（重複して受取られた場合はご容赦ください）
東大物性研の高橋惇です。
東大物性研の赤松克哉さんによるテンソルネットワークを用いた機械学習に関するセミナーを来月、下記の通り開催します。
新たにzoom参加をご希望の方は、ウェブサイトからご登録ください。
kawashima.issp.u-tokyo.ac.jp/mcs/
（以前のセミナーで登録された方は再登録不要です）
皆様のご参加をお待ちしています。
=====
Speaker: Katsuya Akamatsu (ISSP, University of Tokyo)
Time & Date: 13:00 – 14:00, Sep. 5 (Fri), 2025 (JST)
Place: Hybrid (ISSP Room 615 & zoom)
Title: Generative modeling with nonnegative adaptive tensor trees
Abstract:
We propose a structural optimization scheme for a nonnegative tensor tree
model targeting a discrete data distribution [1]. The advantage of imposing
nonnegativity is that it permits a probabilistic interpretation for the
model through an equivalence with hidden Markov models. Thus, we can
operate entirely within a classical framework without passing through the
wavefunction, as is typically done in Born machine-based tensor network
approaches. Previous methods for tensor network machine learning only
consider networks with a fixed, predetermined structure that is based on
information about the dataset, but our approach, based on previous work
[2], also considers optimization of the tree structure in addition to the
parameters.
We show that both our proposed method [1] and the adaptive Born machine
approach [2] can handle a variety of generative modeling tasks on different
kinds of random, synthetic, and real-world datasets of varying complexity.
We find that the Born machine-based approach tends to do better in
minimizing the negative log-likelihood. However, for tasks where
understanding the underlying model is more important, the nonnegative
approach is preferable. We also demonstrate cases involving correlations
beyond two-point correlations, where these adaptive tensor tree methods
work and existing approaches do not.
References:
[1] K. Akamatsu, K. Harada, T. Okubo, and N. Kawashima, arXiv:2504.06722.
[2] K. Harada, T. Okubo, and N. Kawashima, Mach. Learn.: Sci. Technol. 6
025002 (2025).
=====
————————————————-
Computational Material Physics Mailing List
home: www.issp.u-tokyo.ac.jp/public/cmp/
archive: cmp-ml.issp.u-tokyo.ac.jp
X: x.com/cmp_ml
————————————————-