Reverse-Engineering Cognition

We combine Machine Learning and Cognitive Science to train models that can predict the impact of visual assets on audiences, and optimize them. Our in-house models have the best performance in industry and academia.

4M +
Human observations
Higher performance
1.2M +

Step 1
We collect millions of data with cognitive games

We use adaptation of traditional techniques in cognitive science to test humans and their behavior after they are impacted with images and videos.

Step 2
We train Deep-Learning models to reproduce the human cognition patterns

We use the most advanced Computer Vision techniques to build models that reproduce human patterns identified in the data. We learn to see the impact of multiple perceptual and semantic properties.

Step 3
We test against real human behavior in holdout data

We then test our models' predictions against human behavior with new assets. We compare the estimation of our models and the real impact on humans, and the results are the best in the field. This means that given a new asset from a new client, we are able to accurately predict the cognitive reaction of humans to that content.

The result

High Accuracy AI

Mean Absolute Error in Memorability
Accuracy in Saliency (AUC)

Frequently asked questions

Got questions? It makes sense. We're building a new scientific field and the new mainstream tech for optimizing ad creative, so things are getting very different all of a sudden. Let us show you why you'll love it.
Why use AI to do this?

Research shows that human capacity to predict what will be remember is very low: not better than flipping a coin. Supporting human creative judgement with a high-accuracy predictive model can boost the impact of the assets.

How about surveys?

Surveys are what the market resorted to in times when AI was not feasible. However, survey-based studies suffer from several problems: (i) large confidence intervals due to small samples, (ii) they can only be used at the end of the creative process because they take too long and cost too much, (iii) they don’t build better capabilities in time, as AI does with each new test. Memorable's AI offers higher-accuracy testing of all assets, and supports teams in every iteration.

How do you segment by context?

We've done a lot of research on this and the evidence is clear: context matters. A great ad in TV can have very low impact in digital. While demographics don’t shape memorability and first-gaze attention, context does—significantly. Our models differ by temporal context—the typical experience of a user in each media—and spatial context—the elements that surround the ad when it is shown to consumers. You just tell us what type of ad you're trying to build and we'll help you optimize for that media.

Do the models understand complexity?

Our models consider both perceptual and semantic factors. This includes simple things like colors and shapes, and complex ones like objects, scenes, and the relation among them—e.g. their likelihood to co-occur and their emotional charge. Overall, the models consider more than 1500 objects and 330 different actions when they do inference. 

Why do you model an average user?

This is what we do on cognitive impact, since memory and attention predictors are strongly consistent across humans in different demographics and cultures. While viewer A can remember more than viewer B, both are likely to remember things in the same order. This is because these brain operations are more affected by evolutionary traits than by individual human experiences at the scale of a lifetime.

On the other hand, clicks and conversion predictors are not consistent like that and do depend on demographics and culture. For this reason, we train conversion modeles in a client-adjusted way, using the historic data of each client's ads and their performance and generating custom results for each.

I have more questions. Where can I find answers?

We'd love to help. You can contact us at our email or give us a call at 202-817-8195. We will get back to you at the earliest.