Snorkel AI scores $35M Series B to automate data labeling in machine learning

One of the more tedious aspects of machine learning is providing a set of labels to teach the machine learning model what it needs to know. Snorkel AI wants to make it easier for subject matter experts to apply those labels programmatically, and today the startup announced a $35 million Series B.

It also announced a new tool called Applications Studio that provides a way to build common machine learning applications using templates and predefined components.

Lightspeed Venture Partners led the round with participation from previous investors Greylock, GV, In-Q-Tel and Nepenthe Capital. New investors Walden and BlackRock also joined in. The startup reports that it has now raised $50 million.

Company co-founder and CEO Alex Ratner says that data labeling remains a huge challenge and roadblock to moving machine learning and artificial intelligence forward inside a lot of industries because it is costly, labor-intensive and hard for the subject experts to carve out the time to do it.

“The not so hidden secret about AI today is that in spite of all the technological and tooling advancements, roughly 80 to 90% of the cost and time for an average AI project goes into just manually labeling and collecting and relabeling this training data,” he said.

He says that his company has developed a solution to simplify this process to make it easier for subject experts to programmatically add the labels, a process he says decreases the time and effort required to apply labels in a pretty dramatic way from months to hours or days, depending on the complexity of the data.

As the company has developed this methodology, customers have been asking for help in the next step of the machine learning process, which is taking that training data and the model and building an application. That’s where the Application Studio comes in. It could be a contract classifier at a bank or a network anomaly detector at a telco and it helps companies take that next step after data labeling.

“It’s not just about how you programmatically label the data, it’s also about the models, the preprocessors, the post processors, and so we’ve made this now accessible in a kind of templated and visual no-code interface,” he said.

The company’s products are based on research that began at the Stanford AI Lab in 2015. The founders spent four years in the research phase before launching Snorkel in 2019. Today, the startup has 40 employees. Ratner recognizes the issues that the technology industry has had from a diversity perspective and says he has made a conscious effort to build a diverse and inclusive company.

“What I can say is that we tried to prioritize it at a company level, the full team level and at a board level from day one, and to also put action behind that. So we’ve been working with external firms for internal training and audits and strategy around DEI, and we’ve made pipeline diversity, a non-negotiable requirement of any of our contracts with recruiting firms,” he said.

Ratner also recognizes that automation can hard code bias into machine learning models, and he’s hopeful that by simplifying the labeling process, it can make it much easier to detect bias when it happens.

“If you start with a dozen or two dozen of what we call labeling functions in Snorkel, you still need to be vigilant and proactive about trying to detect bias, but it’s easier to audit what taught your model to change it by just going back and looking at a couple of hundred lines of code.”

Read More