From social media videos to pictures taken on our mobile devices, visual content dominates our daily lives and is proliferating at an unprecedented pace. Despite its ubiquity, visual content remains one of the most challenging forms of information to analyze. (Just think about how difficult it is to find specific images on your personal drive or to conduct video searches on the web.) That’s why we’re so excited to partner with Coactive AI—the team that’s bringing structure to unstructured data and helping businesses derive insights from image and video content through the power of machine learning.
Harnessing the power of visual content with machine learning infrastructure
Visual content is a subset of unstructured data, which refers to information that is not arranged according to a pre-set model or schema, and therefore cannot be stored in a traditional relational database. The world’s ability to understand unstructured data is often limited given it is not naturally organized in searchable format and does not adhere to conventional data models. Visual content is uniquely challenging in that:
- Metadata is often lacking or too simple, making data quality a significant time-sink
- Metadata generation can be expensive, slow, and inaccurate
- Embeddings require expertise and maintenance
- Data volumes tend to be large, requiring careful and efficient processing
Like consumers, enterprises are also struggling to leverage their ever-growing visual content stores. According to multiple analysts, unstructured data accounts for 80% of enterprise data. With a large majority of enterprise data being unstructured, this is often a massive, under-tapped resource at companies, especially in industries with a high amount of visual content such as in retail, social media, medical imaging, and autonomous vehicles. Traditionally, enterprises have resorted to two ways to gather insights from visual content: either 1) deploy a significant amount of capital and human resources to build out in-house teams and solutions, or 2) rely on a disjointed multi-step process involving different third-party vendors and services across annotation and labeling, search and query, as well as modeling and embedding.
With the current options being so complex and fragmented, enterprises are trying to address needs that are unmet by the status quo, including speed requirements, cost concerns, scalability issues, as well as quality and coordination challenges. Hearing such feedback from companies inside and outside our portfolio about how the current state is far from ideal, we at Bessemer have been closely tracking breakthroughs in machine learning infrastructure as well as evolutions in the business intelligence and data analytics landscape to identify the next generation of technology that would allow for deeper levels of visual content understanding.
The team bringing AI to unstructured visual content
Through our extensive research of this space, we came to learn about what Cody Coleman’s groundbreaking work on DAWNBench and MLPerf during his PhD in computer science at Stanford, where he was advised by leaders in machine learning, including Matei Zaharia, co-founder, chief technologist, and board member, of Databricks who started the Apache Spark project, and Peter Bailis, founder and CEO of Sisu Data. Inspired by his pioneering PhD thesis, Cody joined forces with his MIT college classmate Will Gaviria Rojas to co-found Coactive AI and solve for the current challenges of visual content, leveraging a data-centric approach to bring structure to unstructured data.
“Smart businesses are failing to use AI on unstructured images. They’re hiring expert data scientists, then asking them to deliver speedy breakthroughs using rusty, unsuitable tools. That’s like asking da Vinci to paint the Mona Lisa with crayons. It’s painful for everyone,” shared Cody. “We want to change this status quo and democratize access to machine learning. Coactive supercharges data practitioners and data-driven teams to unlock insights in unstructured images and video data with unprecedented accuracy, speed, and scalability.”
Having held previous data roles at eBay, Google, Facebook, and Pinterest, Cody and Will know first-hand the painful challenges teams face when analyzing visual content. On top of being world-class technologists, this prior experience has given them an added sense of empathy for Coactive’s customers. This unique combination has influenced their vision of unifying data annotation, search, query, and modeling into one platform that makes MLOps easily accessible to anyone, regardless of background.
Using Coactive, businesses can derive more meaning and insights from visual content in a few simple steps. Enterprises first upload raw images or videos into Coactive’s platform through an API or a secure data lake connection, and the platform will actively embed and index the data with little to no manual supervision. The data is then available to query using Coactive’s fully hosted image search API and SQL interface, which is powered by the platform’s algorithm inspired by the team’s proprietary research on data-centric AI and high performance deep learning. The active learning technique Coactive employs is optimized for raw visual content and allows users to quickly define domain-specific concepts with a few labels in a matter of minutes with minimal manual effort.
When we met with Cody, Will, and the Coactive team, we knew this was a values-driven, once-in-a-generation company building something groundbreaking that had never been attempted before. At Bessemer, we’re looking to partner with companies like Coactive that will define the next century of how we live, work, and do business. Alongside our co-investors at a16z, we are thrilled to co-lead Coactive’s $10.4 million Series A for a total of $14 million in funding, and are excited to partner with Cody, Will, and the Coactive team on this journey.
Want to try out Coactive? Request a demo today!