Recall Potential

What it does

This KPI assesses the likelihood of an image being remembered and recalled, indicating its potential to leave a lasting impression in the viewer's memory.

Why it matters

In marketing, memorable images are key to building mental availability and driving sales. Consistent use of memorable imagery across various touchpoints enhances cognitive fluency and brand consistency, crucial for long-term consumer engagement.

Try it yourself: Memorability Self-Test (2 minutes)

The easiest way to understand Recall Potential is to experience it.
Take the short self-test and see what people remember (and forget), this is what Recall Potential predicts.

Start the Memorability Self-Test

This is a demo. Your responses are not saved, not stored, and not linked to you, the test is only for demonstrating how Recall Potential works.

How it works

Recall Potential is based on the scientific insight that, despite varied personal experiences, people tend to remember and forget the same images. The KPI utilizes a deep learning model trained on human recall data to predict an image's memorability score, ranging from 0 (easily forgettable) to 100 (highly memorable).

How to achieve great results

The illustration shows sample images with model predictions. Top left image has lowest, bottom right image has highest recall potential.  

Recall potential requires clearly focussed images and is supported by uniqueness, surprise and emotion. Emotional or action-oriented content tends to draw attention and linger in memory more effectively than mundane or cluttered scenes.

Examples with assets:

Here is a list of factors driving recall potential: 

WHAT: 

• People: Images that contain people tend to be very memorable
• Atypical Content: Images that have atypical or unusual content, such as a chair shaped like a hand, tend to have high memorability scores.
• Emotion: Images that evoke emotions like disgust, amusement, and fear are more memorable than those evoking awe or contentment

HOW: 

• Color: Images presented in color are more memorable than those in grayscale.
• Brightness: One study found that a neural network designed to manipulate memorability in images tended to make images brighter.
• Colorfulness: Similar to brightness, the neural network tended to make images more colorful.
• Object Size: The network also increased the size of objects in the image.
• Object Centering: Objects tended to be centered within the image.
• Background: The backgrounds of images were made less cluttered.

AI models used

Memorability models at brainsuite are based on the latest deep learning frameworks used in scientific literature for the purpose of memorability modelling, including convolutional deep learning networks and vision transformers. The images used to collect human ground truth data originate from various sources, including social media platforms, large scale webscrapes (Common Crawl) and marketing materials (e.g. ads).

In total we leverage 1m+ images for pretraining, and 65k+ images for fine-tuning our models on the memorability prediction task. Evaluation of our models follows the scientific best practices. Accuracy versus human ground data is 96%. A chi-square test yields a highly significant relationship (p<0.00001) between the ground truth and predicted categories (low/ medium/ high ad recall).

Science Background

There is substantial evidence indicating that memorability is a stable property of images, meaning that it is intrinsic to the image itself, and not dependent on the viewer. One of the early key studies in this field was conducted by Isola et al. (2011) from MIT, who created a large-scale online memory experiment and found a remarkable consistency in which images were remembered by different individuals. The images that were most memorable for some viewers were also the most memorable for others, suggesting that memorability is a fundamental property of the image itself, rather than a function of the viewer.

Several follow-up studies have consistently shown that memorability scores are highly correlated across random splits of large subject pools (see e.g. Rust & Mehrpour, 2020). Memorability effects have been identified in many visual stimulus types, including faces (Bainbridge et al., 2013), scenes (Isola et al., 2011, 2014), objects (Kramer et al., 2023), words (Xie et al., 2020), infographics (Borkin et al., 2013), artwork (Davis & Bainbridge, 2023), and even dance moves (Ongchoco et al., 2023). Interestingly, despite memorability being highly consistent across observers, people are bad at predicting if an image will be memorable or not. Even more surprisingly, human estimation of image memorability is negatively correlated with actual memorability (see image below). Many images that people rate as not memorable will actually be remembered, and vice versa. 

Left and middle: examples of images where predicted (by humans) and actual memorability (MB) diverge. Right: correlation of predicted (human) and actual memorability of images, which is slightly negative (-0.18). @ Nicole Rust

The four quadrants include correctly identified memorable (top left) and forgettable (bottom right) stimuli, as well as forgettable stimuli that were incorrectly judged as memorable (bottom left) and memorable stimuli that were incorrectly judged as forgettable (top right). The first value by each quadrant corresponds to the mean guessed memorability rating of the four images, followed by the mean true memorability rating @ Revsine & Bainbridge 2023

An additional insight relevant to marketing effectiveness is that memorable images are processed faster in the human brain (Ma et al., 2024).

Modelling Memorability

Further studies have used deep learning models to predict image memorability with high accuracy, even for art images (Trent & Bainbridge, 2023), which further supports the idea that memorability is a function of image aspects. For example, Khosla, Bainbridge, Torralba, and Oliva (2015) built on the earlier MIT study by creating a computational model to predict memorability scores for a wide variety of images. Their model was able to accurately predict the memorability of images, reinforcing the idea that certain image features are reliably associated with memorability.   

