2019 Fiscal Year Final Research Report
A Study on Deep Convolutional Neural Network Considering the Mechanism of Visual Information Processing in the Brain
Project/Area Number |
17K00251
|
Research Category |
Grant-in-Aid for Scientific Research (C)
|
Allocation Type | Multi-year Fund |
Section | 一般 |
Research Field |
Perceptual information processing
|
Research Institution | Tokyo University of Technology |
Principal Investigator |
KIKUCHI Masayuki 東京工科大学, コンピュータサイエンス学部, 講師 (20291437)
|
Project Period (FY) |
2017-04-01 – 2020-03-31
|
Keywords | 深層学習 / 畳み込みニューラルネットワーク / 脳 / 視覚系 / モデル |
Outline of Final Research Achievements |
In order to refine the CNN (Deep Convolutional Neural Network) causing the innovation on our society, this research project aimed to add novel functions to the CNN and investigated the properties of the CNN from the viewpoint of brain information processing. First, the usefulness of the medial axis representation which is regarded to be encoded in the brain, is examined by applying medial axis transformation to input patterns. In addition, the model structure was generated automatically by an evolutionary computation to approach the function of the brain having rich structures. On the other hand, we investigated the possibility of U-Net approximating the visual system roughly, to simulate visual functions such as figure-ground separation and visual completions, and performed spine extraction from CT images. Moreover, we investigated the relationship between the accuracy of recognition by CNN and the response distribution, and gained a foothold for efficient model setting.
|
Free Research Field |
脳の情報処理
|
Academic Significance and Societal Importance of the Research Achievements |
昨今,社会の諸領域に変革をもたらしつつある深層学習の中核技術であるCNNは元々,1960年代の一部の脳科学的知見を工学的にモデリングしたものがベースになっている.更なる脳の知見の投入により,深層学習はもっと進化する可能性も考えられる.万能のように扱われる既存の深層学習も現状では狭い問題設定の枠内でのみ有効性が確認されており,人間の脳ほどの汎用性はまだ獲得されていないこと,そして大量のデータがないと高い能力を発揮することができず,少量のデータのみでも内在する性質を適切に学び取れるヒトに及ばないことなどの難点があり,本研究課題の観点はこれらを解決する糸口を提供し得るものと言える.
|