A new market has emerged in the race to advance the immersion of artificial intelligence technologyOpens a new window in the economy as a raft of new businesses draft contributors around the world to train computers to learn the language of the “human world†via crowdsourcing.
For AI to be effective, the ability to distinguish between pictures of cats and dogs, interpret between an angry or a happy customer over the phone or differentiate between a traffic light and a pedestrian in a self-driving car is no less than vital.
To achieve this level, AI software is taught by an army of humans who annotate and tag photos, pictures and voice recordings with the accurate vocabulary and description before uploading them to machine learning systems. (To illustrate the mammoth undertaking, human contributors receive images of, for example, a landscape and must then annotate a mountain, a lake and the sky. With a few thousand versions of these, a machine learning system begins to understand the concept of each. In healthcare, cancer specialists could be paid to upload tagged pictures of tumors.
Forging the New Boom Industry
Eric Schmidt, executive chairman of Google parent company Alphabet, predicts that the power of crowdsourcing in combination with AI will drive a boom industry, and holds that this junction is the heart of the next $100 billion company. In 2016, he forecast that the big opportunity in coming yearsOpens a new window would be in companies that “use the crowd to learn something,†and that those who successfully combined machine learning with crowdsourcing would create a breakthrough business model.
Schmidt illustrated the idea with the skincare industry: By paying dermatologists to identify skin samples with specific dermal conditions and feeding the data into a machine learning system, the AI – with thousands of examples to draw on -could give a better diagnosis than a single practitioner. “If my users teach me and I can sell to them and others a service that is better than their knowledge, it’s a win for everybody,†he told the Startup Grind Europe conference.
New Models Opening for Business
Entrepreneurs are examining ways of combining crowdsourcing with AI to build new business models, with several data-labeling startups winning substantial venture capital backing. Crowdsourced data labeling business CrowdFlower has attracted $58 million in funding (including investment fromUber-founder Travis Kalanick) and is valued at $110 million
Another startup is Mighty ai, a company that has styled its product as ‘Training Data as a Service’ (TDaaS), specializing in delivering training data to build computer vision models for autonomous vehicles. These two businesses, together with Alegion, Scale and Cloudfactory, attracted more than $50 million in venture capital fundingOpens a new window last year.
For a self-driving car to recognize traffic lights, one-way streets and pedestrians, it needs to be shown thousands – possibly millions – of hand-labeled photos. The AI trainers farm out this work to their crowds of labelers, who are paid for the number of pieces of data they label or identify.
A service called Clickworker asks users to listen to conversational responses and evaluate them. The site works with one million people around the globe. An India-based startup called Playment works with some 250,000 contributors who label images, tag pictures and analyze text on behalf of software companies.
The advantage of crowdsourcing is that suggestions can be cross-checked with other contributors so that only the most popular answers are adopted. This lowers, though does not remove, the possibility of errors.
Strength in Numbers
With AI and machine learning expected to power the next wave of the technological revolution, data labeling will be an essential factor of the equation. Some predict it will create a new type of gig-economy employment that soaks up the workers who lose out in the widely-predicted “job apocalypse†ushered in by AI and robotics.
Data labeling contributorsOpens a new window can earn between $10 and $14 an hour, and have the advantage of flexibility, though with few or no labor rights.
It seems counter-intuitive that the AI revolution should depend so strongly on the input of humans. Ideally, machine learning algorithms would teach themselves – show the ML software a million pictures of cats and humans, and the algorithm will build its own rules for differentiating between the two.
This level of sophistication, though, is some way off. Until machines become truly autonomous and can learn under their own steam, the future of AI will depend on lessons imparted by a workforce of millions of humans.