Israeli Scientists Develop AI Training System That Learns Without Human Input

Researchers say the breakthrough could make future AI systems more efficient.

By Shula Rosen

Artificial intelligence may be able to develop essential learning skills without first studying human-created information, according to new research by scientists at Tel Aviv University (TAU) and Stanford University.

The researchers have designed a system in which AI models improve their ability to identify and anticipate patterns by working through problems generated entirely by artificial intelligence, eliminating the need for conventional training material during this stage of development.

TAU announced the findings Wednesday, following their publication on the scientific platform arXiv.

At the center of the research is a different approach to preparing AI systems to interpret information. Rather than supplying a model with extensive collections of text, photographs or other existing material, the scientists created an environment in which two models interact through a series of prediction exercises.

One model constructs sequences of symbols, while its counterpart must determine which symbol is likely to follow. The model generating the sequences gradually becomes better at designing problems that challenge the other system, creating an increasingly demanding learning process.

This arrangement allows the prediction model to refine its abilities through repeated practice without drawing on information collected from the outside world.

To assess whether the training produced useful skills, the researchers introduced the resulting model to a variety of information formats that had not formed part of its original exercises.

These included written language, images, sound recordings, music, software code and DNA sequences. Across the evaluations, the scientists observed improvements in prediction accuracy as they increased the computing resources available for training.

The results suggest that some of the capabilities AI needs to process complex information can be cultivated independently of the large datasets traditionally used to train such systems.

However, the technique does not eliminate the need for real-world information altogether. The researchers stressed that developing an ability to recognize patterns is different from acquiring factual knowledge, which still depends on access to actual data.

The potential advantage lies in preparing AI models to learn more efficiently once they encounter that information, rather than requiring external datasets to develop every aspect of their predictive abilities.

The research comes as Israel considers the broader opportunities and challenges presented by advances in artificial intelligence.

On November 3, President Isaac Herzog launched the “Beyond Intelligence” forum in partnership with the National AI Directorate in the Prime Minister’s Office and the Mosaic Institute for AI Policy to examine the technology’s implications for employment, economic security and society.

Calling for AI development to become a national strategic priority, Herzog declared: “Israel must take its place at the forefront of artificial intelligence. This train will not wait for us, and the whole world is looking our way.”