DATE2026.01.22 #Press Releases

Scheme for building generative models with tensor trees

-Enabling the Elucidation of Causal Relationships such as Biological Phylogenies-

Summary

A joint research team consisting of Katsuya Akamatsu, Project Researcher at the Institute for Solid State Physics, The University of Tokyo (at the time of the research: doctoral student, Graduate School of Science), Professor Naoki Kawashima of the same institute, Assistant Professor Kenji Harada of the Graduate School of Informatics, Kyoto University, and Project Associate Professor Tsuyoshi Okubo of the Graduate School of Science, The University of Tokyo, has proposed a structural optimization scheme for a single-layer Non-negative Adaptive Tensor Tree (NATT) that models a target probability distribution as an alternative paradigm for generative modeling.

The NATT scheme is designed to identify, among probability distribution functions that can be represented in a tensor tree format, the one that best fits a given set of samples. Each sample consists of a list of observed values of many random variables (features); for example, in image recognition tasks, a single image corresponds to one sample. As with general generative models, NATT can generate new “plausible” samples. However, the key focus of this study is that the structure of the resulting tensor tree reflects the information-theoretic structure of the target dataset. This property enables the estimation of causal relationships among random variables.

In this framework, the tree itself is designed to represent a probability distribution function, meaning that the network locally encodes probabilistic relationships among variables. The proposed scheme for constructing generative models based on tensor tree structural optimization represents a fundamentally different approach from the currently dominant paradigms in generative modeling, which form the foundation of artificial intelligence and machine learning applications across many areas of society. In particular, this approach is expected to be applicable to the elucidation of complex causal relationships involving a large number of interacting factors.

Figure：Estimation of Causal Structures Using Tensor Tree Generative Models

Links

The Institute for Solid State Physics, The University of Tokyo,
Kyoto University

Journals

Journal name	Machine Learning: Science and Technology
Title of paper	Plastic tensor networks for interpretable generative modeling