Researchers develop method for training AI without human data — Jerusalem Post

Researchers at Tel Aviv University in Israel and Stanford University have developed a method for training artificial intelligence in which models can train each other without using human-created content or external datasets. This was reported by the Jerusalem Post, citing a statement by Tel Aviv University.

Two models created tasks for each other

As part of the experiment, scientists used two AI models. One of them generated sequences of symbols, while the other had to predict the next symbol. The system did not need texts, photographs, or other external sources for this: the training materials were created within the process itself.

As the exercises were completed, the model that generated the sequences made more complex tasks that could improve the results of the second model. The latter, in turn, developed its ability to predict the next elements by solving these tasks. The study results were published on the scientific platform arXiv.

More current news is available on the UA.News Telegram channel Telegram.

Testing on texts, images and code

To determine whether the acquired skills could be applied outside the artificial environment, the researchers tested the model on written language, images, audio, music, computer code, and DNA sequences. According to the experimental results, prediction accuracy increased when more computing resources were allocated for training.

At the same time, the researchers emphasized the difference between recognizing patterns and acquiring knowledge about the real world. Models can develop prediction skills without external data, but they still need access to relevant data to obtain factual information about real objects and phenomena.