Sora is trained using large amounts of visual data, allowing it to pick up patterns to create images and videos that mimic reality. But it is not trained to understand physical laws like gravity.
“Without a fundamental understanding of the world, a model is basically an animation rather than a simulation,” said study author and Eastern Institute of Technology professor Chen Yuntian.
China’s AI gap with US is widening: ‘We are all very worried’
China’s AI gap with US is widening: ‘We are all very worried’
According to the paper, deep learning models are typically trained using data and not prior knowledge, which can include things like the laws of physics or mathematical logic.
But Peking University and EIT scientists write that when training models, prior information can be used with data to make them more accurate, and “informed machine learning” models can be created that use this knowledge. be able to add to your output.
The team wrote that it was a challenge to “pre-learn” what prior knowledge – which could include things like functional relationships, equations and logic – to incorporate into the model, and to incorporate multiple principles into the models. can lead to falls, the team wrote.
“When faced with a large amount of knowledge and rules—which is often the case, current informed machine learning models struggle or fail,” Chen said.
To address this problem, the researchers created a framework for evaluating the value of the rules and determining which combinations resulted in the most predictive models.
“Incorporating human knowledge into AI models can improve their performance and predictive ability, but the question is whether data and knowledge can be combined,” first author Zhu Hao, a researcher at Peking University, said in a cell. How to Balance Influence”. Press statement.
“Our framework can be used to evaluate various knowledge and rules to enhance the predictive ability of deep learning models.”
According to the paper, the framework calculates “rule importance,” looking at how a particular rule or combination or rules affect a model’s predictive accuracy.
Teaching AI models about such laws — for example, the laws of physics — can make them “more reflective of the real world, making them more relevant to science and engineering,” EIT’s Chen said in the statement. will make it useful”.
The researchers tested their framework by using it to optimize one model for solving multivariate equations, and another to predict the outcome of a chemistry experiment.
Chen said that in the short term the framework would be most useful in scientific models “where consistency between the model and the laws of physics is critical to avoid potentially catastrophic results”.
From AI to EVs, how is the Sino-US rivalry playing out in key high-tech sectors?
From AI to EVs, how is the Sino-US rivalry playing out in key high-tech sectors?
The team hopes to take their framework even further by allowing AI to identify knowledge and rules directly from data without human intervention.
“We want to make it a closed loop by making the model a real AI scientist,” Chen said in the statement. The team is developing an open-source, plug-in tool for AI developers that could allow them to achieve this.
However, the team has already identified at least one problem.
During the study, the team found that when more data is added to a model, general rules become more important than specific local rules, but this doesn’t help in fields like biology and chemistry because they “often There is a lack of readily available general principles. The governing equation”