Almost all modern artificial intelligence (AI) systems learn the same way: they are shown huge arrays of real data - books, websites, images, recordings - and they look for patterns in them. But what if the training data are not taken from the external world but created inside the system itself? A new study conducted with the participation of Tel Aviv and Stanford universities gives the first answer: this is also possible.
The work was carried out under the leadership of Dr. Yoav Levin from the Blavatnik School of Computer Science at Tel Aviv University as part of postdoctoral research by Dr. Kfir Dolev together with colleagues from Stanford. The preprint (a scientific article before peer review) is available on the arXiv server.
The essence of the method is the joint training of two models from scratch. The first generates short computer programs that output sequences of symbols. The second learns to predict the next symbol in each sequence. The generator receives feedback: how much its tasks help learning. Gradually it begins to select tasks corresponding to the growing capabilities of the second model.
A decisive test awaited after training: will the skills acquired on artificial sequences help predict real data that the model has never seen? The researchers evaluated its performance on texts, images, sounds, music, code and DNA sequences.
It turned out that the more computing resources were spent on training, the more steadily prediction accuracy grew - and in many areas at once. The model was also able to derive rules from examples offered to it during testing, without additional training.
The most intriguing thing was found in the material that the system created for itself. During training, familiar mathematical structures emerged in it, including the Fibonacci sequence (a series of numbers where each next one equals the sum of the two previous ones). In other words, the system learned to find ways to generate structured sequences useful for the second model. In another experiment, preliminary training with the new method helped the model master real data faster when they were eventually shown to it.
The work gives a preliminary answer to a fundamental question: does intelligence have a "universal" component independent of external data? And can part of learnability arise from general patterns accessible without acquaintance with reality? The results say that this is possible. However, the authors emphasize that the method is not capable of giving the system facts about the world without exposing it to that world. Rather, it will help AI learn more efficiently from real data.
"For years it was believed that for any useful training of an AI model, it was necessary first to feed it huge volumes of real data. In this study we refuted this assumption and showed that a model can acquire general prediction skills by training only on data created in the learning process itself," Levin concludes. He adds that this is so far a proof of concept on small models, but the results raise questions for future work: what universal mathematical patterns will manifest themselves when AI is strengthened with these methods? How far can one advance without external information? And can models be prepared for more efficient learning from reality with the help of experience they have created themselves?