The Meta group has developed a new artificial intelligence technique, which allows it to react more quickly to problematic content (hatred, violence, conspiracy, etc.) posted on Facebook or Instagram. Authors of this type of message often surf on particular events and constantly seek to bypass existing filtering systems.
However, adapting a classic neural network can take months, because you have to collect and label a large amount of training data. And it is particularly difficult for new phenomena, where examples are still few.
Also to discover in video:
This is why Meta researchers created the “Few-Shot Learner” (FSL) technology, which only needs a small amount of training data to be able to correctly classify new elements. Result: the deployment of the new model can be done after a few weeks, instead of several months. FSL already works in over a hundred languages.
So how does it work ? The idea is to create a language comprehension model that is more generalist, and therefore less dependent on learning data. This model is first trained on “Billions of generic linguistic examples and open source “.
Then, the system is fed with prohibited or problematic content already labeled. Thanks to self-supervised learning techniques, it is then sufficient to provide it with a few dozen examples of problematic content for the system to implicitly and automatically generate the new filtering rule.
The tests carried out have shown that this way of doing things allows problematic content to be quickly exfiltrated, even if the formulations are innovative or ambiguous.
This was particularly the case for posts titled “Vaccine or DNA modifier” (conspiracy) and “Maybe this guy doesn’t need all his teeth” (violence).
Beyond this precise use, the new FSL technology is a small step closer to the Grail of artificial intelligence, namely the creation of a generalist and totally autonomous system, similar to human intelligence.
Source : Meta