| 研究課題/領域番号 |
23K11235
|
|---|
| 研究種目 |
基盤研究(C)
|
| 配分区分 | 基金 |
|---|
| 応募区分 | 一般 |
|---|
| 審査区分 |
小区分61030:知能情報学関連
|
|---|
| 研究機関 | 岩手県立大学 |
|---|
研究代表者 |
藤田 ハミド 岩手県立大学, 公私立大学の部局等, 特命教授 (30244990)
|
|---|
| 研究期間 (年度) |
2023-04-01 – 2026-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
4,420千円 (直接経費: 3,400千円、間接経費: 1,020千円)
2025年度: 1,040千円 (直接経費: 800千円、間接経費: 240千円)
2024年度: 910千円 (直接経費: 700千円、間接経費: 210千円)
2023年度: 2,470千円 (直接経費: 1,900千円、間接経費: 570千円)
|
|---|
| キーワード | machine learnig / Health care prediction / 知能情報学関連 / Machine Learning / Health Care System / Deep Learning |
|---|
| 研究開始時の研究の概要 |
The unique features in this study are: (1) on-line feature selection from data stream using incremental learning on multiclass classification, (2) the fusion of different cross layered CNN and multiclass classifiers collectively representing features extraction of different health symptoms running on GPUs for leaning and testing (3) Ensemble of different deep semi-supervised leaners to predict early health risks, all employing: (a) robust real-time on-line semi-supervised learning systems, and (b) generative Neural Network (GAN) for raw data: for health risk predictions of high accuracy.
|
| 研究実績の概要 |
This project is challenging research areas in machine intelligence based on a combination fusion of CNN in clusters of GPU representing online ensemble deeper learning techniques fused for providing automated predictions of high accuracy. To establish such technology there are several research problems that PI needs to consider achieving the stated objectives. Traditional deep neural models neglect the spatial structure information of the data. When they process the data, they consider it together as a vector, wasting a lot of spatial information. Recent works have shown that CNN with more layers can get better performance than shallow ones. The system proposed in this project is to make full use of the information (features) deep learned from all the CNNs’ cross-layer neurons, to define the logical measures for local feature elicitation. It should have the capability to train effective deeper CNN. The PI is targeting two goals: (1) aim to design a deep learning network that can integrate all the features learned from all the lower-level layers and (2) send them to the higher-level layer for subsequent analysis of such network.
|
| 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
The year 2023 was smooth in progressing to achieve good results, that is published : Toshitaka Hayashi, Dalibor Cimr, Hamido Fujita, Richard Cimler, "Interpretable synthetic signals for explainable one-class time-series classification, "Engineering Applications of Artificial Intelligence, Volume 131, May 2024, 107716, https://doi.org/10.1016/j.engappai.2023.107716, the outcome in this publication is a big milestone in the execution plan of this project.
We made an extensive review: “Critical Review for One-class Classification: recent advances and the reality behind them" and submit this to Information fusion Journal, We made extensive experiments: Considering local cluster balance features for one class image classification without resizing using feature extraction method from clustering results for sliding windows in images. The results submitted to international journal.
|
| 今後の研究の推進方策 |
The main novel contribution; is the transformation subtask from time-series data and image data for high accuracy classification [1,2]. Such a study has significant meaning because images and time series have substantial differences in terms of data structure. The research project is using fused deep learning techniques for time-series data streams-based health risk predictions to construct: (1) Deep Neural Networks (DNNs) based on fused CNN in architecture in cross-layered connection, called as Block 1. (2) fusion of CNNs for learning feature and prediction trained of multi-patch gradients for best classification convergence (Block 2) to construct resilient network employing the methodology reported in PI’s work in [1, 2]. (3) Real Data streaming of online feature selection: it is the source of time series data from multi-sensing which is usually imbalanced and have concept drift from what the system is learned based on pre-labeling. Missing values and data imbalance are big problems in data streaming affecting the accuracy of online learning algorithms: (Block 3) is real-time streaming-based data structure extraction by active learning on proposed CNN architecture. These mentioned three blocks are to be validated in this project. The fused learning features with optimised hyper parameters setting will be tested on public time series data and then on real outsourced data for validation purposes. This is based on data availability and the status of fusion procedure.
|
