Today, we are thrilled to announce Madrona’s investment in Visual Layer, a managed service for curating large-scale image sets to produce higher-quality computer vision models. Madrona co-led the $7M Series Seed together with Insight Partners. Visual Layer is built on top of an open-source package called Fastdup, which is already seeing a growing community of 200,000 early adopters. That community includes enterprises such as Meesho, an Indian social commerce platform where 13 million resellers transact.
Generative AI models are trained on immense volumes of data; the computer vision ones are routinely trained with billions of images. But the quality of AI models is only as good as the quality of the data on which they are trained. Broken or missing images, mislabeled images, and duplicates are all quality problems that can skew AI models. Visual Layer has found that up to 30% of these massive image and video collections, amounting to hundreds of millions of assets, end up reducing the quality of these models instead of adding to them. That wastes scarce talent, compute, and budgets and delays the application and development of generative AI.
This is where Visual Layer comes in. Until today, many engineers and scientists simply had to ignore visual data quality or apply brute force techniques such as having a scientist manually inspect images and videos. Visual Layer is a managed service that automatically curates large-scale sets of images, including quality automation to correct labels, remove duplicates, find anomalies, etc. This can include identifying image labels that are wrong or confusing in order to correct/drop those examples (e.g., this image is labeled “dog” but is likely a “cat”). It can also include finding the most distinct subset of a larger collection to maximize training efficiency. Visual Layer enables scientists and ML practitioners to produce higher-quality models and results.
We see Visual Layer as part of a broader trend by customers to demand a higher quality of data, not just quantity. As Andrew Ng wrote, “focusing on the quality of data fueling AI systems will help unlock its full power.” That theory was reinforced by the immediate excitement we saw around this team’s Fastdup open-source project.
Visual Layer was founded by Danny Bickson, Amir Alush, and Carlos Guestrin. Danny and Carlos are second-time Madrona founders (they were part of Turi, which Apple acquired in 2016). We’re thrilled to welcome them back to the Madrona family and to welcome Amir. This is clearly the right team to take on this ambitious mission. They have each led sophisticated computer vision teams at Apple and Brodmann17, where they dealt daily with these data quality problems and attempted to neutralize them with advanced science. However, they discovered that almost no amount of science could make up for the underlying problems in the data and recognized this important problem demanded a more sophisticated approach.
Madrona has backed AI founders working to make the future happen faster for over a decade. We backed the Allen Institute for AI (AI2) in 2012 to tackle some of the world’s toughest challenges with AI, producing a deep bench of AI talent and companies. Over the years, we have backed companies like OctoML, Turi (acquired by Apple), Lattice (acquired by Apple), Algorithmia (acquired by Datarobot), and Xnor (acquired by Apple). Some of our recent investments in this area include Runway, Numbers Station, and Fixie.
We could not be more excited to work with Danny, Amir, Carlos, and the rest of the Visual Layer team. We’re looking forward to the journey to improve the quality of large-scale image and video datasets, thereby improving models and their results.