500,000 people are working through Amazon's crowdsourcing to train AI

ByRicardo Diaz

Even though AI (Artificial Intelligence) is excellent, there are still many things that humans can handle effectively. As a place for finely assigning such jobs, Amazon provides a service called "Amazon Mechanical Turk" (AMT), but now, a major high-tech enterprise utilizes this AMT user for supporting machine learning .

Amazon Mechanical Turk - Marketplace for crowdsourcing | AWS
https://aws.amazon.com/jp/mturk/


Inside Amazon's clickworker platform: How half a million people are being paid pennies to train AI - TechRepublic
http://www.techrepublic.com/article/inside-amazons-clickworker-platform-how-half-a-million-people-are-training-ai-for-pennies-per-task/

According to the explanation of the official website, AMT is still more human than the evolving computing technology such as "identification of objects of pictures and moving images", "deduplication of data", "transcription of voice recording", "research of data details" It is a service to quickly and inexpensively process tasks that can be effectively performed by assigning it to people called "Turker" registered in AMT.

That cost (consideration paid to Taka) is too low, so in 2014Open letter to Mr. Bezos, "We are not an algorithm" is releasedSometimes it was done.

ByKevan

Major IT companies such as Google, Microsoft, Amazon, Apple, IBM, Facebook, etc. are using AMT and similar crowdsourcing "CrowdFlower"Or, they are preparing similar cloud working platforms at their company and they are working on such work.

In the case of AMT, Takers are about 500,000 people, about 75% are Americans, 15 to 20% are Indians. Most Americans are women, but Indians are mostly men. Many of them are born in 1980 and 1990.

Taskers are assigned the tasks that human beings can do effectively, as mentioned above, but now they are doing jobs like "support for machine learning". One example is the computer vision system used in Tesla Motors' automatic operation ", Amazon's speech recognition assistant" Alexa ", Microsoft's speech recognition personal assistant" Cortana ".

Tesla Motors launches a new demo image showing the power of "fully automatic driving" - GIGAZINE


Speaker type audio assistant who became the biggest hit item in Amazon hardware history "Amazon Echo" The birth of secret story - GIGAZINE


Siri-like "Cortana" made by Microsoft is also available in Japan - GIGAZINE


With these technologies, the computer must recognize that "what is shown on the camera is a person or an indicator" or "what is being instructed by voice", but its contents are very complicated is. So, we will train AI by tagging each content "what is shown in the camera" and "what kind of word was issued". Taka is doing this tagging work.

The amount of data handled is enormous, for example"YouTube-8M" Google announced on September 28, 2016There are 8 million images,"Open Image Dataset" announced on September 30, 2016There are 9 million photos in each, each tagged.ImageNet14 million tagged images are released, but this is the result of processing 1 billion image candidates with 50,000 Takers taking 2 years.

In 2012,A man who continues to monitor children for child pornography or grotesque contentAlthough the story of Takers existed right now, the takers are in this situation right now, recently the organization which called "Islamic state (Islamic state, IS)" was involved, the head stacked in one cup The picture that people are burned, "bloodshed", "cutting / destruction" is also "commonplace". Also, as with the male story of 2012, I also reported that I have been reporting on child pornography all the time.

However, we do not know who the client is for the takers, so after reporting "This is child pornography", it seems that we do not know if the investigation is done or whether it is deleted.

As these tagging progresses, will people no longer need to tag them? The idea comes up, but at the University of SouthamptonGopal RamchurnIn the case of image recognition as an example, "We have not reached the limits yet, as each piece of photograph requires explanation" in which context this picture was taken in each context " Dependence on humans continues. "Even if tagging 50 million images, there are only a few images that can be accurately classified."