It has received widespread acclaim for its sophisticated and lifelike portrayals of objects, being lauded as a significant advancement in generative AI. However, the company has acknowledged its ongoing challenges in replicating certain real-world phenomena, such as the accurate simulation of basic interactions like glass shattering.
Sora undergoes training using extensive visual data, enabling it to recognize patterns for generating images and videos that closely resemble reality. Nevertheless, it lacks training in comprehending fundamental physical principles like gravity.
According to Chen Yuntian, a professor at the Eastern Institute of Technology and one of the study’s authors, “Without a fundamental understanding of the world, a model is essentially an animation rather than a simulation.”
The gap in AI capabilities between China and the US is widening, causing significant concern among experts.
The researchers from Peking University and EIT highlighted that deep learning models are primarily trained using data rather than prior knowledge, such as the laws of physics or mathematical logic.
Incorporating prior knowledge alongside data during model training can enhance accuracy, leading to the development of “informed machine learning” models capable of integrating this knowledge into their outputs.
However, determining which prior knowledge to incorporate, including functional relationships, equations, and logic, presents a challenge. The team noted that incorporating multiple rules could potentially cause model failures.
To tackle this issue, the researchers devised a framework to evaluate the significance of rules and identify the most effective combinations for predictive models.
Xu Hao, the first author and researcher at Peking University, emphasized, “Our framework can be utilized to assess different knowledge and rules to improve the predictive capacity of deep learning models.”
The framework assesses “rule importance” by examining how specific rules or combinations impact a model’s predictive accuracy.
Integrating AI models with rules such as the laws of physics could enhance their real-world applicability, particularly in scientific and engineering fields, according to Chen from EIT.
The team tested the framework by optimizing a model for solving multivariate equations and predicting the outcomes of chemistry experiments.
Looking ahead, the researchers aim to empower AI to autonomously identify knowledge and rules from data without human intervention, essentially transforming the model into an AI scientist.
Despite the progress, the team encountered a notable challenge during the study. As more data is introduced to a model, general rules become more prominent than specific local rules, posing difficulties in fields like biology and chemistry where overarching governing principles may be lacking.
