Facebook's AI Manager: "You need to create fake videos to identify them."

The phenomenon of Deep Pike videos - programs that people have planted things that have never been said - is gaining momentum, but Dr. Antoine Borden is not worried: "We want to test these models, poison can also be used as an antidote, it will not be a big problem in the future."

A big area in your studies is machine learning; what are you doing?

"A major research topic that integrates with various projects is multi language, including the translation and naturalization of text in a variety of languages, and many AI challenges that can be dealt with through multilingualism — language and learning with little information.

"Many scientists think there are problems that there will always be a lack of information to solve them, so we need to find new ways to deal with a minority of information and multilingualism, so if I want to catalog texts in English or French, I have enough information to work with. In Korean, it's harder, and if it's Bengali, then there's no more information at all. "

Give an example of a project in the field.

"One of my favorite projects is an unregulated machine translation, which checks whether you can translate between two languages ​​without any prior knowledge of these languages, for example, if I want to transcribe from French to English, but I have only a lot of text in English and a lot of writing in French, And regardless of whether they can be translated anyway? "

If the texts are not the same at all, a person can not translate. You need some reference point.

"We can train machines to translate between languages ​​without any prior information, which is good at the same level of machine translation based on half a million translated sentences."

How Does It work?

"The world is organized the same everywhere: the clouds in the sky, the beach by the sea, the boats in the sea, etc. When people talk, they relate to objects in the same context, it does not matter if they are in Tel Aviv, Paris, or New York. The program first identifies matching words, which appear together many times, so it can recognize that groups of words in one language look like groups of words in another language, allowing the program to acknowledge parallel words in two different styles, And then the system learns from the mistakes and improves. "

How does the system learn from its own mistakes?

"If we use the phrase" the cat is blue, "it is a real sentence, which can be found in Wikipedia, for example, if it is entered into the English-French algorithm, it can translate it into" chat chat," which is still a lousy translation. , And if you get a result like A cat is red, the system realizes that a mistake was made in translation because it knows what the original sentence is and can update itself, of course, everything happens in the opposite direction.

"The system that created poison is the vaccine."

What else can machines learn?

"We want to create a program that can perform 99% of the learning process without tagged information (tagged information is predefined information for the system, such as a cat image with an information layer that identifies the image as a cat). So that only a small number of variables should be labeled. We want the software to learn what is in the picture, without being able to attach names to objects there. We do it with fun games that the software needs to solve. "

Give an example of such a game.

A simple task is to take a picture, turn it black and white, then ask the software to repaint it, for this program is not a simple task, because it needs to know that the sky is blue, that the sand is yellow and that everything has a specific color, Another mission is to take a picture, cut it into nine parts, change its order, and ask the software to rearrange it, another example is to display five video frames to the system, and then the system needs to predict the next structure. "

What does it give you?

"If I have to train the software to identify cats, I have to show it tens of thousands of tagged pictures of cats, to identify thousands of objects that require millions of images, each of which is labeled as a cat, a car, a person. To train herself, so that when she is presented with a picture, she can already understand what is happening in her, what can move and what does not, so that the software can identify a cat does not need 10,000 pictures, maybe ten pictures. "


It's like a little boy. Show him a picture of a cat once, and he'll recognize every cat he sees on the street.

"Intuitive physics is one of the first building blocks of understanding the world, and we have created a test to see if the software can learn intuitive physics and if we understand how learning machines can help them learn faster."

Translators Beethoven and Mozart

The technology press deals quite a bit with the Deep Pike phenomenon - videos or sound clips created or modified by software, and allows, among other things, to plant things in a person's mouth that he never said and generate a result that people will find challenging to distinguish from the real thing. Is this a problem you are dealing with?

In terms of applied research, we want to develop methods to identify Deep Pike, or as we call it 'information manipulation.' At the same time, we want to push models to generate images, videos and texts - if we're going to respond to what people do, What we can do: We realized that a program that knows how to create Deep Pike is also a program that can recognize it, as a poison can sometimes be used as an antidote.

"Now we're trying to figure out whether to develop an antibody that works against all toxins, which is an important issue and we take it seriously, but it's something that software creates and leaves with signature information that other software can recognize, it's just a matter of finding the right solution. in the future".


What is your Israeli team doing?

"The team here is interested in AI for creators and creativity, but there are also profound scientific challenges, and they have created a project called Universal Musical Translation, which feeds into a music segment algorithm and translates it into another instrument or musical arrangement. To change devices but also to take a piece of music played by one tool and turn it into a section designed for an entire orchestra, in a different but similar arrangement that preserves the original melody. "The algorithm can also translate the piece into the styles of various composers, such as Beethoven's piano or Mozart's symphony.