challenges. One is understanding how machine thinking works and how it compares to human information processing. This will help ensure accuracy. Deep Neural Networks are frequently cited as the best model of human decision-making behavior. They can even surpass human performance in specific tasks. Even the most simple functions of visual discrimination can show apparent inconsistencies or errors compared to humans.

Deep Neural Network technology can be used for applications like face recognition. Scientists must understand how these networks process information and when errors might occur.

The research team did this new study. They used the visual stimulus the Deep Neural Network was presented to model the problem and transform it in multiple ways. The result was similar recognition via similar processing between the AI model and humans.

Professor Philippe Schyns is the senior author of this study and the Head of the University of Glasgow’s Institute of Neuroscience and Technology. He said that AI models should behave “like” human beings. For example, they must recognize faces when they are seen as humans would. This means that an AI model must use the same information as a human to identify the face. We could make the mistake of believing that the AI works precisely like humans, only to find out later that it isn’t.

Researchers created a series of adjustable 3D faces and asked people to rate how similar these faces were to four familiar faces. This information was used to determine if Deep Neural Networks gave the same ratings for similar reasons. It tested whether AI and humans made the same decisions and whether they were based on the exact same information. The researchers can visualize these results as 3D faces that drive network behavior and humans. For example, a network that correctly classified 2000 identities was caused by a highly caricatured look. It identified faces processing different information than humans and helped it identify the faces.

This research is expected to lead to more reliable AI technology that behaves more like humans and less unpredictable errors.