Neural network has created its own language that scientists cannot decipher

(ORDO NEWS) — The DALLE-2 neural network, as it turned out, creates hints for itself, helping to convert text into an image.

At first, these words seemed like gibberish to scientists, but then it turned out that the neural network understands them.

The DALLE-2 neural network, announced not so long ago, impresses with its abilities. So, it is able to convert any text into a corresponding image.

For example, you describe the scene you want to receive with the necessary details, and the neural network itself creates a picture. It turns out she is very professional.

New neural network language

Researchers using artificial intelligence with machine learning (AI) often don’t know exactly how their algorithms solve the problems they face.

Similarly, with DALLE-2: scientists still did not understand what processes take place inside this “black box”.

Having uncovered them now, the researchers came to a surprising conclusion: artificial intelligence has created its own language for building images.

The authors noticed that the DALLE-2 system, which creates images based on input text, will return nonsense words under certain circumstances.

For example, images with the word “airplane” often return some sort of gibberish. At first it seemed that this was a bug, but then the scientists came to the conclusion that the sequence of letters in the returned words was not random.

To test this, scientists simply “fed” the neural network with the words it had previously generated. It turned out that these sets of letters really matter to the program, and they mean the same thing that was previously entered into the system.

This may mean that the neural network for some reason converts human words into its own, so that it is then easier to generate images based on them.


Contact us: [email protected]

Our Standards, Terms of Use: Standard Terms And Conditions.