The question of whether machines can perceive the world as humans do has intrigued researchers for decades. A recent development in artificial intelligence (AI) sheds light on this intriguing possibility. Researchers from the University of Osaka have found that vision transformers (ViTs), a type of deep learning model that specializes in image analysis, can develop human-like visual attention patterns without explicit instruction. This finding, published in Neural Networks, represents a significant breakthrough in AI’s autonomous capabilities.
Visual attention is a fundamental process wherein humans and organisms selectively focus on essential elements while ignoring irrelevant information. This ability develops naturally in humans, but replicating it in AI has long been a challenge. The research team used a self-supervised learning method known as DINO (self-distillation with no labels) to train ViTs. Unlike traditional methods that rely on labeled datasets, DINO allows models to organize visual information autonomously.
Upon testing, these self-trained ViTs demonstrated gaze patterns resembling those of humans, especially when analyzing dynamic video clips. They didn’t process visual scenes at random but displayed specialized attention patterns similar to human visual processing. The ViTs focused on human faces, entire figures, and backgrounds, mimicking how humans segment and interpret visual scenes. This was particularly evident when comparing human eye-tracking data with ViTs’ attention patterns.
The study’s implications are profound. According to senior author Shigeru Kitazawa, ViTs weren’t programmed to recognize faces but learned to prioritize such elements because it optimized their information intake. This suggests self-supervised learning can capture intrinsic ways intelligent systems, humans included, learn from their environment.
These findings could lead to AI systems with enhanced human-like visual capabilities, contribute to more human-friendly robotics, and support technologies aiding early childhood development. By aligning AI systems more closely with human perception, this research not only advances machine learning but also offers new insights into human cognition.