| 研究課題/領域番号 |
22J12703
|
|---|
| 配分区分 | 補助金 |
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
張 一凡 東京大学, 情報理工学系研究科, 特別研究員(DC2)
|
|---|
| 研究期間 (年度) |
2022-04-22 – 2024-03-31
|
| キーワード | 機械学習 / 圏論 / machine learning / disentanglement / category theory |
|---|
| 研究実績の概要 |
In this project, "Algebraic Structures in Weakly Supervised Disentangled Representation Learning", we aimed to develop theoretical tools and practical algorithms for learning abstract and meaningful representations. As the first step, we conducted a meta-analysis of various definitions of disentanglement in machine learning. Using category theory as a unifying framework, we revealed the similarities and differences between different definitions. We also introduced tools to analyze disentanglement in different settings, including equivariant maps and stochastic maps. Our findings can help researchers choose the most appropriate definition of disentanglement for their specific task and discover better metrics, models, and algorithms.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
Currently we are in the first stage of this project. Last year, we focused on exploring the various definitions of disentanglement in machine learning. We proposed that the concepts of the cartesian and monoidal products should serve as the core of disentanglement and argued that modularity and explicitness should be the defining properties of disentanglement. We showed how seemingly distinct formulations were just different manifestations of the same constructions in different categories. The findings can help researchers choose the most appropriate definition of disentanglement for their specific task and discover better metrics, models, and algorithms, contributing to a more comprehensive understanding of disentanglement in machine learning.
|
| 今後の研究の推進方策 |
Now that we have established a foundational understanding of the definitions of disentanglement using category theory, our next step is to develop a systematic way to enrich these definitions into metrics. This will enable us to quantify the degree of disentanglement achieved by a given model and provide a better understanding of how well it satisfies the defining properties. We also plan to investigate partial combinations of factors and unknown projections, as well as the relationships between different definitions. In addition, we are interested in exploring the application of disentanglement to more complex learning problems, such as reinforcement learning, and further analyzing the properties of the monoidal category case.
|
