DATA ANALYSIS SYSTEM, METHOD AND PROGRAM

• Main Authors: KONISHI HIROSHI, KURAUCHI YUKI, SESHIMO HITOSHI, NISHIMURA TAKUYA
• Format: Patent
• Language: English; Japanese
• Published: 31.10.2019
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

To provide a data analysis system capable of appropriately performing analysis while reducing communication traffic amount.SOLUTION: A data analysis system 90 includes an instrument 10 that performs a conversion processing for outputting low-dimensional observation data which is output of an intermediate layer obtained by processing observation data received via an input layer of trained neural network 18A from the input layer to the predetermined intermediate layer, and a device 20 that inputs the low-dimensional observation data at a trained neural network 18B to an intermediate layer next to the predetermined intermediate layer, and performs an analysis processing for setting the output of the output layer to a result of the observation data using the next intermediate layer and the output layer. The trained neural networks 18A and 18B are configured such that a node number of the predetermined intermediate layer is set to less than a node number of the output layer, and are learned in advance so that duplication of probability distributions of the low-dimensional observation data regarding to the observation data that are different in the results of analysis under predetermined control is less relative to when there is no predetermined control.SELECTED DRAWING: Figure 1 【課題】通信量を削減しつつ、適切な分析を行うことができるデータ分析システムを提供する。【解決手段】データ分析システム９０は、学習済みニューラルネットワーク１８Ａの入力層を介して受け付けた観測データを、入力層から所定の中間層まで処理された結果得られる中間層の出力である低次元観測データを出力する変換処理を行う計器１０と、学習済みニューラルネットワーク１８Ｂにおいて低次元観測データを所定の中間層の次の中間層に入力し、次の中間層及び出力層を用いて、出力層の出力を、観測データを分析した結果とする分析処理を行う機器２０とを含む。学習済みニューラルネットワーク１８Ａ、１８Ｂは、所定の中間層のノード数が出力層のノード数よりも少なくなるように構成され、かつ、所定の制約の下、分析の結果が異なる観測データについて低次元観測データの確率分布の重複が、所定の制約がない場合と比べて少なくなるように予め学習されている。【選択図】図１