Sora is trained using large amounts of visual data and can detect patterns to generate images and videos that mimic reality. However, they are not trained to understand physical laws such as gravity.
“Without a basic understanding of the world, the model essentially becomes an animation rather than a simulation,” says study author Chen Yuntian, a professor at Eastern Technological University.
China's AI gap with the US is widening: 'We're all very concerned'
China's AI gap with the US is widening: 'We're all very concerned'
According to the paper, deep learning models are typically trained using data rather than prior knowledge, which can include things like the laws of physics and mathematical logic.
But scientists at Peking University and EIT have developed an “informed machine learning” model that can use prior knowledge along with data to make the model more accurate when training the model, and incorporate this knowledge into the output. It says that you can create .
Deciding whether to “pretrain” a model by incorporating prior knowledge such as functional relationships, equations, or logic can be difficult, and incorporating multiple rules can also lead to model collapse, the team wrote. I am.
“When faced with large amounts of knowledge and rules, as is often the case, current machine learning models based on information tend to struggle or fail,” Chen says.
To address this issue, researchers created a framework to evaluate the value of rules and determine which combinations yield the most predictive models.
“Incorporating human knowledge into AI models has the potential to improve AI efficiency and reasoning ability, but the question is how to balance the impact of data and knowledge,” said first author and Peking University professor says researcher Xu Hao in Cell. Press statement.
“Our framework can be used to evaluate different knowledge and rules to enhance the predictive power of deep learning models.”
According to the paper, the framework calculates “rule importance” and examines how a particular rule or combination of rules affects a model's predictive accuracy.
EIT's Chen said in a statement that teaching AI models such rules (such as the laws of physics) can make them “more reflective of the real world and more useful in science and engineering.” He said that there is a sex.
The researchers tested one framework they use to optimize models for solving multivariate equations and another they use to predict the results of chemical experiments.
Chen said that in the short term, the framework will be most useful in scientific models, where “consistency between the model and the laws of physics is important to avoid potentially disastrous consequences.”
What's going on with the US-China conflict in major high-tech fields, from AI to EVs?
What's going on with the US-China conflict in major high-tech fields, from AI to EVs?
The team hopes to take the framework further, allowing AI to identify its own knowledge and rules directly from data without human intervention.
“By making the model a real AI scientist, we want to make it a closed loop,” Chen said in a statement. The team is developing an open source plugin tool for AI developers that can accomplish this.
However, the team has already identified at least one issue.
During their research, the team discovered that as more data is added to the model, general rules become more important than specific local rules. However, this is not useful in fields such as biology or chemistry. This is because these fields “often lack readily available general rules similar to governing equations”.