Researchers at MIT have developed an image dataset that simulates peripheral vision for use in training machine learning (ML) models, an area where artificial intelligence (AI) notably diverges from human ability. Humans leverage less-detailed peripheral vision to detect shapes and items outside their direct line of sight, an ability AI lacks. Incorporating aspects of peripheral vision into AI may improve hazard prediction and enhance pattern prediction of human behaviour, amongst other applications.
The researchers discovered that integrating this dataset into their ML models bolstered these models’ peripheral object detection capabilities, although their empowered machines still fell short of human performance levels. Unlike human peripheral vision, neither the size of objects nor the level of visual clutter in an image impacted AI’s efficacy in this experimental context.
Exploration of this discrepancy could provide valuable insights for future research, enhancing machine learning models to be more human-like in their operation. The research also highlighted potential applications: superior driver safety, generation of more user-friendly displays, and enhanced ability to forecast human behaviour.
However, when tasked with identifying a target object in the periphery of blurred image pairs, AI models struggled, suggesting a divergence in the strategies employed by AI and humans for object detection tasks. Despite these challenges, MIT’s research raised valuable questions and provided a fresh image dataset for future testing. The publicly accessible dataset is expected to catalyze further studies and experimentation in the field of computer vision.
This research by MIT offers critical contributions to understanding human vision and its application to AI. It emphasizes the need for further AI research, aided by their image database, to learn from the neuroscience of human vision and improve AI’s performance in tasks involving peripheral vision. The project received support from the Toyota Research Institute and the MIT CSAIL METEOR Fellowship.
