How Memorable Is Your Image? Now There's an App for That

Scientists at MIT’s Computer Science and Artificial Intelligence Laboratory have developed an algorithm that predicts the memorability of any picture.

by Nathaniel Ainley

December 18, 2015, 9:25pm

Leonardo da Vinci, The Last Supper

What makes an image memorable? Browsing through CNN’s 25 most iconic photographs, it’s hard to find any universal characteristics. How is it that some images, whether they're paintings or photographs, have stayed relevant through years, decades, even centuries of shifting aesthetic trends and tastes?

Studies have shown that the human brain has a remarkable capacity to quickly process and store vast amounts of visual data. That is to say we can remember a lot of different images and a lot about them. It also happens that, despite unique personal experiences with visual culture, memorable vs. forgettable images have certain properties that are universal to everyone, i.e., that most of us tend to remember or forget the same things. The visual detail in an image of intense action or emotion is easier to recall than that of, say, a pretty landscape.

Using this as a jump-off point, scientists over at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), recently created a web app demo that can analyze any image and predicts its level of memorability. LaMem runs your uploaded image through a deep learning-based algorithm, analyzes it, and then assigns it with a memorability score. The experimenters describe the project as "a novel experimental procedure to objectively measure human memory."

Rembrandt, Self-portrait with Beret and Turned-Up Collar

According to the experiment’s description, LaMem is the “largest annotated image memorability dataset to date (containing 60,000 images from diverse sources.)” The scientists built the LaMem data set through an online crowdsourcing survey where a pool of participants, 27 times larger than the previous dataset, were shown a series of images and asked to identify which pictures they recognized. The results proved that, despite a diverse pool of test subjects, they generally recognized the same images. The data was converted into memorability scores for each of the images used in the experiment and then introduced to the algorithm so it could make its own predictions. The web app also constructs a ‘heatmap’ identifying specific areas of an image that are considered more memorable.