NMFによるcIPS野における立体的特徴認識のモデル化

スポンサーリンク

概要

• 論文の詳細を見る

• Non-negative matrix factorization (NMF) is a technique which learns non-negative bases from many non-negative inputs such that they can be approximated by multiplication of the bases and non-negative weights. Lee and Seiing applied NMF to facial images and showed that the basis images represented parts of faces like eyes, noses, and mouths. Also, Tsunoda and others measured responses of neurons of monkeys which saw various objects and showed that neurons of inferotemporal (IT) cortex responsed selectively to parts of objects. From these it is suggested that computation similar to NMF may be used in IT cortex. Considering the uniformity of the circuit of the cerebral cortex, it is possible that other areas may use the same information processing princple. Shikata and others showed that neurons in cIPS cortex responsed selectively to three-dimensional orientations of objects. In this paper, I hypothesize that 3D feature recognition in cIPS is also done by NMF and validate it by the following simulation. I used images of which pixels have distance information as a model of signals from V1/MT cortex that gives projections to cIPS cortex and I applied NMF to these images. I used the datasets from Middlebury benchmark and some images which are generated by a stereo matching algorithm as input images. I will compare the result with response properties of cIPS neurons and discuss the possibility that cIPS performs computation similar to NMF.

• 社団法人電子情報通信学会の論文
• 2010-03-02

著者

• 細谷 晴夫

  東京大学理学部

• 中村 尚太

  東京大学理学部情報科学科

• 細谷 晴夫

  東京大学理学系研究科情報科学専攻

関連論文

• BESOMモデルに基づく階層的時系列の学習と認識
• BESOMを用いたスパース符号化の一手法
• NMFを用いた大脳皮質MSTd野におけるオプティカルフローのパターン抽出
• プロセスとチャネルに基づく並列言語のための部分計算
• NMFによるcIPS野における立体的特徴認識のモデル化
• 初期視覚野におけるBorder-ownership信号の確率論的解釈
• 抽象カテゴリ学習

スポンサーリンク