Deep-learning algorithm predicts how memorable photos are

Researchers at MIT have been working on a deep-learning algorithm that is able to predict the memorability of a photograph at near-human levels. The researchers are working out of the MIT Computer Science and Artificial Intelligence Laboratory and created the image with the goal of turning it into an app that tweaks photos to make them more memorable. The algorithm is dubbed MemNet.

The algorithm creates a heat map that identifies which parts of the image are most memorable. The team says that understanding the memorability of an image can allow them to build systems that capture the most important information or stores the information that humans are prone to forget.

One researcher describes the system as an instant focus group to tell you how likely someone is to remember your message. MemNet could have applications ranging from improving online ads to helping develop teaching resources or assistants to help you remember things.

During development, the team of researchers fed the algorithm tens of thousands of images from different datasets ranging from LaMem and Sun and Places image sets. Each of those images has a memorability score based on a human subjects ability to recall them in online experiments. The MemNet algorithm performed 30% better than human participants in testing.