Peripheral vision, the ability to see objects outside of our direct line of sight, has been simulated by researchers at MIT to be used with AI technology. Unlike human vision, AI lacks the capability to perceive peripherally. Enhancing AI with this ability could greatly enhance its proactivity in identifying threats, and could even predict if a human would notice an upcoming object.
MIT researchers developed an extensive dataset of images to reproduce peripheral vision in AI models. They discovered that this development improved the AI’s ability to identify objects in the peripheral vision, albeit not to the standards of human abilities. Irrespectively of the size of objects or the amount of visual clutter present, the AI’s performance did not significantly fluctuate.
This could hold the key, according to co-author Vasha DuTell, in developing machine learning models that could be used in a broader range of applications, such as improving safety for drivers or developing displays easier for humans to see.
Peripheral vision modelling could also be used to predict human behaviour more accurately, notes lead author Anne Harrington. Gaining insights into how humans absorb information could help devise new ways of interacting with machines. This is especially useful as many approaches to model peripheral vision in AI simply blur the edges of images, which would not represent the complex information loss that occurs in the human eye.
The researchers used a vast amount of images to train multiple computer vision models and compared its performance to that of humans tasked with detecting objects. In all cases, the abilities of the machines were significantly inferior compared to the humans, particularly in detecting objects in the far peripheral.
This research contributes to our understanding of human vision, demonstrating that it should not be seen as limited because of our finite number of photoreceptors, but rather as a highly optimized process, designed to perform tasks of genuine importance. Current neural network models are unable to replicate this level of function, but with datasets such as those developed by the MIT researchers to replicate peripheral vision, the goal of achieving this is more possible.