Facebook says its new artificial intelligence can identify more problems faster


A few lens learners pre-trained on fire hoses containing billions of Facebook posts and images in more than 100 languages. The system uses them to build an internal perception of Facebook’s content statistics model. The content review was adjusted through additional training on the posts or images flagged in the previous review items and simplified descriptions of the policies that these posts violated.

After it is ready, the system can be directed to find new types of content, such as implementing new rules or expanding to new languages, much less than the previous audit model, Cornelia Carapcea said. Facebook.

She said that a more traditional review system may require hundreds of thousands or millions of sample posts to deploy. Just use dozens of them (the “several shots” in the name), plus simplified descriptions or “hints” of the new policies related to them, and the less-lens learner can be put into use.

Carapcea said: “Because we have seen a lot, we can understand new issues or new strategies faster.” “It is always difficult to obtain enough labeled data on various issues such as violence, hate speech, and incitement; This allows us to react more quickly.”

It can also guide the few-shot learners to find content categories without showing them any examples, just provide the system with a written description of the new strategy-a very simple way to interact with the AI ​​system. Carapcea said that the results of this approach are not very reliable, but this approach can quickly suggest what the new policy will cover, or identify posts that can be used to further train the system.

The immense ability of artificial intelligence like Facebook to create impressive — and many unknowns — prompted researchers at Stanford University to recently set up a center to study such systems, which they called “Base model“Because they seem to be the basis of many technology projects. The large machine learning models being developed are not only used in social networks and search engines, but also in the following industries: finance with health care.

Percy Liang, director of the Stanford Center, said that Facebook’s system seems to demonstrate some of the impressive power of these new models, but it will also demonstrate some of their trade-offs. Being able to instruct artificial intelligence systems to do what you want to do with written text is “exciting” and useful. As Facebook said, new content policies are okay, Liang said, but this ability is poorly understood. “It’s more of an art than a science,” he said.

Liang said that Few-Shot Learner’s speed may also have disadvantages. When engineers do not have to organize as much training data as possible, they sacrifice some control and knowledge of their system functions. “Faith has taken a bigger leap,” Liang said. “The higher the degree of automation, the less potential oversight.”

Facebook’s Carapcea said that as Facebook develops a new audit system, it has also developed methods to check the accuracy or deviation of its performance.

More exciting connection stories


Source link

Recommended For You

About the Author: News Center