China's ambition to develop 'AI scientists'

Chinese researchers have developed a new method to help train machines, hoping that this model can help create "AI scientists".

Chinese scientists say they have found a way to improve the method of training machines with existing knowledge (such as the laws of physics, logic, mathematics), to create 'AI scientists' capable of Ability to experiment and solve scientific problems.

Specifically, according to an article by Chinese researchers published in the journal Cell Press Nexus, deep machine learning models have 'revolutionized the field of scientific research' thanks to their ability to discover relationships between the two. system of large amounts of data. However, the machine model does not have a basic understanding of the world, so the products created are not highly accurate.

To overcome this, the machine model needs to use more available knowledge from humans. The challenge is that too much knowledge introduced at once will cause the model to malfunction. Therefore, Chinese researchers have developed evaluation rules for the machine to self-select and incorporate the most relevant knowledge into combination with data.

China's ambition is to create 'AI scientists' capable of experimenting and solving scientific problems. (Illustration photo: TDS).

One of the examples mentioned is Sora - a model that converts text descriptions into videos by OpenAI company (USA). Sora's developers say it can 'understand how things exist in the real world ', and is capable of advanced, realistic depictions of things.

However, the company admits that Sora still has difficulty simulating some aspects of the real world and cannot "accurately model the physics of many fundamental phenomena".

In fact, the company only "trained" Sora using large amounts of image data, allowing this artificial intelligence to select from a database of images and videos that simulate reality. They haven't been able to help Sora understand physical laws like gravity.

Chen Yuntian, study co-author and professor at the Eastern Institute of Technology (EIT), said: 'Without a fundamental understanding of the world, a machine model is essentially just a copy of data. not a description'.

Xu Hao, study co-author at Peking University, commented: 'Introducing human knowledge into AI models has the potential to improve their efficiency and reasoning ability . The question is how to balance the impact of data and knowledge'.

According to the paper, the new method helps calculate 'importance of rules' , which considers the influence of a particular rule on the predictions of a machine model. The AI ​​will then be taught the most relevant rules - for example, the laws of physics, giving results that are "more relevant to the real world".

The researchers experimented by applying this method to optimize a model that solves multivariate mathematical equations, the same model that predicts the results of chemical experiments.

Chen says that in the short term, this method will be useful for machine models that require consistency with the laws of physics.

The research team also discovered some existing problems. For example, after adding more data to the model, general rules will take precedence over specific rules. In some specific fields that require consideration of many factors - such as biology and chemistry, this can affect the results.

However, the team hopes that the new method will help machine learning models develop further, so that AI can gradually determine its own knowledge and rules, without human intervention.

'We want to turn this into a closed process, thereby turning the model into a real AI scientist ,' the team said.

The researchers are also developing open source tools for AI developers so they can do similar experiments.