Influence of Culture

The consistency of image memorability across species supports the possibility of cross-cultural consistency of image memorability. Indeed, it was recently reported that memorability score for images is consistent across cultures (Jeong, 2023).  

South Korean participants’ recognition accuracy for individual images was compared to that of US participants. Images that Korean participants remembered better were also recognized better by US participants, showing cultural consistency of image memorability. (c) Jeong, 2023

Influence of Demographics

In addition, it has recently been shown that memorability of images remains consistent across individuals at different stages of development (Almog et al., 2023).

Memorability scores per image in adolescent and adult populations. Each dot represents one image. The diagonal line represents
identical memorability scores in adolescents and adults. Points are equally distributed around the line for all categories.  (c) Almog et al. 2023 

In summary, while the viewer's subjective experiences and cultural background can certainly influence what they find memorable, research suggests that there are objective properties within images themselves that make them universally memorable to a wide range of people. Interestingly, it was recently found that this not only applies to images but also human voices (Revsine et al. 2024).

Measuring Memorability

One striking finding in memorability studies is that naively, subjects are bad at predicting how memorable images are: when untrained subjects were asked to predict memorability, their predictions and actual memorability scores were at best weakly correlated (see Rust & Mehrpour, 2020, for review). This implies that we cannot simply ask subjects to rate image memorability in a typical survey. Rather, memorability needs to be measured in behavioral tasks probing novelty versus familiarity of images. Indeed, in memorability science, individual image memorability scores are typically measured from subjects engaged in a visual recognition memory task.

In these memory tasks, subjects typically view one image per trial and report whether that image is novel, or a repeat of an image presented earlier in the sequence. In addition to these “explicit” memory tasks, there are also “incidential” variants where the memory task is introduced as a surprise task after subjects have viewed images under a pretext task (Goetschalckx et al. 2019). Image memorability scores are computed as the subject average performance at remembering a particular image (the hit rate), corrected for the rate of calling novel images familiar (the false alarm rate). The resulting memorability scores are normalized to range from 0 to 1, and they capture the fraction of subjects that will remember seeing an image after first seeing it minutes earlier and after seeing many images since. This approach has been applied to quantify memorability for many images and videos with diverse content. Interestingly, memorability scores maintain their ranks across timescales ranging from minutes to weeks (Goetschalckx et al, 2018).

Sources: 

• Almog, G. et al. (2023), Memoir study: Investigating image memorability across developmental stages. PLOS One, 18(12).  
• Bainbridge WA, et al. (2013) The intrinsic memorability of face photographs. J Exp Psychol Gen 142, 1323–1334 
• Bainbridge WA (2017) The memorability of people: Intrinsic memorability across transformations of a person’s face. J Exp Psychol Learn Mem Cogn 43, 706–716 
• Goetschalckx L, et al. (2018) Image memorability across longer time intervals. Memory 26, 581–588 
• Goetschalckx, L. et al. (2019). Incidental image memorability, Memory, 27(9), 1273-1282. 
• Hagen, T. & Espeseth, T. (2023). Image Memorability Prediction with Vision Transformers. arXiv: 2301.08647. 
• Gygli M., Grabner H., Riemenschneider H. ,Nater F., and Van Gool L. (2013). The interestingness of images. In Proceedings of the IEEE International Conference on Computer Vision, pages 1633–1640, 2013 
• Isola P, et al. (2011) What makes an image memorable? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 145–152 
• Isola P, et al. (2014) What Makes a Photograph Memorable? IEEE Trans Pattern Anal Mach Intell 36, 1469–1482 
• Jaegle, A. et al. (2019). Population response magnitude variation in inferotemporal cortex predicts image memorability. Elife 8, 
• Jeong, S. K. (2023) Cross-cultural consistency of image memorability. Nature Scientific reports, 13. 
• Kramer, M. A., Hebart, M. N., Baker, C. I., & Bainbridge, W. A. (2023). The features underlying the memorability of objects. Science advances, 9(17)
• Ma, A.C., Cameron, A.D. and Wiener, M. (2024). Memorability shapes perceived time (and vice versa). Nature Human Behaviour.
• Ongchoco, J. D. K., Chun, M. M., & Bainbridge, W. A. (2023). What moves us? The intrinsic memorability of dance. Journal of Experimental Psychology: Learning, Memory, and Cognition, 49(6), 889.
• Revsine, C. & Bainbridge, A. (2023). Learning Image Memorability with Feedback-Based Training, bioRxiv 
• Revsine, C., Goldberg, E. and Bainbridge, A. (2024). The Memorability of Voices is Predictable and Consistent across Listeners. bioRXiv
• Rust, N.C. and Mehrpour, V. (2020). Understanding Image Memorability. Trends Cogn Sci., 25(7), 557-568. 
• Trent, D. and Bainbridge WA (2023) Memory for artwork is predictable, PNAS, 150 (28) 
• Xie, W., Bainbridge, W. A., Inati, S. K., Baker, C. I., & Zaghloul, K. A. (2020). Memorability of words in arbitrary verbal associations modulates memory retrieval in the anterior temporal lobe. Nature Human Behaviour, 4, 937–948.