2021 Fiscal Year Final Research Report
Constructing Machine Learning Framework for Set Data Based on Kernel Mean Embeddings
Project/Area Number |
18K18112
|
Research Category |
Grant-in-Aid for Early-Career Scientists
|
Allocation Type | Multi-year Fund |
Review Section |
Basic Section 61030:Intelligent informatics-related
|
Research Institution | Chiba Institute of Technology |
Principal Investigator |
Yoshikawa Yuya 千葉工業大学, 人工知能・ソフトウェア技術研究センター, 主任研究員 (30772040)
|
Project Period (FY) |
2018-04-01 – 2022-03-31
|
Keywords | 機械学習 / 解釈可能性 / 透明性 / ガウス過程 / ニューラルネットワーク / 集合データ / カーネル法 / 人工知能 |
Outline of Final Research Achievements |
We developed machine learning models and a software that can be applied to aggregated data for the problem of interpretable machine learning, which reveals which features contribute to the prediction and to what extent, along with the predictions of target variables. Two main types of machine learning models were developed: the first was based on Gaussian process regression. We developed a new Gaussian process regression model, in which the coefficients (weights) of the local linear model corresponding to each sample are formulated to be generated based on a Gaussian process. The second method is based on neural networks. We developed a neural network-based model that generates the K coefficients (weights) of the local linear model corresponding to each sample from the most important ones. For the developed Gaussian process regression-based model, a software was developed and released so that users can easily use it in their own machine learning systems.
|
Free Research Field |
機械学習
|
Academic Significance and Societal Importance of the Research Achievements |
ガウス過程回帰モデルは、カーネル法における教師あり学習の代表的な手法であり、幅広い分野で応用されている。本研究の成果は、従来のガウス過程回帰モデルの予測精度を維持したまま、現在の機械学習システムにおいて必要不可欠な予測結果の解釈可能性を高めるものであり、学術的・社会的の両面で影響を与えるものである。これは、ニューラルネットワークに基づく手法についても同様である。また、ユーザが人工知能システムに本研究の成果を容易に導入できるようにソフトウェアを開発・公開しており、本邦の「人間中心のAI社会原則」で求められている「AIの透明性の確保」の一助になると考える。
|