Element AI and SSENSE: strengthening the AI commons
Philippe Beaudoin Philippe Beaudoin
October 2 4 min

Element AI and SSENSE: strengthening the AI commons

Early this month at ECCV ‘18 in Munich, we had the great pleasure to announce the winners of our inaugural FashionGen Challenge, a unique text-to-image competition we developed in collaboration with global fashion platform SSENSE.

We’ve now made the dataset from the competition public. Work on the dataset and the competition was led by Element AI fundamental research scientist Negar Rostamzadeh.

Seventeen teams from around the globe took a crack at our release of the SSENSE FashionGen, the world’s first industry-grade text-to-image dataset, and have invented beautiful new methods in the process. While the release of FashionGen is well-worth celebrating on its own, for us at Element AI the ECCV ‘18 challenge also celebrates our first experiment with a new form of collaboration between enterprise, industry, and the AI research community. Working alongside SSENSE to build Fashion-Gen out of their world-class collection of professionally annotated fashion photographs, it was our privilege to guide them through the process of a new kind of investment: cultivating the AI commons.

Cultivating the AI Commons

The guiding principle behind our work on FashionGen is that the more you give the AI commons – the ecology of open-access datasets, models, methods, and researchers that acts as the native environment of fundamental progress in AI – the more the AI commons give you back. What this means isn’t just that everyone can benefit from cultivating the shared garden of the AI commons, but that the companies that make a lasting contribution to the AI commons stand to reap the greatest benefit from progress in AI.

In other words, breakthroughs within fundamental AI research will indeed greatly benefit all players in the long run, but in the short run the first commercial fruits go to the players that achieved the tightest integration with the AI commons.

Importantly, the reason for this “first come, first served” principle in AI progress isn’t simply economic, cultural, or legal, but comes from the nature of the scientific field itself. Cutting-edge neural network methods, despite all their many virtues, are exceptionally temperamental creatures. It can easily take a team of experts months of challenging work to get models developed in the AI commons to work properly in new domains of application.

The big lag between performance in “pure” and “applied” AI is, in this sense, not necessarily a function of the greater difficulty of applied domains, but a function of their current disconnect from the collective power of researchers in the AI commons.

Investing in the AI Commons

Part of our mission here at Element AI, both from a scientific point of view and from a business point of view, is to teach enterprises looking to create a deeper bond between their work and fundamental research in AI how to invest in the AI commons. This is a new form of investment that money quite literally can’t buy, since it depends on genuine participation in the scientific process.

By carefully devising FashionGen to meet the most cutting-edge needs of fundamental computer vision researchers and bringing it into the AI commons, SSENSE and Element AI invite the future of AI research to take place in SSENSE’s own backyard. The goal, for SSENSE, is to facilitate more breakthroughs in the AI research community and continue to nurture awareness around the benefits derived from open access to quality data.

A Step in the Right Direction for Industry-Ready AI Applications

While the AI research community is likely several years away from cracking industry-grade text-to-image synthesis on FashionGen, one of the beauties of AI research is that progress on “futuristic” applications easily spins-off into advanced, industry-ready applications. The cutting edge models that emerge as the AI research community continues to push the state of the art on FashionGen are going to provide a natural foundation for powerful tools in areas like reverse image search, recommendation systems, attribute tagging, predictive pricing and provisioning.

As for ourselves, we’re looking forward to continue our work with SSENSE by adapting fundamental computer vision models into high performance machine-learning tools and innovative UX features. And while we’re not sure when we’ll see a mature text-to-image application for fashion-design professionals, SSENSE has a good head start on getting there first.