New Study Shows Children Outperform AI in Visual Perception

As artificial intelligence (AI) rapidly advances, a new study from Temple University highlights a significant limitation in AI's capabilities. Assistant Professor Vlad Ayzenberg and his team compared the visual perceptual abilities of preschoolers and state-of-the-art AI models, finding that children as young as 3 years old outperform the best computer vision models currently available.

The study, published in Science Advances, demonstrates that the human visual system is far more data efficient than current AI. According to Ayzenberg, "Our findings suggest that the perceptual abilities of even young children are extremely robust, and the human visual system is far more data efficient than current AI."

For the study, 3- to 5-year-olds were tasked with identifying objects from images presented at speeds of 100 milliseconds while their attention was disrupted by factors such as noise. Despite the challenging conditions, the children performed exceptionally well, a result that surprised the researchers. "We thought this task would be really hard for young children because it’s designed for adults," said Ayzenberg.

The study also highlights the potential for cognitive and neural insights from children to improve AI models. Ayzenberg explains, "AI models are useful, but they make mistakes that no human would. They require more training and energy than we do. For instance, training a large language model like ChatGPT has a carbon footprint about 17 times greater than that of a person in one year. If we can understand how young children are able to visually perceive objects, we can then make the models more efficient."

Ayzenberg, who became a faculty member at Temple in July, opened the Vision Learning and Development Lab on Main Campus. In this lab, researchers will use a combination of behavioral, neuroimaging, and computational techniques to understand how the human brain is organized at birth and in early childhood to support the rapid development of cognitive and perceptual abilities. The goal is to eventually create more human-like AI agents based on what they learn from children.

Moving forward, Ayzenberg plans to use functional MRI in awake infants to measure large-scale brain activity while they engage in specific cognitive tasks. "Babies and young children haven’t been in the world very long, but they can do a lot very quickly. We want to understand what neural processes allow them to rapidly develop these sophisticated abilities in the absence of extensive experience," explained Ayzenberg.

He looks forward to conducting this research at Temple. "Temple’s psychology and neuroscience program is strong, especially in the developmental area," he said.