Chinese scholars and overseas collaborators have made progress in the research of stable learning inspired by cause and effect

Figure Causal Inspired Stable Learning Research Framework and Progress

The heterogeneity and uncertainty of real data in an open environment pose severe challenges
to the interpretability and generalization ability of current machine learning models.
How to break through the limitations of the basic assumption of independent and same distribution and explore a new way of machine learning with generalizable mechanism essence and stable performance and explainable performance is an important basic theoretical problem of the new generation of artificial intelligence, which has important application value
for high-risk scenarios such as medical and finance 。 With the support of the National Natural Science Foundation of China (grant number: U1936219, 62141607, 61772304), Cui Peng's research team from Tsinghua University, together with Professor Susan Athey of Stanford University, jointly conducted research on the integration of causal statistical thinking and machine learning framework, and made progress
in the research of causally inspired stable learning theory methods 。 The research results were published in Nature Machine Intelligence in February 2022 under the title "Stable learning establishes some common ground between causal inference and machine learning.
"
。 Article link: _istranslated="1">.

The research team explains that the inherent limitations of current machine learning methods in terms of interpretability, stability and fairness are rooted in their association statistical basis, demonstrates the theoretical feasibility and important challenges of causal inference in breaking through the limitations of the current machine learning bottleneck, puts forward a stable learning framework driven by out-of-distribution generalization by fusing causal inference correlation theory, and explores the theoretical path of advancing the association learning model based on sample reweighting to a causal learning model.
A stable learning method system for linear models and deep nonlinear models is established, and the effectiveness
of the method is proved through theoretical analysis and data experiments.
At present, related methods have achieved significant application value
in scenarios such as smart healthcare and Internet economy.

Machine learning aimed at out-of-distribution generalization is still in its infancy internationally
.
In order to promote the development of this direction, the team built and published NICO, a large-scale image evaluation set that supports out-of-distribution generalization research, and held the first NICO Challenge on out-of-distribution generalization image recognition based on this evaluation set in August 2022, attracting nearly 200 teams from home and abroad to participate and exerting significant influence
.
At present, the evaluation set has been adopted as a standard evaluation set by many research teams such as Stanford University, Massachusetts Institute of Technology, and the University of California, Berkeley
.
Dataset link: https://nicochallenge.
com/
.