While we have known for quite some time how colors, shapes, and distribution deeply affect how we perceive and thus interact with the world, this science had remained on the shelves of academia. Until now. With 41% of marketers publishing visual content 2 to 5 times a week to get us to buy products, optimization for human reception is key to effectiveness. Memorable, a startup incubated in the labs of Harvard and MIT, proposes to master the field of visual impact prediction. We caught up with co-founder and CEO Sebastian Acevedo to learn more :

A little bit about you: What is your background?

Before Memorable, my work was focused on the use of Machine Learning for economic development. I worked on the

portrait of Sebastian Acevedo , co-founder and CEO at Memorable
Sebastian Acevedo , co-founder and
CEO at Memorable

design and execution of machine learning projects in multiple countries, from the Inter-American Development Bank–the largest financial partner in Latin America.

Academically, I come from the social sciences. I have a soft heart for statistics and ML, so I’ve combined all my training in economics and political science with data- and statistics-related courses.

What is Memorable, and what does it do?

Memorable is the new creative co-pilot used by top advertisers to optimize all their ads. We change the game by enabling them to test every single creative iteration and get impact predictions and recommendations on how to improve. All of our tech is based on deep-learning models and cognitive science leveraging millions of data points on human reactions to ads.

Other existing companies offer the ability to quantify the effectiveness of an ad. What is different and better about Memorable?

For starters, we understand how the brain processes visual inputs like no other firm: how it distributes attention, how it stores information, and how it guides the conversion process. Our team continuously collects behavioral responses to visual stimuli in memory and attention games, as well as on the biggest social media and e-commerce platforms. We’ve already collected more than 10M human insights, and the growth rate of our data is very large. The quality of our predictions is our biggest asset and is at the core of our company DNA, with our co-founder, Camilo Fosco, being one of the pioneering researchers in the space.

Memorable can help optimize all assets, in multiple stages of the creative process, with predictions based on real behavior.
Memorable can help optimize all assets, in multiple stages of the creative process, with predictions based on real behavior.

By combining cognitive experiments with real audience reactions on the platforms, we do something quite unique: predicting the impact on both branding and conversion and doing it in a very real way. For instance, we don’t ask customers whether they would buy, we directly observe their behavior online. This is a big contrast with solutions that rely on self-reported measures of things like purchase intent.

Beyond quality and larger applicability, we also differentiate by enabling unlimited tests. We’re really there all along the process–our average clients test 5 iterations of every asset they publish. They do this because it only takes seconds to get scores and recommendations on our platform. So by the time they get to the end, they have optimized all their key creative decisions.

To summarize, we offer high accuracy and complete scalability: Memorable can help optimize all assets, in multiple stages of the creative process, with predictions based on real behavior. 

How is visual content important in cognitive science? How much weight does it carry in your model? And why?

Visual learning has been shown to produce much more memory recall than auditory learning, and that directly translates into the current huge importance of visual content marketing. This was our starting point and the most important part of our products. In addition, we’re now extending to text and audio, so we can help advertisers optimize every piece of their creative.

How do you avoid bias?

We test for bias in several ways. One of them is very typical for assessments of algorithmic bias: checking our levels of error on subsamples of our data to ensure that there’s no systematic error there. 

Beyond that, one of the most important practices is to make sure that our data is representative of the use cases that our platform will have. This is usually accounted for in the performance data we use for training because those data contain reactions from the audiences that are typically targeted by our clients. In addition to this, we are very ambitious in our own data collection through cognitive experiments to make sure that we can remain representative for a large number of audiences. Luckily, most data on cognitive reactions show high consistency across demographics, enabling the models to generalize even better.

Memorable currently predicts five key branding metrics (ad recall, 5 seconds impact, brand association, text saliency, and contextual impact)
Memorable currently predicts five key branding metrics (ad recall, 5 seconds impact, brand association, text saliency, and contextual impact)

Finally, it’s also important to remember that the alternative here is not only biased but has a much harder time measuring and correcting that bias. Many areas of human decision-making suffer from bias that is not measured and is hard to correct at scale. Algorithmic bias can be checked with math and adjusted across all applications while detecting bias in humans and retraining them takes significantly longer, if it’s done at all.

Right now, Memorable quantifies success by measuring memory and recall. What other measurement units are you planning to add?

We are currently predicting 5 key branding metrics (ad recall, 5 seconds impact, brand association, text saliency, and contextual impact) and another 5 for conversion (leads, click-through rate, cost per sale, cost per click, video view rate). In addition to these, we also train custom metrics when our clients require them. 

