‘What I saw didn’t look very much like the future — or at least the automated one you might imagine. The offices could have been call centers or payment processing centers.’

The New York Times reports on the labour-intensive work of turning images and sounds into data. The objective is to make training material for learning-based AI systems.

The job is known as labelling.

It requires identifying a picture or sound to be properly classified, for example, drawing a line around a portion of an image and labelling it ‘dog.’ After a basic machine learning system sees thousands of images of dogs — maybe even millions — it forms its own model of what constitutes a dog.

The training data is a proxy for prior experience. With it, machines can make predictions that are identical to or better than humans.

Like previous computing, the quality of the output depends on the input. The material can be anything: a polyp, a tree, a cough, a laugh; the possibilities are endless. Often many sub-categories are involved; for example, the sound of a laugh could be ‘good’ or ‘evil’ or ‘soft’ or ‘belly’ or many other gradations.

For learning machines, what counts is that the label is accurate and the reference materials are plentiful.

Labelling is an entry-level position — aka low-paying — and the Times account says labelling generated $500 million of activity in 2018. It is expected to be worth $1.2 billion by 2023.

The Times visited labelling operations in five cities.


AI is learning from humans. Many humans.
NEW YORK TIMES | August 16, 2019 | by Cade Metz