Driverless looks very frontier but in fact there is a lot of manpower to teach it

(Original title: Self-driving cars prove to be labour-intensive for humans)


Netease Technology News July 11 news, "Financial Times" website wrote that the driverless car industry looks very cutting-edge, but in fact it is also labor-intensive, behind the data processing requires a lot of manpower investment, and cost is not cheap .

The following is the main content of the article:

Driverless cars sound like a very magical idea. A car concept that can automatically drive and cancel a steering wheel or pedal is like a plot in science fiction.

But like many fantasy stories, behind the driverless car there is also a "witch" - and the number is not small. It turns out that a lot of manpower is required in the journey to fully automated driving.

Many companies involved in the technology employ hundreds or even thousands of people. These employees are often overseas outsourcing centers in India or China. Their mission is to teach robot cars to identify pedestrians, cyclists, and other obstacles. To accomplish this work, these workers need to manually tag thousands of hours of video footage frame by frame. The prototypes of the driverless cars that traveled on test sites in Silicon Valley, Pittsburgh, and Phoenix.

“Machine learning seems to be a myth. Only the wizards can make it.” said Jeremy Conrad, an investor at San Francisco’s Lemnos Labs, “marking the team is very important in every company. We will need their participation for some time in the future. After all, the outdoor environment in which cars are located is too complex and changeable."

The tremendous advances in artificial intelligence, sensor quality, and computational performance have laid the technological foundation for the revolution in unmanned vehicles. However, despite these technological innovations, the development of driverless cars for many years to come will also require humans to contribute silently behind the scenes: marking the trees and highlighting traffic signs to keep these systems up to date.

Challenge greater than other AI applications

Matt Bencke, founder and CEO of startup Mighty.Ai, said, “In my opinion, AI practitioners generally have a blind spot because of arrogance. That is, computers will solve everything.” Mighty. Ai is committed to using the part-time community to filter and mark AI training data for technology companies.

The same problem exists for any AI system: Computers “learn” by absorbing large amounts of manually labeled information, and then use that “model” to identify objects and patterns that appear once again in front of them.

The challenge of training driverless cars is greater than other AI applications because cars may encounter numerous scenarios and situations. Even if the system can adapt to changes in light and weather during different days and years, it may not help because the urban environment may change at any time because of buildings, special events, or accidents.

“The data tagging process is a hidden cost, and people don’t discuss it much,” said Sameep Tandon, CEO of driverless startup Drive.ai. “The process is extremely difficult and troublesome.”

Driverless cars also require higher accuracy than other AI systems. Auto-piloting of cars is achieved by comparing the surroundings they see with cameras and sensors and the detailed on-board 3D maps of the streets around them. Security is at the top of the list: If the Google Photos face recognition system fails to accurately identify people on the map, it will cause trouble; if the Waymo car does not recognize pedestrians, it may cause pedestrians to be killed.

One of the criteria for measuring progress in the struggle to create driverless cars is the number of miles the company’s cars have traveled. Waymo, of Alphabet, once said in May that its car has already traveled 3 million miles on public roads. Tesla also said last year that it has collected more than 100 million miles of mileage data from existing car owners, which will help it develop the Autopilot Autopilot system.

However, the longer the mileage, the more manual tasks the team behind the data processing team have. Driving a few miles can produce dozens of gigabytes of data, which will soon be too large to be uploaded directly from the car. Instead, the data must be saved to the hard disk and then sent to an outsourcing center. For such a frontier industry, the operation of this kind of logistics seems to be somewhat behind.

David Liu, CEO of Silicon Valley startup Plus.ai, which develops driverless car systems, points out that the data generated by each hour of driving can take hundreds of hours to translate into useful data. "We need to get hundreds of thousands or even millions of hours of data from driverless cars that drive everywhere, and thousands of people need to work together around the world to complete this work."

Large technology companies tend not to promote the artificial work part of driverless cars. Waymo, Uber, and Tesla all declined to comment.

"It's not easy to get people to talk about this," said Dan Weld, a professor of computer science and engineering at the University of Washington. "They all prefer to talk about 'magic' like machine learning."

A rare public discussion took place in 2013 when former Waymo and Uber engineer Anthony Levandowski spoke at the University of California, Berkeley. He talked about a Google team in India called what he called a "human robot," which is responsible for marking images from Google Street View services.

This kind of labor-intensive work is not cheap. According to industry estimates, the cost of creating and maintaining such maps for each city in the United States amounts to billions of dollars annually.

Deep learning can replace artificial?

Some start-ups saw business opportunities. Companies such as Plus.ai, Deepmap, Drive.ai and others claim that they can use “deep learning” techniques to reduce this human input while maintaining the accuracy required for the safe operation of driverless cars. Deep learning is a newer, more advanced machine learning technology that seeks to simulate the human brain's analytical process.

According to James Wu, CEO of Deepmap, “With machine learning, it is difficult to achieve 90% or more of accuracy, but with deep learning, it's much simpler to build a model like that.” In May, financing was completed at 25 million U.S. dollars.

However, other industry professionals are skeptical that deep learning will eliminate the need for manual participation. Mighty.ai's Bank pointed out that Facebook, YouTube, Twitter and other social platforms face enormous challenges in dealing with cyberbullying, terrorism and other issues. "If deep learning is so powerful, don't you think they've solved that problem now?" he said. "It's far less complicated than a driverless car. It's a big market."

AI researchers around the world are chasing the goal of “unsupervised learning” (ie, machines can learn without assistance). At the same time, Silicon Valley and Detroit's “Wizards” would like their customers and investors to continue not to pay attention to the heavy human work behind driverless cars. (Lebang)

800 Puffs

Shenzhen Xcool Vapor Technology Co.,Ltd , https://www.szxcoolvape.com

Posted on