Beyond the metrics that are given directly on the platform, we use intermediate models to predict elements like evoked emotions, distinctiveness, dynamism, and so on. 

Aren’t you worried about ethical limitations? If your solution is used to manipulate political opinions, for example?  What kind of protection do you have in place?

Camilo and I have discussed this several times. Ultimately, we’re expanding the knowledge frontier on how memory and attention work, and this is a fascinating area with huge potential in almost any visual industry. To that end, we’ve started testing the impact of our models on other settings like education, and we have plans to help development organizations access this technology to boost the effectiveness of their information campaigns.

We also take inspiration from analogous technologies to find the right path here. Can a tool like Grammarly be used to manipulate political opinions? We’re in a similar position: we’re a copilot guiding a user towards minimizing visual “mistakes” and creating an effective piece of visual content. What that user does with that content is outside of our control, but we’re confident that the good actors far outweigh the bad ones. And we make sure that we do our part to maximize good actors.

Could a generative AI equipped with your data create unforgettable campaign images without any human intervention? Is that where we are heading? 

We usually mention 2 big trends in this space. One is a trend toward automation, represented by Dall-E and other analogous technologies. The other one is specific to the field of AI: we believe that the progress over the next 10 years of this technology will necessarily come from a deeper understanding of the brain and its reactions to content. 

This is where Memorable comes in: we are building technology to build and optimize assets, not just aiming for realism but also for higher cognitive impact by design. We don’t just want an asset that looks cool and was made by a machine, we want an asset that was built and shaped to have a higher impact. And that requires our understanding of what creative trends are behind that impact.

Having said that, humans will still need to be in the loop for key parts of this process. We don’t tell the brands what they should stand for or which story to tell. We help them implement their ideas in a way that maximizes impact by recommending paths through the countless creative decisions that need to be made during the ad creation process.

If all companies use your solution, will all campaigns eventually start looking the same?

No. This is because of a key factor that shapes our predictions: context. Context matters. What’s memorable or salient in a given context can change if the context changes. Our models constantly adapt to each platform’s changing contexts. Let’s say today, on Tiktok, X tends to be memorable. If everyone starts using X, something distinctive might become more memorable in that context. We make sure to adapt to that changing landscape.

This is a key differentiating factor from other firms in the space: we don’t provide general rules that would always work–because we think they don’t exist. Quite the contrary, we have found that the same elements can drive a positive impact in one asset and a negative one in the other, depending on the context and what it’ll be combined with.

To give you an example: we have found dynamism to have a strong impact on many metrics of success. However, if you combine that with the wrong elements, you can end up generating cognitive load and driving results down. Every asset is unique in the things it combines, and that’s why we don’t recommend “golden rules” of success to the clients. This is highly valued by users who want to innovate: we provide them data that helps them de-risk their innovations without limiting them to a set of rules.

You might have seen the Coinbase ad in the last Superbowl. Just a bouncing QR code on the screen that would not have been approved by any of the companies typically enforcing design rules. Our platform actually predicted results 30% above the average super bowl ad, and the ad was indeed a success.

A QR code on black backround
The Coinbase ad during the Superbowl

One final note to consider is how this actually facilitates innovation and change. For every process of innovation, testing new ideas quickly and cheaply is key. That’s precisely what we help marketers do: test multiple ideas, even the crazy ones, and see whether they’d work or not without having to spend a single dollar on media or wait a single day for the results from surveys.

What would you like to see Memorable offer that technology can not yet deliver? 

Like in every company, especially startups, focus is a strength. Our current ambition is to be the copilot of every important marketing campaign, enabling teams to optimize their creative decisions with impact data and while they are still designing, not after publication.

In the future, we’d like to use our insights on brain reactions to empower other industries in addition to marketing. This technology has a general-purpose potential, and we’d love to see that happen. We could power a large set of communication software companies and help their users have high-impact material, whether on a class, a corporate communications video, a pitch deck, or the design of anything in the metaverse. 

 

Main photo by : Photo by ian dooley on Unsplash

Author: Paul Melcher

Paul Melcher is the founder of Kaptur and Managing Director of Melcher System, a consultancy for visual technology firms. He is an entrepreneur, advisor, and consultant with a rich background in visual tech, content licensing, business strategy, and technology with more than 20 years experience in developing world-renowned photo-based companies with already two successful exits.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.