Israeli Researchers Use AI to Decode Human Visual Perception
Translated & summarized from NEWSru Israel by baba
Researchers from Ben-Gurion University have used AI modeling to understand human visual perception, comparing AI predictions to human responses on ambiguous images. The study, published in Nature Reviews Neuroscience, found that the human brain uses both feature classification and a generative approach, unlike many current AI models. This research aims to identify human-like AI systems and improve the study of cognition and AI development.
The story in 5 lines · by baba
- Israeli researchers developed a new method using AI to study human visual perception.
- The study compared AI model predictions with human responses to ambiguous images.
- The human brain uses both feature classification and generative modeling for vision.
- This research could lead to more advanced AI models and better understanding of cognition.
- The findings were published in the journal Nature Reviews Neuroscience.
Scientists from Ben-Gurion University, in collaboration with colleagues, have employed artificial intelligence modeling to better understand the computational processes of the human brain in visual perception. While both humans and modern neural networks can recognize objects, standard tests often fail to distinguish the specific computational pathways used by each. To address this, the researchers developed a series of images designed to elicit conflicting predictions from different AI models, causing them to err. By presenting these ambiguous images to human subjects, the team could then compare the AI models' predictions against actual human responses, identifying which models most accurately mirrored human perception.
One co-author, Tal Golan of Ben-Gurion University, explained the methodology: "We created an image that looked like a 7 to one model and a 3 to another. Then we asked people what they saw. Since in this case, the majority of subjects saw a 7, we can conclude that the model which also sees a 7 more accurately implements human perception."
The study, published in Nature Reviews Neuroscience, found that while rapid feature classification allows vision to quickly grasp basic details like angles and lines for an initial assessment, models relying solely on this approach struggle with ambiguous images. The research suggests the human brain utilizes a generative approach alongside feature classification, constructing a hypothetical model of an object to interpret complex or incomplete visual information more accurately. This dual mechanism enhances efficiency, enabling both rapid recognition and precise interpretation.
The novel testing approach allows for the identification of AI systems that operate most similarly to human perception, paving the way for deeper investigation into computational processes within the living brain. These tests also promise to advance the study of human cognitive abilities and the development of more sophisticated AI models.