We live in an increasingly visual world. We are bombarded with literally thousands of images every day. So, why does the memorability of any one image matter?
With buying cycles as long as 12 months, making your images memorable may impact on the bottom line. If a person forgets your website, yet remembers your competitor’s, you’ve lost a potential customer.
When we talk about measuring an image’s memorability, we’re really talking about measuring people. We might say that an image is 20% more memorable, but we mean a person will be able to recall the image for 20% longer.
Now we know why we need to use the most memorable images possible, how can we identify which images will be memorable?
We can only guess. But computers can find the answer.
“Computers are really good at predicting what people remember, but people are really bad at it,” says Aditya Khosla, of MIT. “When we ask people on crowdsourcing platforms which images are memorable, and then test them later, their answers are very inaccurate. The answers they give are like random numbers. For reliability we need to use machines.”
Aditya and his team at the Massachusetts Institute of Technology, Akhil Raju, Antonio Torralba and Aude Oliva, have done just that. Their tool, LaMem (Large-scale Image Memorability), can predict with better-than-human accuracy which images will be remembered, and which will be forgotten.
Plug an image into LaMem, and it will tell you how memorable it is – and will even overlay a “heatmap” highlighting which parts of the image draw the eye and stick in the mind.
The Science of Memorability
LaMem was trained using a process called machine learning. Much as you’d expect, this is the process of training the machine to recognise concepts and similarities. Aditya gives the example of teaching LaMem about a coffee cup.
“If I were to teach you the concept of what a coffee mug is, I would show you a bunch of mugs. They all look different – they might be different colours or be different sizes, but what makes a mug distinct is that it is a vessel that can hold water or liquid. This concept can be explained and is distinguishable, and these important qualities are what the machine will learn.”
LaMem’s training images were sourced from the most popular images on Flickr, the photo sharing site, to give a real-world example of what humans think makes a “good” image. This was backed up with data from real people to create a program which can accurately judge which images will be recalled.
Aditya was inspired by his mentor’s interest in the subject and a vision of a better world: “I’m interested in making everyone smarter by changing the way content is presented.”
Imagine slashing the time it takes to learn a language; taking the pain out of the rote memorisation required of students, or ensuring people in high-pressure roles really remember how to do their jobs. It’s clear that making our memories work better could improve how we live our lives.
But how does this apply to communications?
As well as examining how memorable each image was, we also examined the representation of men and women in images, and the effect which overlaying text has on the memorability. This latter case is very common on websites –check out our homepage for examples – and as we found it doesn’t always work.
We examined websites from around the world. While we were expecting to discover cultural biases, Aditya said that wasn’t the case:
“Memorability is universal.”
“Memorability is universal. We all have different experiences and come from different backgrounds, but the things we tend to remember or forget are very similar. There’s something in our brains, the way we’re inherently wired, which makes some images more memorable than