2023 Fiscal Year Final Research Report
Teacher-Student Sequential Re-learning Model in Adaptive Structure Deep Learning
| Project/Area Number |
19K12142
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 61040:Soft computing-related
|
|---|
| Research Institution | Prefectural University of Hiroshima |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2019-04-01 – 2024-03-31
|
| Keywords | 構造適応型深層学習 / 構造適応型RBM / 構造適応型DBN / Teacher-Student構造適応型深層学習 / AffectNet / 軽度認知症画像判定 |
|---|
| Outline of Final Research Achievements |
An adaptive structural learning method called Adaptive RBM-DBN, which automatically determines the optimal number of RBM neurons and layers for the training data, has shown high classification accuracy. However, the accuracy cannot be expected to improve in the case of ambiguity or inconsistency, such as medical images or emotions. We proposed the Teacher-Student (T/S) adaptive structural deep learning model as an ensemble learning method using two or more models. We applied the model to facial emotion data and dementia data, and found that the classification performance was improved. Furthermore, we proposed a method for real-time discovery of evacuation routes while avoiding roads damaged by landslides using aerial and satellite images during the torrential rain disaster in western Japan, and demonstrated the effectiveness of the proposed method.
|
| Free Research Field |
深層学習
|
| Academic Significance and Societal Importance of the Research Achievements |
IoTや医療等，データが収集・蓄積されたデータをそのまま画像解析やデータ分析に使用されている。学習時には出現していなかった特徴をもつデータが混在していることがあり，これらを外れ値やノイズとして処理するのではなく，一つの新たな特徴として処理する必要がある。Teacher-Student構造適応型深層学習モデルは実在するデータを無視するのではなく，再学習した知識を蒸留する形で，適切な構造をもつ深層学習法の一つとして提案した。
|
