Tuesday, January 19, 2016

Brain Worm Art: Creating Something Memorable

A lot of factors go into artwork being memorable. It's not something an artist should directly consider when creating, but it's important in gaining notoriety for people to remember seeing your art. I think consumers are more likely to value art if it has a lasting impression on them--even if it isn't the most technically skilled or doesn't spin a good narrative.

So while memorability is a complex, nuanced, subjective system, there are some general rules to it such as famous composition methods. A team at MIT decided to use some modern machine-learning techniques known as convolutional neural networks to teach an algorithm with 60,000 photographs ranked for memorability by humans to teach memorability ranking. They then tested the actual memorability with memorability games and found their algorithm performed on par with humans at measuring memorability. The artificial intelligence, LaMem, is available online and for download. It had all of the heavy-lifting done beforehand in the training, so it can give rank memorability blazing fast and even give a heat-map to define the most memorable areas of an image. Red being most memorable and blue least memorable.

7" Kara Covers

I processed some of Becca's art through LaMem and found they were all highly memorable, between 0.7 and 0.835.

Covers from Becca's children's watercolor comic, 7" Kara. I'm unsure why LaMem thought text would make something more memorable.

Famous Paintings

I was surprised to find many paintings I consider memorable received lower scores, between 0.6 and 0.7, or medium memorability.

M.C. Escher's Relativity. I was sure this would be considered memorable as I personally it has strong staying power. I think a regular person could draw vaguely what it looks like from memory and would certainly recognize art influenced by it.

Edvard Munch's The Scream.
Leonardo da Vinci's Last Supper.
Katsushika Hokusai's The Great Wave off Kanagawa. Finally, something which has a sensible value. Honestly, people probably get this confused with some other Japanese color woodblock cuts, especially in the Hokusai's series, but I still believe it's memorable.

MIT Research Findings

Emotions account for a large variance in memorability. Unsurprisingly, disgust and amusement, and more generally negative emotions, make for some of the most memorable images. The most popular images and consistency of fixation on an image--or the viewer's eyes being drawn to particular location upon first glance--also play a factor.

Astonishingly, aesthetics don't account for any change in memorability. Specifically, 250,000 photographs from the AVA dataset were ranked by hobbyist and professional photographers for their aesthetic allure. They were searching for things such as golden ratio, the rule of thirds, and color harmonies. So maybe design choice isn't correlated with memorability at all; in which case, it would be important to test for memorability when creating promotional material.

Khosla, A., & Torralba, A. (n.d.). Understanding and Predicting Image Memorability at a Large Scale. doi:10.1109/ICCV.2015.27


It would also be interesting to see how placing logos / signatures in work, such as when creating a banner, affect the memorability of it. I imagine after one has taken design into consideration, it would be smart to place one's logo close to a highly memorable area (red hot spot in the images) so there's a higher chance someone will recall your logo with the image when thinking about it. Based on a few simple tests, it seems as though placing a logo in a memorable zone is slightly more effective than placing it in a non-memorable zone. But I have no idea if this extrapolates.

The highest scoring version was actual the final design. Maybe general decisions maximum memorability naturally.
I ran some of Becca's non-cover artwork through as well and found they were still all highly memorable.
Cowgirl painting set. It's interesting to see LaMem focusing on the painted framing and to a lesser extent the girls rather than the landscapes.


There isn't any kind of deeper connection I can draw from this, but scanning through CSAIL's, the MIT research group's, example photos, the lower scored photos look incredibly bland and forgetful and the memorable photos appear professionally shot and composed, but not stock photos.

I would love to hear your thoughts on how memorable you believe your art or artists you like are and see what kind of results you get playing with LaMem. I can't say this will enter Becca's tool-set, but she may use it when deciding upon promotional images after she's narrowed it down to my favorites.

If you would like to learn more about deep neural networks, Joseph Coco, the guest writer, studies them at Vanderbilt University. He was in no way involved with this research. But he may dive deeper into this subject if it draws attention.