“Feature stores,” with a dry, invisible moniker, they may not seem like the most sexy topic.
But it is an important part of the AI systems that businesses – and consumers, for that matter – use every day. That is why they are attracting an increasing amount of attention and investment from creative companies, who see the market as a growing opportunity in the long run.
AI systems are made up of many components, one of which is features. Attributes are individual variables that act as system components. When thinking about attributes, it can be useful to draw a table, where the data used by AI methods are organized into rows of examples (data that the system learns to predict) and attributes columns (data describing those examples). Symbols are attributes used to describe each instance – the AI spam detector may use attributes such as words in the body of the email, for example, or the contact information of the sender.
Working with features seems to be an ad hoc process within a single AI system. But in terms of business scales, where data science is responsible for protecting dozens of thousands of systems, a place to manage and monitor features becomes inevitable.
Enter the attribute store, central repository of ads, storage, and serving features that AI systems rely on. Uber introduced the concept in 2017, feature stores provide a unified space for building and sharing the characteristics of different organizations in the organization.
“Symbolic warehouses sit at the intersection of data and machine learning,” Michael Del Balso, general manager of Tecton.ai, a start-up software development company, told TechCrunch via email. “[Feature stores are] an important part of MLOps packages because they enable data teams to quickly build, reliably build high-quality attributes using real-time data that serves those production characteristics. to evaluate in real time. They serve as a data connection and [AI] model. “
In addition to simply storing data, feature stores allow data engineers to view feature statistics, including features used, locations used, and their impact on models. Feature repositories also modify data, allowing users to compile, filter, and join attributes without the need for code. (Think of the restaurant’s mixed orders to get the characteristic value “number of orders in the last 30 minutes.”)
Del Balso explained: “High-quality warehouses samee automate production pipelines to collect data from batch data sources and real-time sources, convert real-time data, and store data online and online store. [They often also] It includes built-in monitoring capabilities to monitor pipeline health, data transfer, service levels and more. ”
Visual warehouses promise to improve collaboration between teams as they move forward with the development of AI systems. As demand grows, large and emerging technologies such as Tecton are developing products to meet demand – and investors are enthusiastically supporting them.