Small-sample Deep Learning Method Based on Analytic Initialization of Convolutional Layers

2015 Fiscal Year Final Research Report

• Project/Area Number: 26730085
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Perceptual information processing
• Research Institution: The University of Tokyo
• Principal Investigator: Nakayama Hideki 東京大学, 情報理工学(系)研究科, 講師 (00643305)
• Project Period (FY): 2014-04-01 – 2016-03-31
• Keywords: 深層学習 / ディープラーニング / 画像認識 / 畳み込みニューラルネットワーク / 表現学習 / 人工知能

Outline of Final Research Achievements

We proposed a new deep learning methodology that enable fast and stable learning from limited amount training examples. We stacked convolutional layers using discriminative analytic solutions obtained by Fisher weight map to build multi-layer convolutional neural networks. Then we further fine tune the entire network by means of the standard backpropagation to quickly reach better local minima. Proposed method achieved state-of-the-art classification accuracy on some standard benchmarks, namely MNIST and STL-10.

Free Research Field

知能情報処理