A group of researchers from University of Beijing and the Eastern Technological Institute (eit) have collaborated to develop an innovative system that could revolutionize the way scientific research is conducted. According to a study published in the journal Nexus, this system has the potential to pave the way for the creation of “real AI-scientists.”
Deep learning models are known for their ability to identify patterns in vast amounts of data, making them valuable tools for research. However, these models lack a deep understanding of fundamental principles such as physical laws and mathematical logic that govern natural and technical processes.
To address this limitation, a team of scientists led by Professor Chen Yuntyan from EIT introduced the concept of “Informed Machine Learning.” This approach involves integrating specialized knowledge, such as physics laws and chemical equations, directly into the training data alongside the source information.
By incorporating this additional knowledge, the accuracy and effectiveness of neural networks can be significantly enhanced. However, the challenge of selecting and combining the right amount of relevant information had to be overcome to prevent the system from becoming overwhelmed and losing performance.
To tackle this issue, researchers developed an algorithm that assesses the “significance” of information and determines the optimal combinations to improve the predictive capabilities of the machine. According to Chen Yuntyan, integrating human knowledge into AI models can enhance their effectiveness and improve logical reasoning, which is crucial in scientific and engineering domains.
The new system was put to the test in solving algebraic equations and predicting the outcomes of chemical experiments, resulting in a noticeable improvement in the accuracy and stability of the AI models. The ultimate goal of the project is to create a fully autonomous system capable of independently deriving relevant rules and principles without human intervention, transforming it into a genuine AI-scientist.
Although researchers are currently developing an open plugin for AI developers to implement this ambitious vision, they have encountered challenges as the volume of input data rises. As the general rules begin to override local laws in complex fields like biology and chemistry with unique rules not easily described by concise mathematical equations, adapting the system remains a work in progress.