Data-labelling startups want to help improve corporate AI

CORPORATE BOARDS are besotted with artificial intelligence. Worldwide spending on AI is expected to rise from $38bn this year to $98bn by 2023, estimates IDC, a research firm. So far, though, only one in five companies aware of the technology’s potential has incorporated machine learning into its core business. One reason for the slow uptake is the dearth of quality data to teach algorithms to perform useful tasks. The most common form of AI, called “supervised learning”, requires feeding software stacks of pre-tagged examples of, say, cat pictures until it can tell a feline image apart by itself. Data-labelling is the sort of grunt work that corporate AI-users would prefer someone else to do for them. An industry is popping up to help.

The market for data-labelling services may triple to $5bn by 2023, reckons Astasia Myers of Redpoint Ventures, a venture-capital firm. Some outfits, like Mechanical Turk (owned by Amazon, an e-commerce giant), act as middlemen connecting freelancers ready to perform all manner of “micro-tasks”, of which things like tagging pictures is one example, with taskmasters. Other firms specialise. Hive has turned data-labelling into something “like playing Candy Crush”, explains its boss, Kevin Guo, referring to a hit tile-matching game. Its mobile app makes it easy for users to identify objects, earning money instead of points. Its 1.5m players across the world serve more than 100 corporate customers.

Because human data-labelling is labour-intensive, most of it happens in low-wage countries like India, Vietnam and the Philippines. In such places data-labelling “is the easiest way to earn money”, says Hafiz Arslan, a Pakistani software engineer who was recently paid $200 for classifying 4,500 images by the sport they depicted (football, cricket or tennis).

A distributed workforce is, however, prone to human error. That is a problem for AI, which is only as good as the data it learns from. So other startups want progressively to cut humans out of the process. Scale AI, from San Francisco, lets its own algorithms take a first pass at labelling with humans reviewing the work. “We are extremely, extremely quality-conscious,” insists its boss, Alexandr Wang. He says revenues have grown tenfold from a few million dollars last year. Labelbox helps firms gauge the accuracy of labelling.

AI.Reverie goes further, dispensing with human labellers altogether. It uses techniques developed for video games to create and automatically label scenes that can be used to train image-recognition algorithms. Its approach is particularly useful for exposing software to scenarios that might be hard to find in data gleaned from the real world. It can generate scenes set underwater, or featuring heavy fog or torrential rain. The company’s backers include In-Q-Tel, a venture fund for America’s intelligence services.

The industry’s short-term future seems assured. In the longer run a threat may come from developments in “unsupervised learning”, which aims to identify patterns in data that have not been labelled by humans. Manu Sharma, boss of Labelbox, says this remains “primarily an academic pursuit”. How long for is anyone’s guess. ■