Berlin, Dusseldorf, Hannover The view of the screen is rigid and strained. Roads can be seen, cars, scooter riders, cyclists, pedestrians, houses, trees, traffic lights and traffic signs. The objects are outlined and named using a computer mouse, ie “labeled”.
It's the profile of a bone job that hundreds of millions of people do around the world – every day for ten to twelve hours. They are sitting in front of their computers in India, the Philippines, Venezuela, Brazil, Malaysia, Kenya and many other countries.
These digital workers are called Clickworker. Their task: the digital modeling of our environment, called in the jargon semantic segmentation. The goal: Training Artificial Intelligence (AI). The principle can be imagined as with a child exploring the world. Initially, it has to be painstakingly taught what a toy is, what a human is, what is edible and what is poisonous.
The same goes for the AI systems. Many recordings come from car companies, suppliers and IT companies, for example, who want to train systems for self-driving cars. Unlearned systems can not do anything with raw images. The artificial intelligence, which is to be used in the vehicles, must first be taught, for example, to distinguish a playing child from a road sign.
In the end, the systems should surpass humans and be able to detect dangerous situations more quickly – if they have been prepared sufficiently and with good data. Massive amounts of data are being processed – with the help of people who live almost exclusively in developing countries and who record videos for just a few dollars. Precarious working conditions are the rule, fixed employment rare, Clickworker unions do not exist.
The freelancers shimmy from one small order to the next and can hardly make ends meet with their wages. They are, so to speak, the invisible workbench of the automobile industry, the assembly line workers of the present – for whom the companies do not even have to set up a factory anymore.
And they are the key to the trillion-dollar autonomous driving. Manufacturers like VW. BMW and Daimler, Suppliers like Continental. Bosch and Hella and especially the IT companies Google. Amazon and Apple pumping billions into development to get the first fully autonomous car on the road. The market promises a gigantic potential. Analysts of the Swiss Investment Bank UBS estimate it at one trillion dollars by 2030.
But the preliminary work until then is gigantic. Marc Mengler, head of the Berlin-based startup Understand.ai, which specializes in the labeling of video data for the auto industry, knows this well. “Without manual post-processing, the best algorithms are not good enough,” says Mengler. Customers of Understand.ai, which includes well-known automakers, as a glance at the homepage of the start-up reveals, send Mengler the video raw data, which are then labeled image by image using a special program.
“This means that every object that is relevant for autonomous driving must be marked – sometimes with pixel precision,” explains Mengler. Otherwise, the AI would not know what to watch on the video recordings. In other words, the AI would be blind.
A bill gives an idea of the unimaginable workload: A one-hour video shot at 30 frames per second consists of 108,000 pictures. If a car company takes in a hundred thousand hours of video material, which is not uncommon in the field of autonomous driving, this results in almost eleven billion pictures. Labeling complex traffic situations can take up to three hours per image. If a single person were to carry out this work, he would need more than 32 billion hours, or almost 3.7 million years.
“Ghostwork” is what Mary L. Gray and Siddharth Suri call this kind of work in their book of the same name (“Ghostwork”). According to the two authors, the Clickworkers are the new digital proletariat, which, like modern migrant workers, undertakes masses of mechanical, repetitive tasks as needed.
The greater the need, the more likely a person becomes Clickworker
The beginnings of Clickworking date back to 2005, when Amazon launched its crowdsourcing platform Mechanical Turk (MTurk). There, companies were able to outsource digital mini-work to freelancers around the world. At first, for example, they had to curate content for startups such as the Yelp rating portal or edit customer reviews. A short time later, more and more crowdsourcing platforms were established, such as Mighty AI, Hive and Playment, which take on complex tasks in the Clickworking area.
From nowhere, these start-ups were flooded with orders from the car industry from 2017 onwards. Since February 2019, Mighty AI has collaborated with the University of Michigan in collaboration with GM. ford. Honda. Toyota. Intel, LG and Verizon,
Anyone can log in to the platforms and become Clickworker over the Internet. After logging in, the users go through short tutorials, collect experience points with their first small assignments and work their way up to the more lucrative tasks. There is no direct contact between clickworkers from the crowd and clients from the automotive industry.
The auto companies send video raw data to the platforms, which in turn split them into small sections. Each video clip is an order that a Clickworker who is logged in to the portal can accept. For each order usually a few dollars are paid.
Many of these “labelers” are currently from Venezuela, as shown by an analysis by the Hans Böckler Foundation. 75 percent of the total traffic on the portals Mighty AI and Hive come from the South American country. Parallel to the drastic deterioration of the economic situation in Venezuela, more and more people were “trained” to become Clickworkers there. The reason is understandable: The Clickworking portals usually pay in US dollars. The greater the economic hardship of a country, the greater the willingness of residents to become clickworkers.
Christian Papsdorf, a professor at Chemnitz University of Technology and an expert in sociology of technology, criticizes Clickworkers for their insecure working conditions, reminiscent of the early days of industrialization, when hourly wage earners were exploited for low salaries – with one difference: “Now we have it with minute wage laborers do.”
However, Demetrio Aiello, head of the Artificial Intelligence and Robotics research division at Continental, sees no future in the development of AI for autonomous driving as an alternative to human labor. “Machine learning only works in interaction with humans,” says Aiello.
“He has to interpret the images that are taken during the test drives.” To make matters worse, the high safety requirements. “In contrast to the labeling of data in the entertainment sector, the quality standards for autonomous driving are significantly higher,” explains Aiello.
Because of this, attempts to label video footage using the Amazon service MTurk have quickly ceased, as it is not enough for the high standards. Continental resorts to specialized companies. One of these partner companies is Samasource.
(t) Artificial Intelligence (t) KI (t) Clickworker (t) VW (t) Continental (t) Hella (t) Daimler (t) BMW (t) Honda (t) Under-stand. ai (t) startup (t) auto industry (t) suppliers (t) assembly line workers (t) Bosch (t) Google (t) Amazon (t) auto information technology (t) artificial intelligence AI (t) automation (t) Ford (t) Hans Böckler Foundation (t) Hella Aglaia (t) Chemnitz University of Technology (t) General Motors (t) UBS (t) Toyota (t) Yelp (t) Verizon (t) Intel (t) Apple (t) Demetrio Aiello (t) automotive parts (t) automotive industry