Chinese researchers are on the brink of pioneering a groundbreaking approach to developing 'AI scientists capable of conducting experiments and solving scientific problems.
Recent advances in deep learning models have revolutionized scientific research, but current models still struggle to simulate real-world physics interactions accurately.
'Real AI Scientists'
However, a team of researchers from Peking University and the Eastern Institute of Technology (EIT) in China has developed a new framework to train machine learning models with prior knowledge, such as the laws of physics or mathematical logic, alongside data.
The South China Morning Post reports that this approach could lead to the creation of "real AI scientists" who can improve experiments and solve complex scientific problems.
Deep learning models have significantly impacted scientific research by uncovering relationships from large datasets. Despite these advancements, current models like OpenAI's Sora face limitations in accurately simulating certain real-world physics interactions.
For instance, Sora, a text-to-video model, has been widely praised for its advanced, realistic depictions of objects. However, it struggles to model basic interactions accurately, such as the directions that the flame on birthday cake candles goes.
Embedding Human Knowledge Into AI Models
The researchers propose incorporating 'prior knowledge,' such as the laws of physics or mathematical logic, alongside data to train more accurate machine learning models.
Embedding human knowledge into AI models has the potential to improve their efficiency and predictive capability. To address this challenge, the team developed a framework to evaluate the value of prior knowledge and determine its impact on model accuracy.
Their framework aims to assess the worth of knowledge by derived rule importance, considering factors like data volume and estimation range. By conducting quantitative experiments, the researchers aim to elucidate the complex relationship between data and prior knowledge, including dependence, synergistic, and substitution effects.
This model-agnostic framework can be applied to various network architectures, providing a comprehensive understanding of the role of prior knowledge in deep learning models.
What's Next?
The researchers tested their framework on models for solving multivariate equations and predicting chemistry experiment outcomes. They found that incorporating prior knowledge significantly enhanced the performance of these models, particularly in scientific domains where consistency with physical laws is crucial to avoid potentially disastrous consequences.
In the long term, the team aims to develop AI models that can independently identify and apply relevant knowledge without human intervention.
However, they acknowledge challenges such as the dominance of general rules over specific local rules as more data is added to a model, particularly in fields like biology and chemistry where general rules may be lacking.
"We want to make it a closed loop by making the model into a real AI scientist," said Chen Yuntian, an Eastern Institute of Technology professor and study author. The team is developing an open-source plugin tool for AI developers to advance this research further.
Stay posted here at Tech Times.
(Photo : Tech Times Writer John Lopez)
