One way to infuse creativity into your business is to create an AI Powered Creativity Machine.
Given a set of objects (products), this ML engine “imagines” never seen before objects (products), each with several different and several similar characteristics.
Essentially this can be thought of as a type of automated of “Cross Domain Innovation”. Used for centuries, Cross Domain Innovation aims to generate innovative ideas, experiences and values through the convergence of knowledge from different, disparate domains.
An awesome example of this is birds and trains. Japan’s bullet train was inspired by a bird’s beak.
The resemblance is clear.
Japan’s Bullet Train vs the Kingfisher: why recreate the aerodynamics wheel?
So, to implement the above an AI technique known as “disentanglement” is used. The aim is to partially mimic human intuition, imagination and innovation. Disentanglement is not a new idea, and can be used on any type of knowledge and applied to a wide variety of real-world problems.
What exactly is Disentanglement?
Disentanglement is a machine learning model (unsupervised learning) that breaks down (disentangles) any object’s characteristics and encodes them as separate dimensions.
To explain this in non-technical terms, imagine a website developer that needs to add images of people to a site: but not real people so as to avoid any type of claim. So, an AI model is provided with a man’s picture and the expected output is a picture of an imaginary someone who looks similar but is much taller. If the AI Model has learned the ‘height’ dimension independently, then this can be adjusted accordingly to get a picture of a similarly looking but taller person. If instead height and gender were encoded together then that request would result in a picture of a taller woman.
AI “Imagination” Models State of the Art
A University of Southern California (USC) research team led by computer science Professor Laurent Itti and PhD students Yunhao Ge, Sami Abu-El-Haija and Gan Xin, have developed an AI model that imagines an object different from anything else ever seen before.
The 2021 paper
Zero-shot Synthesis with Group-Supervised Learning
(Conference on Learning Representations) details this work.
Other interesting models have been created by
, an artificial intelligence (AI) research laboratory (founded in San Francisco in 2015 by Elon Musk and Sam Altman) conducts AI research with a view to promoting and developing “friendly AI in a way that benefits humanity as a whole”.
Let’s explore some actual and potential application areas of this AI “imagination” category.
Removing AI Bias
Making AI less biased will always be of concern. In a serious incident a few years ago, the search term “gorilla” was blocked by Google’s image recognition service after it starting tagging some non-Caucasian people as “gorillas”.
Similarly, Microsoft and IBM’s face-analysis services performed near perfectly when identifying pictures of white males but not as perfect when analyzing images of black women, reflecting a bias in the training picture set.
One way to address these kinds of issues would be to extract race and gender characteristics altogether from an image recognition model using the disentanglement technique.
Heinz has a large number of products and “varieties”. A Disentanglement model could ingest information about all the products and propose new products, logos and even taglines. These would then be filtered, shortlisted and considered as potential new offerings.
In medicine, disentangling the properties of drugs, for example the core medicinal property as opposed to other properties, and then reassembling them to generate specifications could be used to propose new potential drugs or vaccines.
Soccer (or football in the UK) was born in 1863 when England’s Football Association was formed and simultaneously defined a set of rules for both Soccer and Rugby. Today, AKQA, a digital product agency, has taught an AI Model how to create a new sports game by training it on hundreds of existing ones.
The result is a game called “Speedgate”, imagined and created by the AI. It mixes soccer, rugby and frisbees(!) in a new game that was never thought of or played before. More on this in this
With Speedgate, two opposing teams of six players pass, kick and or throw a ball through 3 gates at either ends of a field, but excluding the middle one (an AI inspired twist).
A gate can be defended by only 1 of the 3 defenders playing. The other 3 players are forwards that try to get the ball through the gates. Unlike American Football, no pushing or bringing a player down is allowed. In addition to creating the game (and all the related rules), the model generated the Speedgate Logo and the Speedgate tagline text (which was the out of the box “Face the ball to be the ball to be above the ball”!)
But, of the games proposed by the AI model, some were clearly non-starters. For example, the model did not take into account the fact that a frisbee should not explode in midair (crazy proposed game), or players should not be dangling from a pole. So, it will always be a team effort: AI and the R&D department.
OpenAI's MuseNet (2019) is a deep neural net that predicts musical notes in music files. It generates songs with up to ten different instruments in ten different styles.
OpenAI's Jukebox (2020) is an open-sourced algorithm that generates music with vocals. Trained on a million samples, the system is given a genre, an artist and a snippet of lyrics from which it outputs song samples.
How about this AI generated country music
Country music Alan Jackson Style?
OpenAI’s DALL-E2 and CLIP
DALL-E2 is a Transformer Model that creates images from textual descriptions. CLIP does the opposite: it creates a description for a given image.
OpenAI’s Generative Pre-Trained Transformer 3 (GPT-3) is an autoregressive language model that uses deep learning to produce human-like text.
The quality of the text generated by GPT-3 is shockingly high, to the point that it is difficult to determine whether it was written by a human or a machine. But the context of what is being written is a different story.
That is where the AI Creativity Machine would step in: to ensure that what is being written by GPT-3 is of true value in terms of the concepts and ideas discussed. More in this
New York Times article
On September 22, 2020, Microsoft licensed "exclusive" use of GPT-3. The public API can still be used freely, but the underlying model is no longer open sourced (:
The Bionic Organization
Martin Reeves, chair of the BCG Henderson Institute, co-authored a book named “The Imagination Machine: how to spark new ideas and create your company’s future”, in which he describes a business that uses AI models similar to the above as “a bionic organization”.
Bionics means the replication of biological systems by mechanical or electronic systems. The term was coined in 1958 by researcher Jack Steele to define the study of biological organisms with a view to solving engineering problems.
The idea was made popular by American science fiction and action television series The Six Million Dollar Man (1973-1978), where a former astronaut, USAF Colonel Steve Austin, portrayed by Lee Majors was rebuilt with superhuman strength, speed and vision due to bionic implants.
Consistent creativity as a core business process is what ultimately gives value to a business by addressing its future in terms of new product offerings.
A company’s Research Department could use a system like the one created at USC to complement researchers’ ideas, filter out the best AI proposals, and ultimately help successfully propel a business forward.
If your company does not have an R&D department, or does not have a way to create,
explore and use advanced AI Models, contact us at
for some out of the box thinking.
If you are a New Product Development professional, you might want to check out