2022 Fiscal Year Final Research Report
Developing a Design Methodology of Deep Neural Networks to Accelerate Paradigm Shit Brought by Deep Learning
Project/Area Number |
19H01110
|
Research Category |
Grant-in-Aid for Scientific Research (A)
|
Allocation Type | Single-year Grants |
Section | 一般 |
Review Section |
Medium-sized Section 61:Human informatics and related fields
|
Research Institution | Tohoku University |
Principal Investigator |
|
Co-Investigator(Kenkyū-buntansha) |
菅沼 雅徳 東北大学, 情報科学研究科, 助教 (00815813)
|
Project Period (FY) |
2019-04-01 – 2023-03-31
|
Keywords | 深層学習 / コンピュータビジョン |
Outline of Final Research Achievements |
Deep learning requires developers to design a network architecture that is appropriate for each individual task that they want to solve, but no methodology or guidelines have been established for designing good architectures. We aimed to solve this issue by designing architectures that achieve high performance in various tasks. We have successfully developed networks that achieve the highest accuracy (at the time of publication) for various tasks, including image restoration, image understanding, 3D geometry estimation, uncertainty estimation, and self-supervised feature learning. While the results have significant impact for each of the targeted tasks, their integration provide a foundation toward establishing the methodology of neural architectural design.
|
Free Research Field |
コンピュータビジョン
|
Academic Significance and Societal Importance of the Research Achievements |
深層学習は,近年発展著しい人工知能の中核技術であるとともに,その他の工学やサイエンスにも大きな影響を与えつつある.その一方で,深層ニューラルネットワークの構造設計に確たる方法論がないという課題があった.本研究は,様々な応用ごとに優れた性能を発揮するネットワーク構造の研究を通じて,それぞれの応用問題の解決に貢献するとともに,ネットワーク構造に関する新たな知見を多く生み出した.これらの成果は,構造設計の方法論の基盤を与えている.
|