2023 Fiscal Year Annual Research Report
Algebraic Structures in Weakly Supervised Disentangled Representation Learning
| Project/Area Number |
22KJ0880
|
|---|
| Allocation Type | Multi-year Fund |
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
張 一凡 東京大学, 情報理工学系研究科, 特別研究員(DC2)
|
|---|
| Project Period (FY) |
2023-03-08 – 2024-03-31
|
| Keywords | Machine Learning / Category Theory |
|---|
| Outline of Annual Research Achievements |
In this project, our primary objective was to develop theoretical frameworks and practical algorithms for acquiring abstract and meaningful representations in machine learning. Our focus centered on the concept of disentanglement, aiming to provide researchers with tools to analyze and quantify the degree of disentanglement achieved by different models across various scenarios.
Our research began with a comprehensive meta-analysis of existing definitions of disentanglement in machine learning. This analysis provided us with valuable insights into the diverse perspectives and approaches within the field. Building upon this foundation, we proposed a new theoretical framework for converting disentanglement definitions into quantitative metrics. Leveraging concepts from topos theory and enriched category theory, we introduced tools to analyze disentanglement and assess model performance.
We developed a novel theoretical framework that bridges the gap between disentanglement definitions and quantitative metrics. By leveraging concepts from topos theory and enriched category theory, we provided a structured approach for analyzing disentanglement across different scenarios.
In addition to theoretical advancements, we also developed practical algorithms to implement and apply our theoretical frameworks. These algorithms enable researchers to assess the degree of disentanglement achieved by their models and select evaluation metrics of disentanglement for their specific tasks.
|
Research Products
(8 results)
