Condé Nast is the parent company to over 20 media brands, including Wired, Vogue, Vanity Fair, and Condé Nast Traveler. It’s hard to imagine the extent and richness of a library that combines all the best of fashion, travel, and other media on the planet, let alone the task of managing it. However, one of Labelbox’s customers is doing just that.
We spoke with Paul Fryzel, who is directly responsible for the ongoing research, design, and implementation of Condé Nast’s international media platform, Formation.
Paul oversees the Formation Research (FORE) team, who leverage machine learning, natural language processing, data science, and engineering methods with a focus on content structure and user understanding.
“One of our researchers significantly improved the performance of one of her models by updating and better managing her dataset. It was then we realized that we needed to put more resources to making that happen systematically.”
Due to Condé Nast being a large enterprise, with many brands and distributed teams of varying technical focus, getting more people to interact with and improve the data being used to train FORE’s models without a specialized tool “seemed like a disaster.” Paul explained, “if you’re a solo researcher or working with a small team, you may be okay with local tools. However, if you’re managing many different users or teams, you really need a centralized system.”
Like many other teams, they considered building their own tool, but realized quickly that they were going to spend a lot of time building something that wouldn’t have all of the features they needed: “This isn’t really what my researchers are here to do. Most of these members are experienced in designing new machine learning models, datasets, and experiments, but not as much in client-side tooling and visual interfaces.”
Even if the team could produce a working, albeit basic application, this would still struggle to meet Condé Nast’s goals. “Depthful data is our bread and butter as a company.” Expert knowledge is a key component of their brand experience, where the difference between a “shirt, blouse, and tunic” and hundreds of other subtly different aesthetic categories, are concepts that their models need to understand.
“Incorporating Labelbox into our process was a huge step forward for our team.”
At this point, Paul started looking into tools for his team, and found Labelbox. He reached out and was impressed by the team and their technical understanding of the job he was trying to do.
“Hands down, the team at Labelbox was the biggest selling point. Our relationship expedited when I started talking to Manu, Dan, and Brian and realized that we were already on the same page. I can talk to them on a technical and tactical level. The people on my team talking to them do have domain expertise, and realized that the Labelbox team knew more about this solution. Incorporating Labelbox into our process was a huge step forward for our team.”
Paul gave examples of how Labelbox is improving Condé Nast’s datasets and models, such as their work in fashion object detection for brands like Vogue and GQ. “The dashboard is one of my favorite parts about Labelbox. We used it for our runway detection tool [for Vogue]. As we started labelling, we saw that there were 10 times as many shoes as compared to dresses. We were able to adjust in the moment to assure that we would have more reasonable distribution within our labels. It gives us a real-time view of how we are doing, and that tight feedback loop is what will help us build and maintain high-quality datasets.”
Using Labelbox gave Paul’s team a different view of what was possible. “We went from having a static idea of our data and models, to an real-time, interactive view where we can consistently create checkpoints and reevaluate.”
“I’m very excited to see the roadmap of how Labelbox will grow. Teams of experts like mine will be working with this tool on the front line of machine learning technology, and interfacing with the experienced team at Labelbox will make great things possible.”
And it was a good time: “Using Labelbox was fun for my team. It became a bit of a game to see how much could be labeled and how our data affected the performance of the models.”“I’m very excited to see the roadmap of how Labelbox will grow. Teams of experts like mine will be working with this tool on the front line of machine learning technology, and interfacing with the experienced team at Labelbox will make great things possible.”