Unexpectedly, driverless cars are actually a labor-intensive industry.

Car Audio Crystal 3.2*2.5mm 3225 24M (24.000MHZ) 20PF 10PPM 20PPM 30PPM

Single chip microcomputer STM32L151CCU6

One-piece inductor

When the patients were reading magazines and brushing Instagram, Shari Forrest opened an app on the phone and began training an artificial intelligence that would be used on the unmanned vehicle. Let's take a look at the related content with the car electronics editor.

Unexpectedly, driverless cars are actually a labor-intensive industry.

The 54-year-old Forrest is not an engineer or programmer. She relies on writing materials to make a living. Still, when it's free, Forrest will still log in on Mighty AI and spend his time tagging pedestrians, trash cans, and anything else you don't want to hit.

Mighty AI is an artificial intelligence crowdsourcing service platform. Here, Forrest's "colleagues" have 200,000 people worldwide. The main services provided by this company include computer vision, autonomous driving, natural language processing and more. In January of this year, the company just announced the completion of 14 million US dollars of B round financing.

"If you can earn a few dollars in the queue, why not?" Forrest said.

For the unmanned vehicle industry, this kind of data annotation work like Forrest is indispensable before the autopilot era really comes. To train the AI ​​that autopilots rely on, the amount of data required is often beyond imagination.

Companies like Google and GM rarely mention this, but it is the workforce of people like Forrest that forms the cornerstone behind the sizzling data centers.

Most companies that develop autonomous driving technology employ hundreds of employees and then outsource the work of recognizing pedestrians, cyclists and other obstacles to India or China. These workers need to mark thousands of hours of video frame by frame.

“Machine learning is a myth, like The Wizard of Oz,” said Jeremy Conrad, an investor at Silicon Valley hardware incubator Lemnos Labs. “The tagging team is critical to every company, and we still need them for a while because of the outdoor environment. Very varied."

The tremendous advances in artificial intelligence, sensor quality and computing power have laid the technical foundation for the revolution of unmanned vehicles. However, despite these innovations, in the years to come, humans still need to be quietly behind the scenes in order to keep these systems in the best condition.

“In my opinion, AI practitioners are generally arrogant and think that computers can solve everything,” said Matt Bencke, founder and CEO of Mighty AI. The same problem exists in artificial intelligence systems: computers need to “learn” through the huge amount of data that is manually tagged, and then use this 'model' to identify similar objects and shapes in the future.

The challenge of training an unmanned vehicle is greater than other artificial intelligence applications because the scenes and environments that the car faces are open. Even after adjusting for lighting and weather conditions at different times, the urban environment will change overnight due to construction, events or accidents.

“The labeling process is usually a very hidden cost and is rarely mentioned,” said Sameep Tandon, CEO of Autopilot startup Drive.ai. “It’s very painful and very troublesome.”

Unmanned vehicles require more precision than other artificial intelligence systems. The reason why cars can drive automatically is achieved by comparing them with the car's 3D map by using the surroundings of the camera and sensors. Safety is a top priority: If Google Photos' facial recognition can't confirm the person in a photo, it's no big deal; but if Waymo's unmanned car doesn't recognize the pedestrian, it can cause a fatal accident.

In the process of competing to develop unmanned vehicles, the measurement standards of enterprises are their accumulated mileage. Waymo said in May that their cars had been tested on public roads for 3 million miles, and Tesla said last year that they had used the Autopilot system to collect more than 10 million miles of driving data from existing owners.

However, the more mileage, the more data processing personnel behind these companies need to do more work by hand. A few miles of driving miles can generate tens of gigabytes of data, which can quickly become too large to be transmitted over the car's wireless network. Instead, it must first be saved on the hard drive and then transferred to the outsourcing center.

It is indeed surprising that such an industry with cutting-edge technology adopts such an old model.

David Liu, CEO of Silicon Valley Autopilot Systems Development Inc. PlusAI, says it takes hundreds of hours to convert it into useful data for every hour of driving.

He said, "We need hundreds of thousands or even millions of hours of data" to get unmanned vehicles to drive around and need "hundreds of thousands of people to achieve this goal."

But large technology companies often refuse to disclose the manual procedures involved in unmanned vehicles. Waymo, Uber and Tesla declined to comment.

“It’s hard to hear people talking about it,” said Dan Weld, a professor of computer science and engineering at the University of Washington. “They all like to say that machine learning is amazing.”

Anthony Levandowski, a former Waymo and Uber engineer, in a public speech at the University of California at Berkeley in 2013, called Google's Indian team a team of "human robots" that tagged various images captured by the Street View service. .

The cost of such labor-intensive work is not low. Industry insiders estimate that to create and maintain a map of each US city, the annual cost is about billions of dollars.

Some startups have seen opportunities, and PlusAI, Deepmap, and Drive.ai all claim to be able to use "deep learning" to reduce labor costs (about Drive.ai's plan to do drones entirely with deep learning, poke here), but The accuracy required for safe driving can still be maintained. But other industry insiders still suspect that deep learning is difficult to completely rid of dependence on humans.

Mighty AI's Bencke pointed out that the challenge for Facebook, YouTube and Twitter is how to deal with all kinds of negative content such as bullying and terrorism on their social platforms. "If deep learning has this ability, don't you think it should solve this problem now?" He said, "This problem is not as complicated as an unmanned car, and the market is very big."

At the moment, the wisdom of AI is still more dependent on the quality of the data it uses to train. Bencke said that his own platform will use his own machine learning technology to determine the areas that everyone is best at, and to distribute work. No one becomes rich, but for Mighty AI, it doesn't matter.

Forrest said that he earned $300 on this job last year, most of which was used for online shopping. She has never seen an unmanned car, let alone a ride. But she knows she is making them smarter and confident about the future of this technology.

The above is about automotive electronics. I never imagined that driverless cars are actually a labor-intensive industry. If you want to know more information, please pay more attention. Electronic engineering will provide you with more complete and detailed information. , updated information.