Researchers at MIT have created an artificial intelligence system that can predict how "memorable" a photo is and what parts of it will be best remembered. The "MemNet" algorithm is based on "deep-learning" techniques, a new style of AI that uses pattern recognition and processing inspired by the brain's own networks.
The MemNet system, developed by the school's Computer Science and Artificial Intelligence Lab, was exposed to thousands of photos, each with a score based on how well humans had proved able to remember them. By closely analyzing the visual features of all these, MemNet determined a pattern of what photos were memorable and what objects within them contributed most to that.
When presented with new images unaccompanied by a memorability score, the system predicted that score very nearly as well as humans did. You can test it out with this online tool, which will analyze any photo you upload and estimate its memorability and show a heat map of what items it has determined contributed most to it.
"Understanding memorability can help us make systems to capture the most important information, or, conversely, to store information that humans will most likely forget," said lead study author Aditya Khosla. "It's like having an instant focus group that tells you how likely it is that someone will remember a visual message."
MemNet's creators describe the systems behind it in a paper (PDF) they presented at the International Conference on Computer Vision.